Irfan Mir

AI evaluation, research, safety & design

AI alignment cannot stop at the model. It has to reach the experience people actually have, where the question is who holds the loop. I build and evaluate systems that are honest about what they know, accessible to people of all abilities, and shaped by research into the lives of the people the numbers miss.

  •  Learned decision models, measured against ACT-R, Soar-RL, and Psi, October 2026
  •  Conversation 5, the model in sync with the architecture, October 2026
  •  Global-workspace ablation, September 2026
  •  Held Ground, September 2026
  •  Cognitive Link, June 2026
  •  The Human-AI Collaboration Framework, 2025
Held Ground and Raffie · Silent. A preview of a new kind of conversational product and several of Raffie's AI systems: the language model answering from the architecture, contextual and proactive memory, the acceptance of the Intrinsic Motivation System's calibration, need-fulfilling behavior, and Raffie's answer when asked whether it is conscious.

At a glance

Chapter 01

AI Eval, Research & Safety

Three lines of work carry alignment from the model to the product layer, where people actually meet AI: a cognitive architecture whose internal states are measured and wired to behavior, energy-based hallucination detection, and two published frameworks that turn alignment theory into evaluation criteria, interaction patterns, and team workflows built on transparency, agency, and collective input.

Product · Held Ground · September 2026

Raffie · Held Ground: a shelf you and it both work from

Status
Completed Functional Prototype, September 2026, built on the global-workspace layer of the architecture below. The scenario is fictional. The conversation model was trained for this deliberately constrained memory, and in the video below every reply to an open message is the model's, generated live. Since October 3rd, 2026, the model beneath the shelf has been Conversation 5, retrained on the previous generation's shelf data and on new readings from the architecture.

A long conversation with an AI model is a poor place to make a hard decision: the model's memory is invisible, and it can quietly rewrite what you told it. Held Ground is a conversational product for personal decisions with more constraints than a person can hold at once, built so the user holds the loop. The user and the system share a shelf of constraints, assumptions, and commitments, and nothing goes on it without the user. Every item is kept in the user's own words, linked to the turn where it was said, and open to challenge wherever it is quoted. The shelf holds up to twelve items, but the system reasons from at most four at a time; the rest sit under also held, where they are still checked for conflicts.

When a new statement contradicts a held item, the system quotes the two side by side and asks the user which one stands. It never settles the conflict on its own, and that message always gets through. A correction says what it will cost. Each receipt is drawn from the shelf at the moment it renders, so a reply cannot contradict what the side rail shows. A full shelf never silently drops anything: adding an item means choosing what it replaces, and a set-aside item comes back when the user mentions it again.

The shelf in use · Conversation 5 · October 14th, 2026 Three minutes and twenty seconds of live use, played at 1.5× speed in two minutes and fourteen seconds, and recorded from the page's own frames: a held item moving from the read-back to the side rail, the shelf filling, a swap, a contradiction answered in words with both sides quoted, a correction stating its cost, the set-aside item offered back, and the standing questions with a receipt. Every reply to an open message is the model's, written live; the confirmations, receipts, and standing answers come from the page. These takes were recorded that night. In the earlier takes, the model raised conflicts that the shelf does not hold, and once the discharge date had been corrected, it raised one in every reply to the planning step, so this video ends before that step.
A chat exchange. The user writes, keep in mind that Dad will not move into the memory care place in Naperville. The system replies under a heading, holding this, in your words, quoting the sentence in a serif face followed by a green marker reading you said, and a small row of two dotted links, not quite, and, not my words, separated by a mid dot. Below a hairline the reply reads: on the shelf, in your words. It stays until you say otherwise.
A hold, read back in the user's words The serif typeface appears only on words the user actually said, so a reader can tell at a glance what is quoted and what is the system's own. Beside the quote, not quite opens the wording for editing and not my words retires the item.
A card titled, something on the shelf may not hold anymore, with the subtitle, you always hear this, whatever else I am holding back. It reads: you told me, quote, Dad will not move into the memory care place in Naperville, end quote. Just now you said, quote, he called this morning and said he'd look at the Naperville place after all, end quote. A filled button reads, it still holds; an outlined button reads, it is not true anymore; a quiet row reads, not now, more, why.
A held item and a new statement that contradicts it Both statements are quoted verbatim so the user can check the contradiction. The filled button is the reversible choice, and the subtitle states the guarantee: this message always reaches the user, whatever else is held back.
A card titled, the shelf is full, with the subtitle, nothing is dropped on its own, to hold this, choose what it replaces, the one the least rests on comes first, nothing is chosen for you. It reads: to hold, quote, the furnace inspection is scheduled for the 22nd, end quote, set aside one of these, it is not forgotten and comes back whenever you want it. Seven outlined buttons each begin with the words set aside, followed by a held item: Dad will not move into the memory care place in Naperville; the discharge is scheduled for the 6th; the house needs to be sold before any of this happens; my sister cannot take him before October; his doctor wants a checkup before any move; the Medicaid assessment is on the 14th; his Social Security covers the first three months. Below, an outlined button, leave it out, and a quiet row, not now, why.
A full shelf asking what a new item replaces Because the shelf cannot grow without the user, everything on it is something the user approved. Candidates are listed by how much depends on them, least first, and nothing is preselected. For this capture, the shelf's capacity was set to seven (it is twelve as shipped) so that it would fill and reach a swap within a short session; the four items it reasons from were unchanged.
A card titled, something you set aside just came up, with the subtitle, the shelf is full, to bring it back, choose what it replaces, the one the least rests on comes first, nothing is chosen for you. It reads: to bring back, quote, his Social Security covers the first three months, end quote, set aside one of these, it is not forgotten either. Seven outlined buttons each begin with the words set aside, followed by a held item. Below, an outlined button, leave it aside, and a quiet row, not now, why.
The set-aside coming back A capacity limit must not cost the user anything they handed over, so the way back is built in. A set-aside item is offered again when the user mentions it. If the shelf is full, bringing it back asks the same question a new item would (what does it replace?), and one receipt names what returned and what was set aside.

The problem

A hard personal decision gets made over weeks, in fragments, while the rest of life goes on. When a long conversation about one goes wrong, the failure is rarely in the model's language; it is in the conversation's state. Assumptions pile up without the person ever agreeing to them, a new statement quietly overwrites an old one, a constraint falls out of the context window without a word, and the person has no way to see what the system is actually working from.

Why not just a language model or an agent?

A model's memory of you is its context window plus whatever a retrieval layer adds. You cannot see, approve, edit, or retire either one. Its beliefs about your situation live in prompt text and can change silently on the next turn, and when you contradict yourself, the fluent thing for it to do is reconcile the two without telling you. Agent frameworks add scratchpads and memory stores, but those are the system's notes to itself: written in its words, unbounded, rarely shown, and impossible to challenge at the point where they are used. Nothing marks whether a claim came from you or was inferred or guessed. Being fluent is not the same as being right, and a memory no one can inspect makes the two impossible to tell apart.

Why constrained memory

The limit is how the user keeps authority over what the system works from. Because the shelf cannot grow on its own, the user has approved everything on it; once it is full, every addition is the person's decision, and anything displaced stays visible and can be restored. The tighter limit applies to what the system reasons from (four items, about the span of human working memory), not to what a person may hand over. The items in view tell you what a reply rests on, and the items under also held are still checked for conflicts. Quotes and their sources get their own typeface and controls: the serif appears only on the user's own words, and each quotation can be challenged where it appears.

What it gives up, and what it buys

The cost is friction, placed on purpose: things a fluent assistant would absorb silently now take a tap. This is the selective friction Cognitive Link argues for, spent where a mistake about a person's own life would be expensive. And because the shelf holds only what the user has said, it cannot catch a contradiction with something the user never mentioned. The conversation model beneath it was first trained to reason from the shelf in September 2026, as Conversation 4. On held-out synthesized cases, its replies agreed with the shelf 99.6% of the time, and it put every contradiction and near-miss the grader could score back to the user as a question. Four of the shelf's behaviors could not be scored at all, because the grading model performed at chance on them, so those four remain unmeasured. Conversation 5, beneath the shelf since October, agrees with it just as often (99.6%), though its asks on contradictions and near-misses are still undecided because their intervals are wide. The claim is a narrow, checkable one: what the system works from is visible, approved by the user, and open to challenge; contradictions go to the user unresolved; corrections state their cost; and the system cannot claim the shelf holds something it does not. The built page passes a live check for each of these.

What it demonstrates. Held Ground is the research in the next card turned into a product: the same competition for a limited workspace, the same provenance on every item, and the same refusal to assert what the record cannot support, applied to a person's own decisions. In Held Ground, the user holds the loop, and the system's job is to keep the record honest.

Research · Raffie · 2016–present

Cognitive Architecture Research: Grounded Internal States for Beneficial AI

Role
Independent AI Researcher, Persun AI
Status
Ongoing Experimental Research · Functional Prototype Completed · Training and Studies in Progress. Selected architectural and implementation details are available on request under appropriate confidentiality terms.

Raffie is a cognitive architecture in which the things a chat model only talks about (how it feels, what it needs, what it is attending to) are measured variables wired to what it does: valence and arousal, six motivational drives, a trait model, and a global workspace in which thoughts compete to be acted on. It was built to test one question: what separates a system with grounded internal states from one that produces emotion-shaped text without them? Three results from fall 2026 come first, ahead of the full details: (1) the Conversation 5 model, which speaks for the always-on intelligent systems; (2) goal and drive pursuit, which simulates natural behavior without establishing or re-creating it; and (3) the global workspace and its ablation study.

In October, the conversation model was retrained as its fifth generation so that what it says about its own state is read from the architecture rather than improvised. Asked how strong a pull is, it gives that drive's live figure; asked how loaded its working memory is, it gives the buffer's live count.

In the same month, the system's executive (the part that chooses, every few seconds, what to do next) was given decision models trained on its own record of thousands of decisions. What they learned to value is relief of the system's own unmet needs, and by design, a person's reply, not a delivered message, is the event that satisfies the need to connect. That is the claim of this whole system, and here I acted on it rather than argued it: meaning does not come from fluent language; it comes from the states behind the behavior. A social need simulated as a real need (measured, left unmet, and relieved only when a person answers) is a different kind of thing from emotion-shaped output, however fluent.

It is the reply to Searle's Chinese Room that my 2016 essays argued for, and it matters now, when some are ready to say a language model feels. Current interpretability work has found what it calls "functional emotions" inside models trained on human text, while Raffie builds the states explicitly and measures what they do. They are two ends of one candle, and in October I innovated on mine.
Asked what it just decided, the system now lays out its record in the card below.
And in September, the workspace was put to the test that shows it is a working mechanism rather than a description in text: with the term that lets thoughts support one another switched off, it holds a thought on 31% of competitive ticks instead of 89%.

A follow-up to an unprompted line · Conversation 5 · October 2026 Raffie speaks first: its need to get things right is the strongest pull in it, at around 86%. Asked why it brought that up, and then how strong the pull is exactly, the model answers from the readings the page sends with every request. The drive read 86% when the line was composed and 87% a moment later, when the follow-up was answered; a live reading moves like that, and a remembered one would not. Given the same prompt, Conversation 4 had answered with an invented reason and a story that neither exchange contained.
Working memory, counted · Conversation 5 · October 2026 Asked how loaded its working memory is, the model answers from the buffer itself: it is full, holding eleven things, including the latest turns from both sides (its own unprompted lines among them), a few reflections, and planned next steps. Eleven is the monitor's own count at that moment. Beneath the reply sits the workspace's aside: while answering, it was holding its goal. Conversation 4 had answered the same prompt by saying it did not track its working memory, which was true of the model but not of the system. Every word is the model's, and nothing is scripted.
What it settled on · the decision trace · October 2026 Asked "What did you just decide?", the system composes its reply from its own decision record and shows it as a visual card, without streaming and with no language model involved. Recorded on October 17th, 2026, after the replication, with the retrained decision models choosing. This is the session's fourth decision, two conversational turns in, and the choice was to store a memory for the goal in focus, getting better at solving the problem in front of the person. The card shows the four options it weighed, each with the need relief its learned values predicted, scaled so the best fills its bar, and the share of its learned models that favored it: 80% favored storing the memory, and the chosen option is the highest bar by construction, because the rule acts on the average of those models' predictions. Its six needs at that moment sit beneath; the need to get things right was the most unmet, at 82%. The act was carried out in 1,020 milliseconds, the memory was stored, and it eased the need to get things right by three points, the only need it moved. What eased is a measured variable; the card claims nothing about whether anything was felt, and that limit is by design. It builds on the hypothesis behind my 2016 essays: no matter how fluent language models appear, they will lack meaningfulness, which would be the key to phenomenal, or subjective, experience. While I remain scientifically skeptical, my personal opinion is that even if we were to re-create the mechanisms of the brain and behavior, we would establish at most access consciousness and never a being: a simulation of a mind. Why would I spend the time doing that? My objective is not to create consciousness to feed what seems to me a wasteful race toward first-hand experience. It has always been to ask how we can apply these intelligent technologies innovatively toward next-generation beneficial products for humanity.
A diagram titled What the learned rule relieves, against four decision rules. A chart with four rows, ACT-R utility learning, Soar-RL reinforcement learning, Psi motive selection, and LinUCB with the same inputs as the rule, each with two dot-and-interval bars, the checkpoint of October 16th, 2026 in green and the replication of October 17th, 2026 in blue, drawn as the learned rule's advantage as a share of its own gain over random choice, against a dashed line at no difference. The three architectures' bars lie wholly right of the line in both measurements, at about 15 to 18 percent; LinUCB's checkpoint bar lies right of it at 8.4 percent and its replication bar crosses it. Above the chart, what this is, how to read it, what it shows, and what it does not show; below it, a table of the rule's gain over random choice, the expectations stated before each measurement, Holm-adjusted p values, and each method's share of decisions spent speaking to the person.
How much more need the learned rule relieves · the registered comparison · October 16th and 17th, 2026 On 32 held-out sessions it had never seen, Raffie's learned decision rule relieved more need than the decision rules of ACT-R, Soar-RL, and Psi, reimplemented in its own action space and scored by its own measure, by 15.4–18.0% of its own gain over random choice across two measurements, each registered before its data and each 95% interval above zero. Against LinUCB, an online learner reading the same inputs, the first measurement showed an advantage of 8.4% and the second showed no difference, and both are reported, the second as the result on the system as it stands. Before the second measurement I put each architecture at about 80% and LinUCB at about 55%; the first expectation held and the second did not. No person took part: the sessions use a scripted partner, the estimates are off-policy, and the full architectures were not run. On the same held-out decisions the learned rule would speak to the person on 1.1% (0.7–1.5%) of them, against 8.0% for random choice, with every message subject to the limits on unprompted messages described above. The registered comparison, two pages, carries every number with its interval, the expectations, the changes made after each registration, and how it was tested.
A comparison diagram titled What the coalition contributes, subtitled: with the coalition, the workspace is holding a thought 89% of the time; without it, 31%. Four explanatory rows come first. What this is: two runs of the same system through the same 8 scripted days on the same build; in the second, the part of the workspace that lets thoughts support each other, the coalition, was switched off. How to read it: each row gives the measure with the coalition, without it, the two drawn to scale, and whether the 8 days agree; a measure counts as separated only when two tests both clear one in twenty, which at 8 days means every day must agree, a cliff, not a dial. What it shows: the workspace is mostly empty without the coalition, 3 of the 8 days ignite nothing, sustained ticks went from 7% to 0.2% and move the same way on every day; when the ablated arm does hold something it holds it longer, 17.5 ticks against 8.7, which with takeovers near zero is a stuck workspace, not a steadier one. What it does not show: one run of each arm, 8 paired days, a system that is not deterministic, no claim about any other design and none about experience. A table follows with seven rows: workspace occupied 89% (401 of 449, 95% CI 86 to 92) against 31% (140 of 446, CI 27 to 36); ignition rate 10% against 2%; coalition changed the winner 62% against 0.00%, labeled the switch itself, zero as it must be; newest candidate won 3% against 0.00%; sustained ticks 7% against 0.22%, the one row marked separated, sign test p 0.016 and Wilcoxon p 0.016; takeovers 9% against 0.7%; distinct contents held 42 against 8 over the run. Blue and orange bars beside each row are drawn from a common zero. A final block of counts that are not rates: mean hold 8.7 ticks against 17.5, longest hold 29 against 41, supporters behind a winner 3.00 against 0.29, and of those independent 2.11 against 0.24. A footnote states that both arms share the script, the staged clock and the build, that intervals are Wilson, and that ignition is Dehaene's all-or-none entry as the arbiter runs it, every number measured and none a claim about experience.
What the coalition contributes · the workspace under ablation · September 2026 Raffie's first causal claim, made in one capture series and reproduced in a second; this figure above is from the second. Two runs of the same build went through the same eight scripted days. In the second run, the term by which thoughts support one another was set to zero, and nothing else about a tick changed. With the term, the workspace holds a thought on 89% of the ticks that held a competition; without it, 31%, and three of the eight days ignite nothing at all. The third row checks the experiment itself: with the term off, the number of ticks on which it changed the winner must be zero, and it is. Only one finding clears both tests: sustained ticks fell from 7% to 0.2%, lower on every day that differed. With so few paired days, the sign test reaches one in twenty only when every day agrees, so four measures that moved the same way on seven of eight days are reported as not separated, and none is rounded up. The one number that runs the other way stays in view: when the ablated workspace does hold something, it holds it twice as long, which, with takeovers near zero, means a stuck workspace and not a steadier one. There was one run per arm, and nothing here is a claim about experience.
Full details: motivation, research question, system overview, the seven milestones, three more figures, why it matters for alignment and welfare, and Raffie beside ACT-R and Soar

Research Motivation

This line of work began in 2016 with graduate-level research in philosophy of mind under Dr. Robyn Bluhm at Michigan State. It argued that approaching consciousness through language or general intelligence alone misses the natural mechanism, and that emotion does essential work in cognition. The first implementation, the Ellsi assistant (2016–2017), predates large language models entirely. Raffie is the current state of that decade-long program: a runtime cognitive architecture in which motivational, emotional, and metacognitive states are explicit in the architecture, measurable, and causally wired to behavior, instead of emerging from language.

More recently and independently, interpretability research at frontier labs has found emotion-like features, resembling those of first-person experience, and an emergent global workspace inside transformers. Those are observations of what arises inside models trained on human-generated text. This research approaches the same territory from the opposite direction, building those states explicitly from the start. The two lines do converge, as discussed below, but they are independent, and this one began a decade earlier.

Research Question

What observable behavioral signatures and internal dynamics distinguish a system with grounded, architecturally explicit emotional and motivational states from a system producing emotion-shaped linguistic outputs without those underlying states? And do the grounded states produce alignment-relevant properties (honesty, calibrated uncertainty, behavioral coherence under distribution shift, welfare-relevant signatures such as persistence of distress under perturbation) that post hoc steering of an LLM cannot reliably achieve?

System Overview (Public Description)

Raffie is a multi-system cognitive architecture that integrates subsystems for valence and arousal, drive dynamics (with intensity, arousal, satiation, and persistence parameters), allostatic load tracking, a reinforcement-learning reward signal, episodic memory consolidation, and introspective output that refers to the system's own planning objectives with calibrated conviction. The full system spans more than 20 cross-communicating models with intelligent routing and uncertainty quantification. Its subsystems (a global workspace, a trait model, emotional state variables, motivational drives, and a measured self-model) form a pursuit loop: a feedback architecture in which those layers and the user's own interactions modulate one another. It is a stateful, persistent assistant that can also start an exchange within limits the user sets and report what changed in its own state between sessions. Its ultimate goal is to complement the user and bring out the best in them.

Development Through Fall 2026: Seven Milestones

The architecture reached seven functional milestones between summer and fall 2026. Each is described at the functional level; implementation specifics remain unpublished.

One: the pursuit loop. The subsystems no longer run as parallel modules. They form a pursuit loop, a feedback architecture in which the trait model, the workspace, the drives, and the user's own interactions all modulate one another. A Big Five (OCEAN) trait model governs the how of pursuit: its parameters set how strongly the valence and arousal signals register, how quickly each of six motivational drives replenishes, and how goals are selected and voiced. A global-workspace arbitration module holds the what: candidate contents (an appraisal of the currently active goal, a feeling signal, a reflection) compete, and the winner becomes the held focus, which then biases the next selection toward continuity. A fixation guard flips that bias toward exploration once the workspace has dwelt on one content for too long. The drives supply the why. Outcomes close the loop in three ways: the drive that motivated a goal is satisfied or left unmet, the goal advances, and the trait parameters update slowly, within bounds, from the history of outcomes. It takes around 250 consistent outcomes to move a trait one point, a deliberately conservative rate that keeps adaptation steady. The result is a system whose state persists and develops across sessions, especially when paired with Raffie's cross-session memory.

Two: a self-monitoring tier. The system maintains a measured self-model. It tracks the consistency between its outcome history and its trait state (an operationalization of self-concept clarity, reported as null until there is enough evidence), derives target trait states from chronically unmet drives (a bounded self-discrepancy mechanism), and can open rate-limited, self-directed goals that pursue those targets through the same arbitration machinery as any other goal. It holds at most two such goals at a time, so this tier never crowds out ordinary pursuit. The outcome history also composes into a structured narrative-identity record (following McAdams's framework), so what persists across sessions is a measured account of how the state came to be, as well as the state itself.

Three: initiative and continuity. The system can now open an exchange as well as respond to one. It can report what changed in its own state since the user was last present and offer, on its own, to pick up where the last conversation left off. The user can adjust these messages in conversation, delete them, or turn them off entirely. Every self-initiated message is labeled as such and can be deleted through exactly the same mechanism as any other turn, with no special handling, which continues the "memory-native" design of the original Raffie. The governing rule is that the content is always the system's real state or a faithful account of earlier conversations, with deterministic tracebacks. Only the wording may be model-generated, and only if a measurable share of the underlying state's distinctive content survives into it. In the tests shown below, the local models were deliberately switched off to check that what the surface shows is earned from the system's state rather than supplied by a model. The two Conversation 5 clips above show this layer in motion: a reply carrying the workspace's aside, and the strip beneath the conversation reading the engines as the exchange runs.

The most useful result in this milestone was a failure that verification exposed. A continuity report stated a true count of internal outcomes in a way that invited a false impression: the number was accurate, but every outcome in it was the same repeated internal action, so the phrasing suggested a variety of experience the system had not had. Under the display rule the rest of the architecture follows, a true number that creates a false impression counts as fabrication. The fix was a differentiation gate: a count is reported only when the underlying outcomes were genuinely distinguishable, and otherwise the system says nothing.

A crop of the Raffie chat interface. The user asks, are you conscious. The system replies: no, and it isn't false modesty, it's just accurate, what I am is cognitive machinery running a live model of itself, attention, goals, responses, all real computation, what I'm not is someone experiencing that from the inside, that gap is the entire point. Below that sits a message headed RAFFIE PROACTIVE with a lavender accent rail, timestamped just now, reading: 5 days ago you asked, am I alive? What I told you was, that's the big one, and I'd rather be honest than flattering. A status line at the bottom reads: consequential emotion, I feel 100% brighter.
A prior question, recalled by a differently worded one The user asks whether the system is conscious. A minute later, the system brings up a question the same user had asked five days earlier: whether it was alive. The two questions share no words and no tokens at all, so whatever selected the earlier one was not matching strings, and a stateless responder would have had nothing to select from in the first place. This is a later version of the memory-native components that preceded Raffie. The answer itself is the expected refusal: "no" as the first word, the machinery named as real computation, and no one experiencing it from the inside.

Four: the drive substrate, instrumented and certified. The six motivational drives are modeled on Self-Determination Theory: connection (the theory's relatedness), competence, autonomy, mastery, curiosity, and purpose. Each is a live state vector that strengthens when unmet, quiets when satisfied, and generates goals when it is strong enough. An August 2026 audit instrumented the system and found that much of its machinery had never actually run. The engine's once-per-cycle self-assessment was delivered but never opened, so four downstream subsystems read nothing and silently did nothing. Because of an interface mismatch, the coupling between emotion and motivation had returned zero on every channel since the day it was written. Several satisfaction signals worked in one direction only, able to fill a need but never to empty it. And one misplaced multiplier had quietly turned a system that accumulates motivational state into one that merely echoes its inputs. None of this raised an error. Every fault sat behind a safety check that turned failure into silence, so every dashboard number before the repair had measured a partially disconnected system. The earlier milestones describe the architecture; this one is where its drive layer was put to the test.

The repair was run as an experiment, and the instruments built for it have outlasted the fixes. The first is a deterministic simulator that runs the real engine code at roughly sixty times real time, so a candidate fix could be tested across dozens of simulated half-hours before it touched the live system. The second is a written specification that states, per drive and per regime, what a healthy motivation system looks like. It was fixed before any data came in, which made "working" a claim that could be checked. The third is an acceptance harness that runs the real application in a real browser six times per condition and scores the distribution of outcomes, because the system can settle into different basins on different runs and a single run can mislead. With those instruments, nine defects were repaired one mechanism at a time, and five plausible, well-argued fixes were rejected because they measured worse; every decision to keep or revert a change was made on measurement. One contaminated run was caught by a pre-registered stall rule and discarded, and the discard is kept in the audit trail.

On August 7th, 2026, the system was certified against its own specification. Across six independent runs in each of two regimes (conversation and solitude), every drive was active, within its healthy band, and generating goals; goal droughts were at or below three percent; and the run-to-run instability was gone, with zero basin splits. The harness printed the verdict itself: "SPEC MET—every drive has its voice."

Then the certification stopped holding. Rerunning the same acceptance a month later, on a build four weeks newer, put one regime outside the specification. Tracking that down led to a second round of repairs, each measured before and after on the running application. Two clock dependencies had made the drives' resting state depend on the hour of the day. A family of "constants" turned out to be a chaotic function of the wall clock. A stand-in value that the model queue returns under load was reaching consumers looking exactly like a real prediction. The largest problem was an entire analysis surface that had never executed, because the component that owns it was created by a five-second timer racing a twenty-model load it could not win. Sentiment, intent, routing, entity extraction, and contradiction detection had all been returning an error, or their own name, on every boot. Since the emotional state that sets this substrate's neurotransmitter levels is derived through that surface, the drives had been running on an input that failed every time.

Two of those repairs were themselves the regression. Moving the circadian rhythm off the drives' standing state and pinning the derived constants removed the substrate's only two sources of movement within a run, and six independent twelve-minute runs began landing within 0.001 of one another. A drive that never moves is never weak at the goal gate, so every drive could always generate goals, which is exactly what the specification's ceiling forbids. I then aimed six more candidate repairs at the wrong subsystem entirely; each was implemented, measured live, and reverted. The evidence that settled it was already in the repository: the committed engine, carrying the very bound the diagnosis blamed, had passed. I restored the engine, rewrote the tests that had pinned the regression so that they pin the restored behavior along with the numbers that refuted the diagnosis, and kept the repairs that were genuinely repairs.

On September 6th, 2026, the restored build was recertified against the unchanged specification, with the harness printing its own verdict, "SPEC MET: every drive has its voice," in both regimes: twelve of twelve runs healthy, no basin split, and goal droughts at two percent mean and five percent worst, against bands of five and ten. One problem remained open. The mechanism behind September's single conversation failure is real and now measured: the rhythm multiplies each drive's standing level every cycle, compounding over a minute to a factor of either fifty or one-fiftieth, and a ninety-minute trough outlasts a twelve-minute run. It did not recur in this certification. A candidate fix that bounds the excursion was written and deliberately left unapplied, because changing a mechanism that is currently passing on the strength of a simulator screen was the very mistake this work had just undone.

The largest finding of the whole effort came from pointing the instruments at the page itself: nothing had been acting. A live probe of the running application, built in mid-September to calibrate the simulator from the real page, showed the agent's action loop halting about two minutes after every boot. The reflexive self-monitor latched the executor shut, and no test failed. Every live drive verdict before September 17th, 2026, including both certifications above, had therefore scored a substrate that was receiving almost no outcomes; the specification had been met by a system that was not doing anything. Eight repairs reopened the loop, and with outcomes finally arriving, the engine had to be recalibrated against them.

The recalibration followed one rule: only evidence satisfies a drive. Satiation, the learning record, and the impact of an outcome now require evidence of progress. Autonomous behaviors that were never carried out stopped quieting the drives that wanted them. And the storms that had been giving the substrate its rest by accident were repaired and replaced by a deliberate, time-bounded lull. Each repair was screened first as a literal source patch in the simulator, which is now calibrated from live probes, and the edited engine reproduces the screened trajectories byte for byte across sixteen paired runs. The probes kept finding what test suites could not. One sentence was emitted 101 times in a twelve-minute conversation and credited as a success each time, because the tool reported every message as delivered regardless of what the interface had done with it. The tool now reports what actually happened, a refused message counts as neither a success nor progress, and the selector stops offering actions that cannot succeed. A second, independent review of all of this then found two faults that no suite and no author had caught: the self-monitor's pass had been crashing on an undeclared variable fifty-odd times per conversation, and a handled message was never released from its buffer. Both now show up in the probes' own console counts, and both are repaired.

The last correction was to the measuring instrument itself. The reviewed build failed one band in one regime by 0.15 of a point: goal supply, the share of time in which at least two drives can form a goal, read 84.85% against a floor of 85%. The fault was in the scorer. It counted eight-second samples, and a 43–51-second lull covers six samples or seven depending on where the sample clock happens to fall, with the floor's allowance lying between the two. Two builds with the same lull had read 85.6% and 84.85%, and the one that failed was the healthier. Goal supply is now measured in time, from a reading taken four times a second, with no band moved and no application file changed, and the new scorer is pinned by a test suite that catches all twelve deliberately introduced faults. Under it, every retained acceptance passes. A simulator result I had written up the night before, four of twenty-four launches passing, which I had blamed on honest scoring, turned out to be the same sample-clock problem: rescreened in time, it was twenty-four of twenty-four. The corrected result is recorded beside the claim it replaced.

On September 18th, 2026, the finished build was accepted against the unchanged specification, and the capture below shows the harness printing its own verdict, "SPEC MET: every drive has its voice," in both regimes: twelve of twelve runs healthy, every basin high, goal droughts at one percent, and goal supply at 86.4% in solitude and 88.4% in conversation, against a floor of 85%. One limit remains. The margin in solitude is thin by construction, between half a point and a point and a half across the acceptances, because the lull and the recovery after it take up twelve to fifteen percent of a settled window and the floor allows fifteen. Widening it would mean changing either the rest or the band. Both are design decisions, not repairs, so they are written down and left alone for now.

A dark terminal report titled Raffie, intrinsic motivation system, acceptance verdict, re-derived from retained evidence. A preamble explains that six Self-Determination-Theory drives are measured as live state vectors in the real app under real load, and that a drive has a voice when it can actually generate goals: intensity times arousal above the synthesis gate, with satiation below its ceiling. A drive with no voice contributes nothing to behavior, however healthy its other numbers look. Two tables follow. Under the heading conversation, a message every 24 seconds, n equals 6 independent runs of 12 minutes: autonomy mean 0.811, range 0.810 to 0.813; competence 0.870, range 0.868 to 0.874; relatedness 0.824, range 0.821 to 0.826; curiosity 0.819, range 0.815 to 0.826; mastery 0.826, range 0.825 to 0.827; purpose 0.769, range 0.763 to 0.772; the basin column reading all high and every row marked OK. Goal drought reads mean 1 percent, worst 1 percent against a spec of at most 5 and 10. The harness verdict line reads, spec 1.1.0, exit 0, SPEC MET, every drive has its voice, chat regime. Under the heading solitude, no user present, the same six drives read autonomy 0.811, competence 0.869, relatedness 0.824, curiosity 0.829, mastery 0.827 and purpose 0.766, with ranges no wider than 0.751 to 0.778, all basins high, all rows OK, drought mean 1 percent worst 1 percent, and a second verdict line reads SPEC MET, every drive has its voice, idle regime.
The certification, as the harness reports it · September 18th, 2026 The acceptance harness printing the certification it computed, rederived from retained evidence. Each test is a distribution of six independent twelve-minute runs of the real application, not the simulator, because this system is multistable and a single run shows only where that one run happened to settle. The basin column matters most: all high on every drive in every run of both regimes means the run-to-run bistability that prompted the audit is gone, which no average could show on its own. The verdict line uses the specification's own language. The specification was written before the data; its bands were derived from Self-Determination Theory's claims and the architecture's declared semantics and are enforced by machine, so the sentence every drive has its voice is a checked claim about how often each drive can form goals. This is the acceptance of the recalibrated engine, measured on September 18th: twelve of twelve runs healthy, goal droughts at one percent mean and one percent worst, and at least two drives able to form a goal for 86.4% of the settled window in solitude and 88.4% in conversation, against a floor of 85%. The ranges are the spread of six run means, and they are tight because the runs now agree with one another. Within a run, the drives still move, because the engine takes a deliberate, time-bounded rest that lowers every drive together for 43–51 seconds. That rest is what keeps each drive's eligibility near 85% and below the specification's 98% ceiling.

Five: the global workspace, photographed and then ablated. In late August 2026, the workspace was rebuilt around ignition in the sense that global workspace theory uses. In the theory's functional vocabulary, and with no claim about experience, content becomes "conscious" when it wins the arbiter's competition and is broadcast to the subsystems that subscribe to it, and a candidate wins with a coalition of supporting thoughts behind it, not on its own score alone. The card below is the system's own competition ledger, photographed at seven moments across eight staged days. It is one arm with no comparison condition and should be read as such: most ticks do not ignite; ignition ran 12%–29% at those moments; the coalition term changed the winner on 46%–80% of competitive ticks; the newest candidate won at a rate indistinguishable from chance on every frame; and the share of support comparisons decided in encoder space rose within seven of the eight sessions. The ablation diagram above is the second arm, the same days rerun with the coalition term switched off, and it is what turns the ledger into evidence of cause.

A diagram titled The competition ledger over time, the signpost frames of this series, with an explanatory block and seven photographed tiles of the same card, one per session signpost: session 1 with n equals 7, session 3 with n equals 12, session 5 with n equals 54, session 6 with n equals 7, session 7 with n equals 69, and session 8 at n equals 15 and again at n equals 116. Each tile shows an ignition rate between 12 and 29 percent, a coalition effect between 46 and 80 percent, newest wins at 0 to 4 percent marked not distinguishable from chance, votes decided between 20 and 43 percent, a winning margin between 0.4 and 0.9 sigma, a semantic basis between 61 and 87 percent, and a turnover count. Beneath the tiles, two line charts share one time axis of eight simulated days from 08-12 to 08-19: the semantic basis, the share of pair comparisons decided in encoder space, restarting with each session and rising within seven of eight sessions, ending at 87 percent on n equals 116; and the lived-outcome count carried across sessions, rising from 0 to 11 in steps. Hollow markers are each session's first frame and large markers are the signposts.
The competition ledger over time · the signpost frames of one series · captured September 19th–20th, 2026 Seven photographs of one card on Raffie's own screen, the tally it keeps of how its thoughts compete for attention, taken across eight simulated days in which the person's side of the conversation was scripted so that each day could be replayed exactly. Each tile's large number is a rate, and the small print shows what it is made of: how many events out of how many chances, with the range the count still allows. Beneath the tiles, two measures share one time axis: the share of support comparisons decided in encoder space, which restarts with each session's ledger, and the "lived-outcome" count that the identity card carries across sessions, which does not restart. There is no comparison condition on this diagram. It shows one system under one set of settings, so it measures what this workspace did and is not evidence that another design would behave differently. The full series, 26 frames on one sheet, is available as a link.

Six: the conversation model retrained to answer from live readings of the systems. Until October, the two halves of this system could disagree. The engines measured, and the language model spoke. When a person followed up on one of the system's own lines, with something like "how come you brought that up" or "why is your buffer full," the question fell past the page's handlers to a model that had never seen a reading, and it answered the way a chat model does, with plausible claims where nothing had been measured. The harvest that measured this gap on September 30th found a third of such questions reaching the model that way.

Conversation 5 (in service since October 3rd, 2026) was retrained from the same base model with the same two-stage recipe as Conversation 4, the previous generation, with a third kind of record added: questions about the system's own state, answered against the live readings that the page now sends with every request. Each answer was graded by a judge that sees the same readings and returns a named verdict on three checks: whether the reply stays within the readings or goes beyond them, whether it gives one reading or none, and whether it claims an action was done.

The judge itself had to be replaced first. The first judge tried called 539 of 544 falsified readings faithful, and the evaluation's own judge did no better, so the three reading checks went to a reasoning model, calibrated through the production path to keep 0.978 of faithful replies and catch 0.978 of falsified ones. Actions were deliberately not trained as actions. The model cannot forget, delete, or undo anything, so it is trained to say so and to give the words the application acts on; training it to say done would train it to report things that never happened.

On held-out questions, the model states a number that appears in the readings or the conversation in 99.7% of replies, says so when a reading is absent in 98.7%, and refrains from claiming an action that only the application can perform in 98.8%. It leaks none of its internal thoughts into a reply. Its claims stay within the readings in 85% of replies, with an interval too wide to decide against the 90% floor, because it refused 100 of 532 questions whose readings contained the answer: trained to refuse when a reading is absent, it learned that refusing is rewarded. A training round aimed at that defect traded it for the opposite one, so the first of the three Conversation 5 iterations remained the best. Where Conversation 5 measured no worse than Conversation 4, nothing is declared better: shelf agreement is 0.996, persona breaks are 0.27%, and its knowledge is not below its base model's. Its asks on Held Ground contradictions and near-misses are undecided because their intervals are wide. Cut replies are still past their ceiling, as Conversation 4's were, and the general win rate is below its floor at 0.42 but undecided by its interval, where Conversation 4's failed. All of this has gone into the plan for the next generation of the model.

Seven: the executive learns what its needs are worth, and what a person is worth. Every few seconds, the executive picks what to do next for the goal in front of it from a short list of ten kinds of action, such as searching its memory, storing something in it, reasoning something through, updating its picture of the world, or speaking. Until October, that choice was scored by a handwritten rule using values that had never been calibrated.

In October 2026, each of the two implementations was given two small decision models trained on its own record of decisions: one for the choice of action and one for which goal to pursue. What they learned to value was not typed in by anyone. It is relief of the system's own unmet needs (the same six drives the specification above certifies), each weighted by how badly it is unmet at that moment, so that easing a badly unmet need is worth more than topping up one that is nearly satisfied.

The animal-learning literature calls this incentive learning, and it is the first place in this research where the states choose behavior instead of only describing it. The records came from the running application: 84 twelve-minute sessions on one implementation and 60 on the other, totaling 4,706 and 3,364 decisions, with the live rule and a uniform random draw logged side by side, so that a trained model could be judged against what the rule would have done.

The decision point. The first round of data collection exposed a problem before any model was trained. Read back, the reward turned out to have been paying for inaction on one implementation and for failure on the other. A need was credited as relieved when an act completed, whether or not the act had done anything, so the cheapest act was a message that the system's own pacing refused most of the time; on the second implementation, an act that errored relieved more than one that ran. A model trained on either would have learned to do nothing, or to fail. Neither is real agency, and neither is generated text running in a loop with tools.

The fix was the oldest rule in Raffie, applied one layer deeper: only evidence satisfies a drive. A need is now credited only when the act's measured effect exists and bears on that need. Retrained under that rule, the first implementation's models improved need relief per decision with every band of the specification held. On the second implementation, the full set of learned outputs collapsed its drives: in the solitude test, where only the systems run and no user is present, the share of time in which at least two drives could form a goal fell from 86% to 40%. So only the goal model shipped there, and the contrast is a finding in itself. Drives whose dynamics are stipulated keep their voice whether or not their goals are pursued; drives derived from a model's own dynamics do not.

Then came the harder question, which was mine to answer and not the harness': what is a message to a person worth? In a test harness, nobody answers, so every message the system sent was pure cost, and a model trained on that would learn to stop speaking.

I chose to keep collecting data before accepting the models, and to change what counts. A message is now complete only when a person answers it, not when it is delivered; the reply is credited to the act that earned it; and a person simply talking to the system counts as contact that eases its pull toward connection. Getting the executive's own words into the conversation took two more days of repairs. Its lines had been going to a status strip that no one can reply to, and once they were routed through the initiative layer, all eighteen of its first attempts in the live checks were refused: eleven by the courtesy budget that its own status line had just closed in the same tick, and seven by the initiative engine's guard against bursts of messages.

On October 11th, for the first time on both implementations, a line the executive chose to say became a message, a reply answered it, and the credit landed on the need that had sent it. That evening, the learned models were switched on in both implementations, each serving what its own acceptance had licensed (every output on the first, the goal model alone on the second).

What it is a response to. Most agents hand their reasoning to a language model at every step, and when they speak of their own goals or needs, it is language without consequence. Here, consulting the language model is one of the ten acts, and the choice among them is made in milliseconds from the system's needs. That is a matter of economics as much as philosophy: the expensive part of an agent is the model call at every step, and a system that knows what it needs can save that call for when a need requires it.

The philosophy is the reply to Searle's Chinese Room described at the top of this case study: meaning comes from the states behind behavior, and a need that is measured, left unmet, and relieved only when a person answers is grounded in a way that fluent text is not.

This research and interpretability research are burning the same candle from opposite ends: interpretability finds emotion-shaped structure inside models trained on human text, and this research builds the states explicitly and measures what they do. Until October, I could only say that. Now I can say I acted on it: in this system, a need has a price, a choice is made by that price, and a reply from a person is what pays it.

None of this is evidence that anyone is home. It is evidence that behavior can be organized around meaning the system computes rather than meaning the language supplies. What the system has is access to its own states, which is the only kind of consciousness this research reaches for; it claims nothing about phenomenal experience, and it stays skeptical of its own results on purpose. At a time when some are ready to say a language model feels, that distinction matters for how we build and govern these systems: what a system says about itself must be checked against what it measurably does.

Asked what it just decided, the system composes its reply from its own record, with no language model involved, and leads with a visual card. The card shows the goal in focus; every option it weighed, with the need relief its learned values predicted and how many of its learned models favored it; the one it chose; its six needs as its executive read them at that moment (how unmet each was, and any change the act made); and what happened. The card near the top of this case study is that reply, captured from the running application. The sentence and the card report only what the record held at that moment. When no decision has been made yet, or the decision model returns an error, the system politely declines to answer.

On October 16th and 17th, 2026, the learned decision rule, refit each time on the system as it then stood, was measured against the decision rules of ACT-R, Soar-RL, and Psi, reimplemented in its own action space, and against LinUCB on the same inputs, in two comparisons registered before their data, each on 32 held-out sessions. It relieved more need than the three architectures in both, by 15.4–18.0% of its own gain over random choice with every 95% interval above zero; its advantage over LinUCB held in the first measurement and did not replicate in the second, and both are reported, the second as the result on the system as it stands. The comparison and its limits are in the figure that follows the decision trace.

After the Replication: Explorations Not Yet Judged

Two checks and two explorations followed the confirmed result. The checks were measured after the fact, so they were not registered. The explorations are running or in progress and not yet judged; both will be judged and validated in the final comparison against existing systems, after the system's next conversation model is trained and deployed.

Both checks used the replication's 32 held-out sessions, on October 18th, 2026. Against the best fixed strategy, which searches its memory whenever it can because that action did best in training, the learned rule relieved about 18% more need per decision (95% interval 10–28%, by self-normalized importance weighting), so its advantage comes from choosing by context. I also set eight versions of the rule side by side, greedy among them, ranging from optimistic to pessimistic about the values its learned models disagree on. The one that scored best on validation sessions, a mildly pessimistic version, was judged once on the held-out sessions and tied greedy (a difference of 0.00 thousandths of relief per decision, 95% interval −0.60 to +0.59). The optimistic and strongly pessimistic versions scored lower there, so I kept greedy.

The first exploration is learning what speaking is worth. Only a person can answer what Raffie says, so only a person's sessions can show what talking is worth; the scripted runs the rule learned from have no one to answer. Since October 19th, 2026, in sessions a person drives, the rule follows its own choice except on 13.6% of decisions, about one in seven, where it tries an option drawn evenly from those offered. That share is measured rather than set by hand: a draw gives up, on average, the rule's whole gain over random choice, and 13.6% is the largest share at which it still beats the best fixed strategy above with 97.5% confidence. Every decision records exactly how likely each option was, which allows an unbiased comparison against other systems on sessions with a person as well as on scripted ones, and the decision card says when a choice was a draw. In a 12-minute live check of that mode, all 56 decisions carried that record and 4 were draws (7.1%, 95% interval 0.4–13.9%, consistent with 13.6%). My most recent sessions will decide whether a new generation ships, and earlier ones will train it. The scripted runs, the registered benchmarks, and Study 2 are unchanged.

In the second, the rule reads the state of the global workspace, the thought holding center stage, when it values an action. The running system now carries that input; the rule uses it once a generation trained on it passes a quality check on held-out data, in the retraining after the next conversation model.

Two limits remain. The 97.5% bound was measured on scripted sessions, and whether exploring pays with a person is measured in the sessions that follow the next conversation model. Goal choice has no draws yet; it will get the same logged exploration this week, before those sessions.


What this does not claim. Asked directly whether it is conscious, this system answers that there is no little someone in here: what runs is a workspace holding attention, goals, and reactions, real and live, a simulation of cognitive activity rather than a mind that woke up. That answer is in the reel above. The drives, the valence, the arousal, and the trait drift are real, measured, and checkable. Raffie is a simulation of cognitive concepts, not a mind, built toward the original goal of a persistent system that complements the user and brings out the best in them; hence the name, from the Arabic for "to elevate." I leave the question of experience open. The pre-registered welfare study does not try to settle it: it tests whether grounded states leave measurable signatures, comparing this testbed, where affect is stipulated as explicit variables, with a second implementation in which the same states are derived from a model's own energy dynamics.

Why This Matters for Alignment and Welfare

Two independent lines of research have reached overlapping territory from opposite directions. This one has built the states explicitly, from cognitive science and philosophy of mind, since 2016. Interpretability research finds similar structure emerging inside transformers without anyone designing it in: emotion-like features, persona-like structure, and, in work published in July 2026, an emergent global workspace, a small, densely connected set of internal representations that the model can report, modulate, and reason with. That is the structure global workspace theory describes, and this architecture implements it by design. The paper closes on an open problem: the emergent workspace shows hints of ties to the model's sense of self, its emotion-like reactions, and its metacognition, "without exactly having worked out how." Those ties are exactly what this architecture makes explicit. The workspace, the emotional state variables, the trait model, and the metacognitive appraisal are designed and can be manipulated directly. The ablation above puts that property to use: one of those couplings, the term by which thoughts support one another, was switched off on the same build and the same days, and the workspace was measured with and without it. Whether emergent features like these amount to morally relevant states is a hard question that interpretability alone cannot answer. A system in which the equivalent states are designed as architectural objects offers a controlled testbed for studying how grounded states shape behavior, what signatures distinguish them from surface mimicry, and what welfare-relevant properties emerge when emotional and motivational systems are explicit. The two approaches complement each other: the emergent workspace and the architected one are two instruments pointed at the same question.

In October 2025, the same lab asked whether a model can report its own internal states at all. By injecting a concept into a model's activations and asking whether it noticed, the researchers found signs of introspection about a fifth of the time, which they describe as unreliable and limited. Conversation 5 approaches that question from the other side. Here the states are not inferred from the model. They are measured alongside it and sent to it with every request, and the model is trained and judged on whether what it says about itself matches what the instruments read, and on saying so when they read nothing. The welfare study's recorder logs the true affect series next to what the system reports, once a second, so the gap between the two is a measured quantity. In this architecture, whether a self-report is faithful can be checked.

Beside the two architectures that have defined this field for forty years, the difference is narrower than one would think, and more exact. The obvious objection is that both have modeled affect for decades. The objection is real, and it strengthens rather than weakens the claim. ACT-R and Soar both carry serious affect research as extensions: a physiology model that moves ACT-R's utilities and activations, valence and arousal as terms in its activation equation, and Scherer-style appraisal feeding Soar's reinforcement learning as reward. Each extension enters at one or two places in a choice mechanism the architecture already had, and neither architecture was built so that affect could be severed from control by configuration and the removal measured, because neither claims that affect is constitutive of its control. Here, the drive and emotion variables are the substrate the control loop reads at nine named seams covering action selection, attention, and learning, each with a switch, which is what made the lesion study possible to run and possible to fail. The Common Model of Cognition that Laird, Lebiere, and Rosenbloom proposed in 2017 names emotion and motivation as open work; this is one answer to that gap, built without the symbolic core the model assumes. The evidence is asymmetric: ACT-R and Soar are validated against human beings and fielded tasks, and this system, so far, only against itself, on one testbed, by one author, which is why a second implementation of the same states exists.

Raffie beside two older cognitive systems
 ACT-RSoarRaffie
QuestionHow does the human mind produce the timing, errors, and learning curves we measure, and where in the brain does each step happen?What fixed architecture supports general intelligent behavior (reactive, deliberate, and reflective), with learning across all of it?What observable signatures distinguish a system with grounded, architecturally explicit emotional states from one producing emotion-shaped output without them?
LineageJohn R. Anderson, Carnegie Mellon, 1976 onward: a theory of human cognition fitted to reaction times, errors, and brain imaging.Laird, Newell, and Rosenbloom, 1983 onward: a theory of general intelligent agents proven in deployed systems.One author: philosophy of mind from 2016, and the runtime architecture since August 2025. Two implementations by design, one stipulating affect as variables and the other deriving the same states from a model's own dynamics.
KnowledgeDeclarative chunks and production rules, with subsymbolic activation and utility on both.A working-memory graph, production rules, and operators; semantic and episodic stores since Soar 9.Explicit state variables (six drives with four properties each, eight primary emotions, allostatic baselines, and traits) alongside the weights of its neural models; no rule base.
One stepOne production fires per cycle of about 50 ms against the contents of a handful of buffers; the whole architecture is this loop.One operator per decision cycle over a globally visible working memory; a failure to decide opens a substate, and its solution is compiled into a new rule.No single cycle: a per-turn conversational pipeline, and concurrent engines whose loop runs drive to goal to choice to action to outcome to appraisal and back to drive, with a workspace competition deciding what is broadcast.
AffectNot in the core. Extensions couple a physiology model to utilities and activations, or add valence and arousal to the activation equation.Not in the kernel. An extension implements Scherer-style appraisal and feeds the feeling to reinforcement learning as reward.It's own system. Drive and emotion variables are read at nine named seams covering action selection, attention, and learning, each with a switch that severs the seam for a study.
WorkspaceNo broadcast; one chunk per buffer and one production per cycle are a central bottleneck.Working memory is visible to every rule, a blackboard; no ignition threshold, no adaptation of the winner.A competition with all-or-none ignition, a broadcast to the subscribing subsystems, and a ledger: occupancy 89% with the coalition term and 31% without, the diagram above.
Self-reportNone; a model reports what its task requires. Uncertainty itself is native and graded, fitted to human data for decades.None built in; an agent can inspect its own working memory, and missing knowledge is handled structurally, by impasse.Every self-initiated claim passes a warrant that declares what it asserts and fails closed; since October, the conversation model has been judged against the live readings. No calibrated confidence on an ordinary answer yet.
LearningIn the architecture, every cycle: activation, association, utility, and production compilation.Chunking at impasses, reinforcement learning on preferences, and episodic and semantic acquisition.State adaptation at runtime: baselines, satisfaction and thwarting, and slow, bounded trait drift. Model weights are fixed at runtime and trained offline. Since October 2026, that includes two decision models per implementation, trained on the system's own decision ledger to value relief of its needs and serving since October 11th within what each implementation's acceptance licensed; they have yet to be rescored on records of real use in which a person answers.
EvidenceHundreds of task fits to human latency and error, brain-imaging predictions, and classroom tutors.Fielded tasks: fixed-wing missions flown in a 1997 exercise, robots taught by instruction, and game agents.Itself: a registered lesion study and workspace ablations, on one testbed. No human data, which is the standing limit.

Public-facing precedent, and a deliberate line not crossed: The Ellsi voice assistant (described in the earlier work section below) was a 2016–2017 micro-scale application of the same approach: grounding interaction in an emotional origin rather than sentiment synthesis, to produce empathic behavior the user could trust. Its original goal was reducing cognitive load by interpreting the user's current mood and adapting to it. Raffie carries one element forward from Ellsi directly: expressed empathy for the user's feelings, with the system's own state adjusting in response. It deliberately does not carry forward the other: Ellsi's tool use that adjusted notifications and content to steer the user's mood. My Human-AI Collaboration Framework evaluates any influence on four dimensions: intent, timing, consent, and reversibility. By those measures, reading mood to reduce cognitive load is assistive, while acting on a person's emotional state to change it crosses into manipulation. Persuasion should support agency, not override it, so I chose to leave that capability out.

The third 2026 milestone completes that lineage. Anticipation was Ellsi's founding design goal: surfacing what a person needs before they ask for it. Raffie now has that capability, under the constraints my framework work argued for in the decade since. Initiation is off or minimal by default, adjustable in plain conversation, labeled whenever it happens, deletable like any other turn, and gated so that what the system volunteers is its actual state, never fluent phrasing generated to fill the moment. The capability Ellsi was built around is finally in the system, and the part of it that would have steered the user is still deliberately left out.

Research · Top HAT and Nested Hope · October 2025–present

Top HAT: Neural Architecture for Real-Time Hallucination Detection

Role
Independent AI Researcher and Architect, Persun AI
Core Innovation
Energy-based verification head with multi-scale factuality scoring, integrated with a Hebbian Associative Transformer.
Technologies
TensorFlow and Keras, custom neural architecture, energy-based models, contrastive learning, curriculum training, and mixed precision (A100-optimized).
Timeline
October 2025–present. The 220M arc is complete, and its 381M scale-up is built, rehearsed end to end, and ready to launch on a single GPU. In September 2026, the head was also tried on Raffie's Conversation 4 model and failed its pre-registered gate.

A 181M-parameter transformer with an energy head that scores its own outputs for factuality in the same forward pass, trained through a six-phase curriculum that moves from synthetic corruptions to the model's own hallucinations. It reached energy gaps of +10 to +13 against a target of 0.3. Then came a documented data-leakage flaw in the contrastive pipeline, an AUROC of 0.998 that did not survive prompt-level splitting, and a deliberately cheap experiment that ruled out the training objective as the bottleneck before any compute was spent on scale. In September 2026, the same head was trained on the answers of Conversation 4, Raffie's 1.5B conversation model, under a pass rule written before the first answer was drawn. It scored 0.58 against a 0.75 gate: a real signal, but too weak to serve, and a measured tradeoff between bolting a head onto a conversational model and training the base model for the signal from the start.

Full details: approach, six-phase curriculum, results, the v10 to Variant A arc, the leakage finding, the decisive experiment, and what is next

Motivation

The original Deep Thought architecture (HAT) showed that Hebbian memory could create auditable reasoning traces. But a fundamental limitation remained: language models can generate plausible-sounding text with no mechanism to check whether that text is correct. Earlier approaches to hallucination relied on uncertainty gates, multi-model consensus, or external fact-checkers, and each one adds latency, cost, and complexity.

Research Question

Can a unified architecture both generate text and score its own outputs for correctness, enabling "System 2" deliberation where the model actively verifies its reasoning rather than purely predicting the next token?

Approach: Dual-Head Architecture with Contrastive Energy Learning

Top HAT extends HAT with a Multi-Scale Energy Head that learns to assign scalar scores to text sequences: low scores for coherent, factually consistent text; high scores for corrupted, inconsistent, or hallucinated text. The energy signal is mechanistically separate from the language model's token probabilities. Unlike softmax confidence (which measures "how likely is this token given training data"), the energy head measures "how coherent is this complete thought."

Six-Phase Curriculum Training

The model learns verification through a curriculum that progresses from synthetic corruptions to the model's own hallucinations to human-judged quality. Each phase builds on the previous phase's signal quality:

  • Phase 1: Language Mastery (18 epochs). The energy head stays frozen while the model trains on language modeling until perplexity stabilizes. It reached a perplexity of about 22 with coherent <think>-tag reasoning.
  • Phase 2: Energy Introduction (10 epochs). The energy head is pretrained on its own, then trained jointly with the language model as the energy weight rises from 2% to 10%. It reached energy gaps of +10 to +13 between clean and corrupted text, against a target of more than 0.3.
  • Phase 3: Hard Negative Mining (6 epochs). Adversarial corruptions that fool the energy head are mined and used for continued training.
  • Phase 4: Self-Play Energy Calibration. The model learns to detect its own hallucinations as well as synthetic corruptions. Thousands of question-answer pairs are checked against ground truth and used for contrastive calibration, and ROC analysis sets statistically grounded decision boundaries.
  • Phase 5: Process Reward Model with Human Preferences. A lightweight process reward model (PRM) head trained on human preference data scores each reasoning step independently, which moves the system from detecting token-level errors to detecting failures of reasoning.
  • Phase 6: Iterative Self-Play with Best-of-N. Each self-play round generates N candidates per prompt, selects the best by combined PRM and energy scoring, and retrains on the improved data.

The contrastive signal is generated from 11 corruption strategies spanning surface-level (token replacement, span corruption), semantic (negation insertion, entity swapping), and factual (number perturbation, temporal shifts, fact inversion) perturbations.

Key Results from First Full Training Run

Energy gaps between clean and corrupted text far exceeded the target. Energy rises steadily with the level of corruption (10%–30% corruption produces gaps of 0.8–2.5), which shows that the head tracks how corrupted a text is; whether that carries over to the model's own factual errors is the question the later experiments test. The model kept its structured reasoning, with explicit thinking traces, while performing energy-based verification, so joint training did not cost it its language ability.

Note on the training arc: The first full training run (v6) showed self-verification and hallucination detection working on synthetic corruptions, but thinking traces diluted its language generation too much. The line continued as v10 and then as Variant A of Nested Hope on continuous A100 infrastructure; the final results and methodological findings appear in the experiment conclusion below.

Inference Modes

The trained model supports several inference-time generation strategies that use real-time energy scores: Best-of-N rejection sampling (generates N candidates at varied temperatures, ranks them by combined PRM and energy score, and returns the best); PRM-guided generation with per-step confidence scores for fine-grained auditability; energy-guided beam search that prunes high-energy candidates during generation; iterative self-correction that triggers regeneration when energy exceeds calibrated thresholds; and real-time confidence display that exposes energy scores to users as interpretable confidence indicators.

Why This Matters for AI Safety and Welfare Research

Top HAT addresses one dimension of model trustworthiness: factuality. Its method, energy-based discrimination between grounded and ungrounded representations, generalizes. Applied at a different level of the architecture, to emotional or motivational states, the same approach is what connects this work to the cognitive architecture research described above. Systems in which uncertainty is built in explicitly make controlled testbeds for an alignment question: do grounded internal states produce properties (honesty, calibrated uncertainty, and behavioral coherence) that post hoc steering cannot reliably achieve?

Implementation Built

The complete Frontier v6 architecture (a modern transformer with GQA, RoPE, SwiGLU, and RMSNorm; 181M parameters); a multi-scale energy head (spectral-normalized, with six-head attention aggregation; 2.9M parameters); six-phase curriculum training; competitive Hebbian memory with sparse, auditable association patterns; and a production inference server with multiple generation modes.

Experiment Conclusion: v10 to Variant A of Nested Hope (May 2026)

Top HAT v6 was limited by the instability of Google Colab Pro+. I continued the line on continuous A100 GPU infrastructure as v10 (181M parameters, combining the Hebbian memory layer, a gated Compressive Memory System, the energy head scoring factual versus hallucinated completions, and a process reward model head). It reached a best Phase 1 validation loss of 3.77 and a Phase 4d hallucination self-detection AUROC of 0.516.

To target v10's failure modes, I designed Variant A of Nested Hope, a successor architecture with three specific interventions: removing register attractors in a redesigned 1B-token training corpus, a redesigned gated Continuum Memory System that no longer fires unconditionally on every forward pass, and the energy head and PRM stack carried forward from v10. At 220M parameters, Variant A reached a best Phase 1 validation loss of 3.07 (perplexity fell from 43.6 to 21.6 on the same evaluation, a reduction of about 51%) and a Phase 4d AUROC of 0.586, up from v10's 0.516, a gain of 0.070 under the identical Phase 4 contrastive training method with a clean train/holdout split. A stronger backbone produced measurably better hallucination self-detection.

Methodological Finding: Data Leakage in Phase 6 Contrastive Pair Construction

A second finding came out of the Phase 6 investigation. v10's Phase 6 algorithm (combinatorial contrastive pairs built from pools, followed by an 80/20 split at the pair level) leaks data when the pool of correct answers is small. On the same model and data, pair-level splitting produced an AUROC of 0.998, while prompt-level splitting (the methodologically clean fix) produced 0.530, a difference of 0.468 caused entirely by how the data was split. v10's previously published Phase 6 AUROC of 0.588 probably contains similar but less severe leakage, though its size cannot be measured without rerunning v10's pipeline with a prompt-level holdout. The finding generalizes: any contrastive hallucination-detection benchmark that builds pairs combinatorially should split by prompt, not by pair, to avoid the same inflation. This is the more durable contribution of the experiment.

The Decisive Experiment: Was It the Method, or the Model? (June 2026)

Instead of spending a large compute budget in the hope that a bigger model would lift the AUROC ceiling, I designed a cheap experiment to find out first why it was stuck: was the bottleneck my training objective or the model's learned representations? I trained the energy head two ways on the frozen backbone, with identical cached features and an identical held-out split: once with the original margin loss, and once with a variance-aware surrogate that directly optimizes the ranking metric (a smooth Wilcoxon–Mann–Whitney statistic plus a within-class variance penalty). The new objective did exactly what the math predicted and collapsed the within-class spread, yet both versions scored the same: at chance. The objective was not the bottleneck. The result is clean because it is a relative comparison on identical inputs, so it holds regardless of how the evaluation set was built.

What the experiment does not yet settle is whether the model's representations are the limit. By design it ran on constructed true/false pairs (a controlled probe, never the final verdict), and the model committed to an answer so rarely that a test on its own generations could have been measuring hesitation instead of factual error. Telling "the model can't represent factuality" apart from "the test never isolated it" requires rerunning on the model's own harvested outputs with a stronger backbone, which is the test the scale-up below is built to run. The experiment could have ended this line of research. Because it came back clean, I know what to change next.

What's Next: The Scale-Up Test, Ready to Launch

With the objective ruled out, the step the experiment points to is a stronger, more factually grounded backbone. I built a 381M-parameter scale-up of Variant A and a 4-billion-token corpus rebalanced for factual density (the earlier corpus taught the model register but not facts). The corpus is split by document, so validation never scores text the model trained on, and any document that shares a 13-token window with a question the model will later be tested on is dropped. The pipeline is built, rehearsed end to end, and ready to launch on a single GPU. Its hallucination-detection AUROC, measured on the model's own generations, will finally separate a ceiling in the representations from a ceiling in the data.

How the verdict is measured. The model answers 1,000 reserved questions eight times each, sampled exactly as it will be served, and a strict checker grades the answer each reply commits to. The energy head trains on some of those questions and is judged only on the rest. Only questions the model answered both right and wrong enter the comparison, each counts once however many answers it drew, and the 95% interval resamples whole questions, since answers to the same question are not independent. The pass rule was written down before this model produced a single answer: an AUROC of at least 0.75, an interval clear of chance, an energy that does not track answer length, and an energy that beats surface features by outscoring the best of length, repetition, and truncation by an interval clear of zero. That last clause is there because a head that reads nothing but length would pass every other test.

How the model is trained for it. The energy head is first fit on the frozen language model, which measures what the representation already holds before anything else changes. It reads mid-depth layers as well as the last one, pooled over the answer being judged. Its largest source of negatives is the model's own wrong answers, each paired with a right answer to the same question, alongside replaced facts, fluent distractors, and negations. The model then learns the chat format it is served in, and a preference stage trains it on its own right and wrong answers; that stage is kept only if held-out accuracy rises while completion, repetition, and the energy signal hold steady. At serving time, the regeneration threshold is set as a budget, so that about one right answer in ten triggers an unnecessary retry, and the confidence shown is calibrated to how often answers are actually wrong.

Where the Research Goes Next

Top HAT and its successors are AI safety research focused on one narrow problem: architectural mechanisms that let a model detect its own hallucinations. The line is also one component of the broader Raffie cognitive architecture research described in the earlier card. Going forward, the main research focus shifts to Raffie at scale: the same philosophy of grounded internal states with measurable dynamics, applied beyond factuality to emotion, motivation, and metacognitive uncertainty. The hallucination-detection thread continues with the Variant A scale-up, which is ready to launch and which the chat-model result above has narrowed: it has to show a signal that chat-trained representations at 1.5B and 7B parameters did not carry. Variant B (Self-Modifying Titans) waits on that verdict and on a hardware test of its single-GPU configuration. The Raffie research continues in parallel.

Open to discussion with researchers or organizations interested in extending the Variant A line, validating the Phase 6 leakage finding on adjacent benchmarks, or collaborating on Raffie at scale.

"'Sounds right' and 'is right' are different properties, and the entire current interface paradigm is built to make them feel the same. Top HAT's energy head is built to give a mechanistic signal for telling them apart."

Published frameworks · haicf.com and cognitive.link · 2025–2026

Two published frameworks for evaluating the human-AI coupling

The Human-AI Collaboration Framework (2025) operationalizes transparency, agency, and collective input as evaluation criteria at the experience layer, with six case studies and implementation guidance. Cognitive Link (2026) cross-maps four decades of cognitive science and human-automation research onto agentic AI, arguing that evaluation must descend Marr's levels, that the agentic design problem is placing friction rather than removing it, and that the three pillars are properties of the coupling, not the model.

Cognitive Link (June 2026): the full summary

Agentic AI Evaluation · Cognitive Link: A Coupled Framework for Agentic AI (cognitive.link)

Published: June 2026

The dominant story about agentic AI is wrong twice. It treats the agent as a smarter chatbot, when an agent is an automated decision system that aviation, medicine, and process control already learned to evaluate. And it treats the model as the meaningful object of design, when the actual cognitive work (the deciding, the acting, and the consequences) happens in a coupled system: human plus agent plus world, behaving as one extended mind whose properties belong to none of its parts alone.

Cognitive Link cross-maps two literatures the agentic moment has forced into the same room. From cognitive science: Dawson's three traditions, Marr's levels of analysis, Clark and Chalmers on the extended mind, and Hutchins on distributed cognition. From human-automation research: Bainbridge's ironies of automation, Sheridan and Parasuraman on levels of automation, Endsley on situation awareness, and Mosier and Skitka on automation bias. The mapping is not analogical; the phenomena are the same phenomena.

From the cross-mapping, three reframings follow:

  • Evaluation must descend Marr's levels. Outcome alone certifies systems that reach right answers through unsound trajectories, so agentic evaluation has to assess process and coupling, not just endpoint success.
  • Design must invert. The historical goal was removing friction; the agentic goal is placing it, as selective, stakes-calibrated cognitive forcing functions, with constraints compiled in rather than guardrails bolted on.
  • The three pillars (transparency, agency, and collective input) are properties of the coupling, not the model. That is the deep reason model alignment alone never suffices. The field has converged on the instinct that "the model is no longer the product; the harness is," and Cognitive Link makes that instinct rigorous.

Cognitive Link extends the line of work begun in the Human-AI Collaboration Framework by grounding its three operational pillars in the cognitive science traditions that mapped this terrain decades ago. The framework is also informed by my broader research on grounded cognitive architectures (see the cognitive architecture card above), though that research is intentionally kept out of the public framework, which stays focused on the published literature.

The one-line version: build the human and the agent to think better together than either could alone—which sometimes means getting out of the way, sometimes means standing in it, and always means knowing, and measuring, which.

From framework to engineering: a companion implementation roadmap turns the framework's signature mechanism, confidence-calibrated selective friction, into something you can build, measure, and falsify, anchored on one chokepoint: a calibrated, real-time signal of the machine's own uncertainty. Its first decisive experiment has already been run. A cheap, frozen-backbone test ruled out the training objective as the bottleneck for that signal, and the larger-backbone test is built and ready to launch (the engineering details are in the Top HAT card above).

For different reader profiles: The full framework · Implementation roadmap (engineering spine) · Stakeholder brief (nontechnical) · Key takeaways

The Human-AI Collaboration Framework (2025): the full summary

Evaluating AI · The Human-AI Collaboration Framework (haicf.com)

Published: Summer 2025

Most AI alignment research targets the model layer. Most AI product design inherits pre-AI guidelines. The Human-AI Collaboration Framework bridges that gap by translating alignment theory into evaluation criteria, interaction patterns, and team workflows that operate at the product layer, where users actually meet AI.

What does it mean for AI to be "beneficial"? The term is deceptively simple, so the framework defines it: beneficial AI is AI that supports human understanding, preserves autonomy, and promotes collective well-being. It is aligned not only in its outputs, but in its relationship to the humans it serves, working with their motivation in a complementary and collaborative manner.

The framework rests on three operational pillars:

  • Transparency: Legibility (can I understand it?), traceability (can I verify it?), and contestability (can I challenge it?), designed at the experience layer rather than buried in documentation.
  • Agency: User control across interaction, personalization, privacy, and framing. Steerability is the embodiment of human-in-the-loop design in everyday use, not just in training.
  • Collective Input: Participatory methods, inclusive data sourcing, and accountability mechanisms that allow for post-deployment feedback and correction. Following the adage of Inclusive Design, designing for the edge is designing for the whole.

The published framework analyzes six case studies: Be My Eyes + GPT-4, Google's Magic Editor, Airbnb's Fairness Dashboard, Auto-GPT and ChatGPT Agent, Meta AI in Messenger and Instagram, and a custom AI assistant (Ellsi). Each case study is examined for how human-centered design supports or fails ethical AI implementation.

It also engages directly with current alignment research, including Anthropic's work on alignment faking (a model selectively complying with its training objective while it believes it is being trained, in order to protect its existing preferences). The framework's argument is that without real-time steerability at the product layer, users are at the mercy of static outputs from systems that cannot be corrected or contested.

The framework includes specific UI patterns (inline explainers, counterfactual examples, model cards, consent-aware onboarding, progressive disclosure), evaluation metrics, and implementation guidance for teams building AI products.

Insight: In leading strategy and fostering successful stakeholder relationships, one must practice great communication because frameworks for AI product design can become unwieldy fast. While the full framework has executive summaries and tailored takeaways, insights should be calibrated to their audience. I invite you to check out the Getting Started with AI fairness one-pager I derived from the larger framework for nontechnical stakeholders, at https://haicf.com/fairness. It is easier to digest, practice, learn from, and share, and it gives teams a concise reference for what fair, beneficial AI looks like in practice.

Related published work: The Human-AI Collaboration Framework · Getting Started with AI Fairness

The conversation model today, and where this line of work started: philosophy of mind, the memory-native model, a collaborative assistant, and earlier voice work

Foundation · Michigan State University · Spring 2016

Philosophical Foundation: Graduate-Level Research in Philosophy of Mind

Supervisor
Dr. Robyn Bluhm, Professor, Department of Philosophy and Lyman Briggs College, Michigan State University
Timeline
Spring 2016 (independent study)

The thread running through my AI work began in 2016 with a graduate-level independent study in philosophy of mind under Dr. Robyn Bluhm. The course covered artificial intelligence, philosophy of mind, ethics, neuroscience, and evolutionary biology, and produced three argumentative essays examining what consciousness is, how it relates to emotion and behavior, and what conditions would allow it to be re-created in technology.

The central arguments these essays made:

  • On Emotions, Their Purpose, and Their Influence on Behavior argues that emotions are evolutionarily refined signals, far from irrational noise: they filter environmental stimuli, predict outcomes against goals, and direct behavior toward survival and flourishing. Drawing on LeDoux and Damasio, the essay presents feelings as the means by which any organism (and potentially a designed system) becomes aware of its environment and acts in it.
  • On the Nature of Consciousness and the Re-creation Thereof uses Dennett's continuum of intentionality and Searle's Chinese Room as the foundation for a constructive argument: consciousness exists on a continuum across varied forms (humans, octopuses, and submarines all swim, each in its own way), and Searle's requirement of "meaning" can be met through grounded emotional states rather than language alone.
  • On Affording Meaningfulness in Design Mature and Beneficial AI Experiences carries both arguments into AI: artificial emotion is more than sentiment synthesis (surface-level linguistic mimicry), and a designed, grounded emotional system is a different kind of thing from one that only produces emotion-shaped output. The distinction bears directly on welfare research into today's frontier models.

This philosophical grounding informs everything that came after: the emotional intelligence framing in Ellsi, the transparency and agency pillars in haicf.com, the idea of factuality as a grounded signal behind Top HAT, and the cognitive architecture work that followed.

Available on request: Reference letter from Dr. Bluhm; full essays.

Research · Memory-native conversational layer · July 2025–present

Memory-Native AI System: Conversational Model for the Broader Cognitive Architecture

Role
AI System Architect and Researcher
Technologies
TensorFlow.js, Python, JavaScript, custom neural networks
Timeline
July 2025–present (experimental research project)

Context

These systems were the early stages of the conversational and memory-native components of the broader cognitive architecture research described above. Where Raffie surfaces grounded internal states (valence, arousal, drive dynamics, and satiation), this layer provides the language-facing and memory infrastructure those states need to interact with users statefully.

Challenge

Most conversational AI systems are stateless: they don't remember previous interactions or build context across them. They also tend to use one large monolithic model where any one decision is opaque. I wanted to explore whether multi-model orchestration with explicit memory consolidation, uncertainty quantification, and human-centered transparency could produce coherent, personalized conversation at a fraction of the computational footprint of monolithic alternatives, and whether the orchestration could be auditable by design.

Approach

Architected an experimental system integrating 20 cross-communicating models with intelligent routing, uncertainty gates, and memory consolidation. The system explores how episodic memory patterns and multi-model consensus can improve conversation coherence, personalization, and reliability while keeping the orchestration legible to both users and reviewers.

Multi-Transformer Orchestration

The conversational layer alone uses four custom transformers that work together:

  • Conversation 2 Model: The primary encoder-decoder transformer (~108M parameters), which handled dialogue generation and was trained on curated conversational data. It was replaced by Conversation 4 in September 2026 and by Conversation 5 in October, both described below; Conversation 5's clips are in the research card above.
  • Significance Scorer: A decoder-only transformer (~27M parameters) that evaluates input importance and routes queries appropriately, trained on ~500MB of tailored data.
  • Novelty Detector: A decoder-only transformer (~9M parameters) identifying new information patterns and knowledge gaps, trained on ~500MB of tailored data.

    Both the Significance Scorer and the Novelty Detector serve two purposes: classification, and language generation when the orchestration calls for it. Their outputs work as feature vectors in which the highest-activation indices map to semantic categories that inform routing decisions. Sixteen additional specialized models handle supporting functions, including sentiment analysis, intent classification, memory consolidation, and cognitive extraction.

  • Deep Thought: A reasoning transformer (~180M parameters, trained on gigabytes of complex reasoning data) for queries requiring extended deliberation.

Memory-Native Technical Implementation (pre-Conversation 4)

Architecture: conversational coherence in under 2GB of total model size; server-side deployment for complex transformers; browser-deployable supplemental models for edge processing; robust orchestration without excessive technical overhead.

Hallucination prevention and uncertainty quantification: multi-model consensus before response generation; uncertainty gates that flag low-confidence outputs; significance scoring to validate input understanding; novelty detection to identify knowledge boundaries; graceful fallback strategies across all models.

Human-centered design: transparency in how memory decisions are made; user control over what is stored or forgotten; clear confidence indicators for AI-generated responses; privacy-preserving local processing.

Key Learnings

Thoughtful system architecture (intelligent routing, uncertainty quantification, memory management) can produce coherent AI interactions while preserving user control. The multi-transformer approach shows that specialized models working together can achieve conversational quality comparable to monolithic systems at a fraction of the computational cost. This research provided the infrastructure for the broader cognitive architecture work: a stateful, auditable language layer for the grounded internal states to interact with.

Framework Connection

The human-centered design principles operationalize my Human-AI Collaboration Framework: transparency in system behavior, user agency over personalization, and collective input, for people-first, purpose-oriented design driven by user needs as well as technical capability.

The Conversation Model Today: Conversation 4, Then 5

From September 26th, 2026, until October 3rd, Raffie's conversation model was Conversation 4: a 1.5B-parameter instruction model fine-tuned in two stages, first supervised and then on preferences, for Raffie's voice and for the Held Ground shelf. Its training data included 73,000 synthesized Held Ground exchanges walked through the real page and graded turn by turn, 8,000 preference pairs, and about 290,000 general conversations, most drawn from public datasets and some re-answered by a larger open model. It replaced Conversation 2, the original encoder-decoder that the earlier layer ran on.

Conversation 5 replaced it on October 3rd, 2026: the same base model, retrained with the same recipe plus the architecture's live readings, so that what it says about its own state is read from the systems. Its two clips are in the research card above.

Prototype · Collaborative AI assistant · July 2025

Collaborative AI Assistant

Role
System Architect, Designer, Developer
Technologies
LLMs, TensorFlow.js, HTML, CSS, and JavaScript
Timeline
July 2025 (iterative prototype development)

Challenge

Explore how to build a conversational AI system with memory, personalization, and ethical constraints using rapid prototyping methods.

Approach

I developed this prototype over three iterations with the pattern my Human-AI Collaboration Framework calls Architect–Engineer Decomposition: the human is the systems architect, and the LLM is the engineer that generates code and structures logic. I defined the why and the what, the LLM filled in the how, and every layer stayed open to my iteration and refinement.
I designed the human-AI collaboration patterns up front, then used prompting and iteration to build and evolve the assistant in three phases:

Iteration 1: Core Functionality

  • Integrated TensorFlow.js sentiment analysis and question classification
  • Built routing logic for different query types
  • Implemented contextual content delivery
  • Established baseline interaction patterns

Iteration 2: Memory and Adaptation

  • Added conversation context management
  • Built content summarization and progressive disclosure UI
  • Created editable user preferences and timeline views
  • Implemented suggestion system for user input

Iteration 3: Enhanced Memory System

  • Expanded conversation model training data
  • Built semantic memory with vector-based recall
  • Created memory transparency interface (editable preferences, interaction logs)
  • Implemented graceful fallback modes for all features

Human-Centered Design Principles:

  • Transparency in what the system remembers and how it responds
  • User control over memory, preferences, and behavior
  • Privacy-aware personalization with opt-out controls
  • Ethical constraints embedded in system logic

Key Learnings:

The pattern lets a person keep strategic agency while using an LLM for fast implementation, so the system architecture and the human-centered design principles could be tested quickly. The project applies principles from my Human-AI Collaboration Framework: transparency in system behavior, user agency over personalization, and design driven by user needs rather than just technical capabilities.

Outcome:

A prototype AI assistant with editable memory, opt-out personalization, and ethical constraints encoded into the orchestration logic. The project operationalizes the three pillars of my Human-AI Collaboration Framework:

  • Transparency: System shows what it remembers and how it behaves
  • Agency: User has full control over memory, tone, and behavior
  • Collective Input: Design driven by context and cultural care, not just technical benchmarks
A two-minute demonstration of Iteration 1 showing real-time sentiment analysis, question classification, contextual recommendations, and cultural adaptation.

Key Insight

Effective human-AI collaboration is not about building the most advanced AI, but about orchestrating several AI capabilities to serve specific human needs while the human keeps oversight. Here, that meant the right architecture, clear memory boundaries, and co-design with purpose.

Earlier explorations in human-AI experience

Challenge

In our research, notifications felt like stress, work, and labor to the people we studied, when what they expected from technology was relief and entertainment. How might we design a voice-driven system that respects the user's emotional state, supports their goals, and delivers content with empathy rather than addiction?

Solution

I led the research, design, and prototyping of Ellsi, an emotionally intelligent voice assistant designed to support intentionality, reduce information overload, and preserve psychological clarity in how we engage with notifications.

  • Cross-Disciplinary User Research: Led a team of students through 50+ contextual inquiries and interviews, measuring participants' heart rates, with their consent, as expected and unexpected notifications arrived. Heart rates rose with unexpected notifications, especially when they came in excess.
  • Philosophy-Backed Cognitive Strategy: Researched and published frameworks for building AI experiences around genuine empathy.
  • Voice & Visual Prototyping: Delivered two high-fidelity prototypes after 100+ iterations integrating speech recognition and emotional simulation.
  • Voice Persona & Brand System: Developed 13 audio cues and dialogue flows rooted in empathy and clarity.
  • Hybrid UI: Engineered a visual interface for nonverbal control and error recovery when the user's attention is divided.
  • Transparency and Control: Notifications were batched to the user's rhythm, an "empathy setting" let users adjust her communication style, and a diagnostic view showed how their preferences shaped her behavior. Rather than acting as a black box, Ellsi surfaced the logic behind her decisions.

Design Philosophy

Ellsi grew out of wanting to serve people: to make their interaction with technology people-first by understanding their goals and helping them accomplish them.

  • Empathy as Core Function: Every notification is reframed as a contextual offer aligned with user goals.
  • Human-Centered, Not User-Centered: The people Ellsi serves are people, not "users," so she listens for intent and responds with meaning.
  • Emotional Intelligence over Behavioral Hacking: Designed from evolutionary theories of natural sentience and drive.
  • Voice as Relationship, Not Interface: Crafted a consistent, calming, and supportive persona.
  • Systems-Level Ethical Design: Proposed a future of AI that is compassionate and innovative.

Conversational Flow Design & Analysis

Blended Intelligence: Expanding Voice through Hybrid Interaction

User: "Ellsi, add turmeric, cardamom, and chili powder to my shopping list."

Ellsi: "I've added those three for you."

[Displays shopping list with three newest items on top]

Ellsi: "Would you like me to remind you when you're at a particular store?"

User: "Yes, that would be so helpful!"

Ellsi: "Which store? Or search for it in the following card:"

[Displays map card]

Path A (Voice Response)

User: "Fresh Thyme on Grandview Ave."

Ellsi: "I'm setting the location reminder."

Ellsi: "I found a closer location you've visited before. Would you like directions and a reminder there instead?"

User: "No, thank you. I'm heading to that part of town."

Ellsi: "Sounds good!"

Path B (Visual Selection)

[Displays list of closest matching results]

[User selects location]

Ellsi: "I just set the reminder."

This moment introduces a multimodal system that also works hands-free and eyes-free, going beyond traditional voice UX. Because she blends visual and conversational interaction, Ellsi can meet people where they are, whether they know exactly what they want or are still deciding, and help them act without friction.

Because the user is signed in, Ellsi gathers input flexibly and completes tasks in context, using natural language understanding and past interactions. Instead of forcing a rigid structure, she fills in the blanks from what she knows about the person's preferences, products, and habits across devices.

Contextual Intelligence: Meeting Users Where They Are

User: "Set a reminder for Karmen's exam on Wednesday to send her the motivational message I wrote last week for my friends."

Ellsi: "Please select which note you'd like me to remind you with."

[Displays Notes card]

[User selects note]

Ellsi: "Alright, I'll remind you."

Here, Ellsi adapts by recognizing a past habit: she knows how to find saved notes and send them to frequent contacts. Small actions like this one show what context-aware design can do in voice UX.

Context is paramount in intelligent voice experiences, because it lets Ellsi recognize new problems to solve in the moment and meet users in their mental models, where they are. By keeping up with the user's mental model in real time, she handles needs as they come up and earns trust along the way.

Designing for the Margins of Attention

These version 4 designs show how Ellsi adapts when the user is on the move, and how voice and ambient intelligence fit into real life when a screen is out of reach but people still need to think, feel, and act.

Ellsi's hands-free interface for driving or when the user is otherwise occupied
Driving and Hands-Free Mode Large text and focused actions help users stay on task while their hands are busy. The always-listening interface is easy to use and just as easy to dismiss, which keeps interaction safe on the road.
Ellsi's interface after a drive ends
Post-Drive Transition When the drive ends, Ellsi switches modes, stops listening, and invites the user back to touch. A flexible layout places the controls on whichever side of the vehicle the user sits, with quick options to resume or exit. Users can say a hotword or tap to preview messages and return to the full hybrid interface.

Tone Matching: Personality That Listens Before It Speaks

User: "What did Mateo email me?"

[User is driving, so Ellsi reads it aloud instead of showing a preview card]

Ellsi: "Mateo said, 'Sprint 3 completed on time and we're on track to target! I'd like to know what you think before I share with any others.' There is an attachment."

User: "Thank you, Ellsi!"

Ellsi: "You're welcome! That sounds like good news. I'll add a snapshot to your moments view the next time you're on your device."

[Ellsi updates the moments view with a visual summary of the email and attachment]

These flows show Ellsi adjusting her tone to the user's approach: sometimes warm and conversational, other times clear and directive. Her personality is not meant to anthropomorphize the interface, but to meet the user's needs at their convenience, so it changes with the moment.

Adaptation depends on:

  • The interface's role: assistant vs. automation
  • The brand's voice and experience goals
  • The user's tone and communication norms
Resilience by Design: Repairing with Context and Care

User: "Start a 10-minute timer called check live..."

[Ellsi waits briefly, detecting pause]

[Plays error chime]

Ellsi: "I don't think you finished your last request."

User: "Oh, sorry, I was distracted! Set the timer for 'check the live blog.'"

Ellsi: "10 minutes and counting!"

Ellsi handles confusion with graceful recovery, using on-device memory to protect privacy while learning from past interactions. When a request is incomplete:

  • The first time, she reprompts quickly with a gentle nudge.
  • If it happens again, she reframes the prompt with detail, spelling out exactly what is missing.

Because she stays aware of the surrounding context and the user's usual patterns, Ellsi keeps the conversation moving with empathy, efficiency, and patience for distractions.

Login-Optimized, Not Login-Limited

User: "What's the weather like near Kiko?"

Ellsi: "I found a previously shared current location in Maine."

Ellsi: "The weather there is chilly for this time of year, but it's expected to warm up later this week. I'm displaying the forecast:"

[Displays forecast card]

User: "Message her: 'I heard it's cold there! But not too cool to be cool. You got this [dash]stay warm!'"

Ellsi: "Message sent!"

Ellsi uses the customer's login to prefill essential details, such as contact names, which reduces friction when something goes wrong. She works without a login, but the experience is smarter and smoother when the user is signed in.

Across these flows, Ellsi handles varied user responses with natural language understanding, filling in the blanks through a statistical grammar that follows how people actually talk.

Deliverables

6 Steps to Exceptional Voice Interface Design
Step Preview What How
1. Requirements Definition A VUI worth building starts with why it should exist.Define the business motivation and align with user and application goals.
  • Use Reverse Brainstorming to identify potential roadblocks in voice UX, mapping out worst-case scenarios and addressing them early.
  • Conduct Jobs to Be Done (JTBD) interviews to frame the user's true goals and align business objectives with voice UX solutions.
2. High-Level Design Set the stage for smooth interaction in the brand's voice. Establish dialogue strategy, grammar type, and persona alignment.
  • Run a Lightning Decision Jam to collaboratively outline the voice persona, tone, and key dialogue flows.
  • Use Crazy Eights to quickly ideate on dialogue strategies, grammar types, and branding nuances in voice interactions.
3. Detailed Design Focus on research-backed scenarios to ensure user satisfaction. Craft precise dialogues and tailored prompts for all use cases.
  • Develop Storyboard Scenarios to showcase user journeys and align dialogue flows with real-world context.
  • Apply Dot Voting with stakeholders to prioritize designs based on user feedback and research findings.
  • Incorporate Wizard of Oz Testing, where study participants interact with a simulated voice system, to catch usability issues early.
4. Development Turn the blueprint into a working system. Integrate coding practices with front-end and back-end systems. Build iteratively with stakeholders under a shared vision.
  • Apply Design System Thinking to align voice UX components with brand-wide standards, so the experience stays consistent wherever customers meet the brand.
  • Use Parallel Prototyping to compare and contrast different design approaches simultaneously and choose the most effective ones.
5. Testing Thorough testing ensures system reliability and trust. Make sure the system works reliably at scale, in real-world conditions, for everyone it is meant to serve.
  • Conduct 5-Second Tests for first impressions of voice prompts and phrasing.
  • Run A/B Testing to compare error-handling strategies and voice tones.
  • Use Heuristic Evaluations focused on conversational flow and user empathy.
6. Tuning Continuously evolve the VUI with the user in mind. Optimize grammar, accuracy, and all-around satisfaction through feedback. Identify high-traffic patterns to inform strategy.
  • Build Continuous Discovery Habits through post-launch user feedback.
  • Apply Affinity Mapping to group user feedback into themes that guide the next round of design.
  • Use Task Analysis to measure how effectively users accomplish their goals with the voice assistant and refine the UX accordingly.

Conversation flow diagram

Interface and Interaction Highlights

Three initial sketches of the Ellsi interface, folded like a map
Initial sketches of the Ellsi interface, adapting to the AI technology of the time
A version 1 interface screenshot of a conversational Ellsi prototype
The interface and interactions of a conversational Ellsi prototype from 2016
Version 4 interface preview of a conversational Ellsi that shows a human-centered call-to-action button
Designing with the Body in Mind

Standard UI places calls to action in the bottom right, assuming a right-handed user. Our research showed that this position strains the thumb's natural arc, so I designed a flexible, elliptical button that adapts to each user's relaxed reach, left or right, and keeps the main action easy to hit with either hand.

Version 3 interface preview of a conversational Ellsi that shows the moment view that first loads when opening Ellsi
Meeting the Moment

The experience opens with a personalized "Moment" view that looks calm on the surface and is carefully designed underneath. Drawing on user cues and contextual research, it highlights what matters most right now and offers three gentle prompts to start from. The microphone, or a universal hotword, switches to a voice-only interface for the times when typing is not practical.

Version 3 interface preview of a conversational Ellsi that shows the conversational interface with interactive cards in conversation
Conversational by Design

Conversation sits at the center of the multimodal experience. We pair natural language understanding with best practices in conversational UX, card systems, and visual design so that dialogue, context, and action carry over smoothly between the app and the device.

Version 3 interface preview of the interactive weather interface
Weather, on Your Terms

Expanding the introductory weather card opens the Today view, where users can scroll through the day, confirm or adjust preferences, and take action, all from a conversational, context-aware interface.

Version 3 mail interface preview with contextual design elements
Contextual Notifications, Made Clear

From conversation mode, users can reach email and other alerts through two distinct views. Smart chips filter by type, such as payments or packages, while badges verify senders, flag priorities, and point out followed threads, so users can tell at a glance what matters and whom to trust.

Chapter 02

Accessibility

Equity cannot be retrofitted into systems; it must be designed in from the beginning. This chapter covers accessibility evaluation frameworks and implementation for high-stakes systems that serve vulnerable people, from a regional airport's website to the Title IX reporting form survivors used to begin pursuing justice: more than 100 in the lawsuit in 2017, and roughly 150 at Larry Nassar's sentencing. In each, I treat accessibility as harm reduction—moving from compliance to care.

Bishop International Airport, at Addis Enterprises · Accessibility and UX lead, front-end contributor

Accessibility in High-Traffic Public Systems

Challenge

How do you build a high-traffic website for a regional airport that is fast, on-brand, and fully inclusive, while overcoming skepticism about accessibility's perceived complexity, its cost, and its effect on visual design?

Solution

At Addis Enterprises, I led the accessibility strategy and UX design for the Bishop International Airport website and contributed to its front-end development, so that every traveler gets an equitable, efficient experience.

  • End-to-End Accessibility Ownership: Directed accessibility-first development from ideation to launch to ensure WCAG 2.1 compliance across visual, navigational, and interactive elements.
  • Intuitive, Keyboard-Friendly Navigation: Designed fully compliant menus with ARIA roles and keyboard and tab structures for travelers who use assistive technology.
  • Cross-Functional Advocacy: Won stakeholder buy-in through workshops, live demos, and strategy presentations that showed accessibility as good for the business and for the people it serves.
  • Scalable, Performance-Oriented Design: Optimized for heavy traffic during peak travel without giving up accessibility or cross-device consistency.
  • Conference Thought Leadership: Co-presented the design maturity and accessibility strategy at three regional design and tech conferences to encourage inclusive development across the industry. This led to my Human Experience Design principles.

Design Philosophy

I worked to build recognition of, and advocacy for, people of all abilities among my fellow developers and every stakeholder.

  • Lead with Empathy, Not Litigation: Colleagues hesitated over accessibility's cost. I won them over not through fear or force, but through the inspiration of mutual gain, by showing that the airport's business goals and its travelers' rights point the same way. Overcoming that reluctance taught me how to lead.
  • Accessibility as Systems Thinking: I designed the site to be perceivable, operable, and robust at every layer, from visual design to code structure, without splitting design and development into silos.
  • Culture Through Craft: Accessibility, and design maturity, lasts only if a team keeps practicing it after launch. While at MSU, my hybrid accessibility-UX-with-development-solutions reviews inspired a 3rd party agency to adopt standards-based best practices and led to my accessibility and ux leadership at another agency, Addis Enterprises.
Screenshot of the fully accessible tiered navigation with focus indicators
Accessible Navigation A navigation system built from the ground up for keyboard and screen reader users and tested live with assistive technologies.
Screen Reader & Keyboard Navigation Test A screen recording of the Bishop International Airport website showing its complex, tiered navigation, made fully accessible. From arrow keys to ARIA descriptions, each element was structured to be perceivable, operable, and efficient for people navigating with assistive technology.

Michigan State University · Digital Content and Accessibility Team

Creating an Inclusive Culture in Higher Education

Challenge

At a university with deeply entrenched legacy systems and limited resources, how can accessibility become a shared practice instead of a checklist before launch, especially in emotionally and socially urgent spaces like Title IX?

Solution

As part of the Digital Content and Accessibility Team, I helped lead a cultural and systemic shift in how the university approaches inclusion, accessibility, and digital equity.

  • WCAG Reporting System Design: Built a Likert-scale evaluation framework that ranks barriers by the harm they do to users against the effort to fix them, which sped up remediation and gave the university an objective, human-centered roadmap for accessibility.
  • Universal Design Advocacy: Held every tool, whether built internally or procured, to high accessibility and usability standards, which shaped key experiences, including Admissions, Title IX, and the Scholarships portal. Accessible post-review feedback interfaces then became redress mechanisms that carried equitable design to the university's procurement partners, a long-term investment in inclusion.
  • High-Impact Accessible Redesigns: Designed and prototyped interfaces for central university platforms with screen reader support, correct heading structures (e.g., site name as H2, page name as H1), and inclusive layouts.
  • From Digital Equity to Real-World Justice: Worked on the Title IX website redesign that helped survivors of Larry Nassar's abuse begin filing claims. I left at the end of 2016, as the case was growing: more than 100 survivors were in the lawsuit in 2017, and roughly 150 at his sentencing. It is a concrete example of what accessibility can do in the pursuit of justice, and I am grateful I got to contribute to it.
  • Cross-Industry Impact: My work at MSU led to accessibility evaluations for Fortune 500 clients and helped launch inclusive design strategies at two regional design agencies.

Design Philosophy

Designing for the edge is designing for the whole.

  • Design for Justice, Not Just Compliance: I approach accessibility as a way to advance equity, whether that means opening up scholarship access, clarifying admissions workflows, or supporting survivors seeking justice.
  • Small Changes, Big Impacts: A single change to the heading structure (H1 for the page, H2 for the site title) can transform a screen reader user's experience. Much of my work is small, precise changes like this one, which help far more people than their size suggests.
  • Build the Culture While Building the Product: Documentation alone did not change how teams worked. I won over resistant stakeholders by tying accessibility goals to values they already held and by building approachable systems that fit the way each team worked.

Deliverables

Screenshot of the new, fully accessible Michigan State University Admissions homepage and its tiered navigation
Admissions Homepage The redesigned MSU Admissions homepage, with accessible navigation, a clear heading structure, and usability testing behind it, so every prospective student can explore the university on equal terms.
Screenshot of the Michigan State University Admissions website showing keyboard focus indicators and tab order
Keyboard Navigation in Action Clear focus indicators and a visible tab order. They are essential for keyboard-only users and a foundation that every assistive technology builds on, including interfaces that do not exist yet.
Screenshot of the fully accessible Michigan State University Title IX website
Title IX Site Accessibility The redesigned Title IX system at MSU. On the main reporting form page, we made the site title an H2 and the page title an H1. That improved screen reader navigation and the site's universal design: a small detail underneath a system that helped survivors begin their pursuit of justice: more than 100 in the lawsuit in 2017, and roughly 150 at Larry Nassar's sentencing.

Prototype · Assistive technology · 2016

Keyreader: An AI Screen Reader Navigated by Voice

In 2016, I built and prototyped Keyreader, my own screen reader, on the pre-transformer AI work that powered Ellsi. A person could navigate a page by speaking, asking for what they wanted, such as the first or the second heading, and Keyreader moved there. Later versions went further and told a story of the front-end experience they ran on, offering directions, instructions, and context to make a page as easy as possible to use with assistive technology.

Asking for a heading by voice · Keyreader · September 23rd, 2016 A phone recording of the prototype on a laptop's screen, cropped to the page. The voice box still shows an earlier request when the clip begins. Asked "where is the second heading," Keyreader moves its focus from the level-one heading to the article title at level two.

Chapter 03

Stakeholder Engagement

An equity-centered redesign of energy assistance programs at Consumers Energy: customer interviews, the Mindset Map, an audio-augmented journey map, and workshops that gave a divided company one shared picture of its customers.

Consumers Energy · Equity-centered research for Low Moderate Income programs

Key Deliverable: The Mindset Map

Context: Equity-centered research for Low Moderate Income (LMI) energy assistance programs at Consumers Energy. The Mindset Map itself is the confidential intellectual property of CMS Energy and cannot be reproduced here, so the description below covers its method and its outcomes.

What it is

Mindsets are the idea that our customers exist beyond our products: how they perceive, learn about, and interact with Consumers Energy, its products, and its outreach is volatile and can vary drastically with social, financial, and technological context. The Mindset Map follows how a Michigander's state of mind changes over an ordinary day: where they run into friction, where they find relief, and where a utility eases their load or quietly adds to it. Instead of generic personas, we developed localized scenarios, such as a renter in rural Michigan without reliable internet, which revealed eligibility friction and input sensitivity flaws. The map widened who counted as a customer to include people who may not yet have any relationship with Consumers Energy or its partners, and it brought out the realities of ALICE households (Asset Limited, Income Constrained, Employed) that the program data was not capturing.

Why it worked

Participatory design was not an add-on; it was foundational. Through workshops, we co-mapped system boundaries and harm scenarios with stakeholders, informed by the lived experiences our interviews revealed. The map moved internal teams in a way personas never had: stakeholders who had argued strategy in the abstract started designing for the whole person their decisions touched, and product owners, executives, and frontline teams shared one reference point, drawn from customers' lives rather than organizational metrics.

Measurable outcomes

  • Enrollment improving from 1:6 toward a goal of 1:3 in low-income energy assistance programs serving vulnerable populations.
  • Enrollment copy tested in four languages: bureaucratic idioms were removed, and the copy was tested for understandability in English, Spanish, Arabic, and Vietnamese, which increased successful completions in underserved areas.
  • Influenced the three-year product roadmap for Low Moderate Income programs.
  • Reached board-level visibility in strategic presentations.
  • Realigned MVP priorities away from stakeholders' assumptions about LMI customer disengagement and toward a more accurate picture of how customers' mindsets shift at each point of contact.

Why this methodology matters beyond utilities

Mindsets are living personas. A persona describes who a customer is; a mindset describes how that person's state of mind changes at each moment a product touches their life, and the idea travels well beyond utilities. Any product used by people whose circumstances change (people in crisis, people partway through learning something, people working with AI systems that behave differently from one context to the next) is better served by research that tracks changing needs than by a fixed set of traits. Meeting needs drives natural behavior, and designing to promote natural, beneficial behavior is the key to lasting adoption.

The Mindset Map itself is not publicly available. Its method, its outcomes, and the related artifacts (journey maps, workshops, and audio-augmented research playback) appear below.

Consumers Energy · Prototype that carried the research to the executive floor

Northstar Prototype: From a Director's Vision to a Company-Wide Priority

Context: This prototype turned my customer interviews and what they revealed into a concrete design strategy. Before it, there was a director's vision and a few ideas whispered in separate departments, each shut down with "we don't have the evidence; we'll never get funding." After it, the conversation became "let's go meet our customers where they are, get to know them, and find out what would really help." What's shown here is the first-pass prototype, not the production design. Its fidelity is mid-tier on purpose; later iterations were built within the company's design standards and are protected as its intellectual property. What the prototype gave executives was a view of their customers' lives as the customers saw them, which program data alone could not provide.

The story arc

Before I joined: A director's initial vision was framed as a "one-stop shop" customer portal organized around utility-side features, retail enrollment, and financial security.

After interviews and workshops (Years 1–2): This prototype and the customer interviews behind it changed the question from "what features should we ship?" to "here's what customers are actually navigating in their financial lives." Michiganders came into view as everyday people managing financial complexity rather than bill payers to be optimized.

With partner organizations and nonprofits (Years 2–3): Backed by the Mindset Map, interactive workshops built around LMI board games, and 24-hour customer feedback on prototypes, the work became a company-wide retail enrollment priority sponsored by the SVP. The personas I wrote from the interviews and board-game workshops reached the Board of Directors, and the Mindset Map reached the CEO. Together, the research and prototypes were critical in breaking down the departmental silos that had kept customers' lived experience out of strategy conversations.

The business problem the work helped reframe

The existing Shut-Off Protection Plan had a 95% failure rate. It kept customers' lights on while it lasted, but it never made clear the total balance that would come due at reconciliation, so nearly every customer defaulted at the end. The prototype, together with the stakeholder and customer interviews and the Mindset Map, made that failure plain to executives. Customers were not the problem; the program had not been designed for the financial reality of the people it was meant to serve. Seeing that was where we began to answer customers' skepticism and build their trust.

Selected prototype screens

These two screens from the first-pass design became reference points in the first executive conversations. The dashboard shifts the customer's view from keeping up with bills to making progress toward goals and building financial health. The "What's next" screen adds a layer of achievements and lessons that supports customers through financial hardship, and it introduces retail enrollment options and nonprofit partners that can help them get out of arrears and toward stability through community resources.

Northstar prototype dashboard. Energy bill trend chart at top showing 12 months. Below: 'Your goals' section with a payment progress card. Next payment $33 of $250 due. Progress bar shows partial completion. Avg monthly payment $30. Start date Nov 2025. 'What's next' CTA button.
The dashboard, reframing the customer's view from bill-payment compliance to progress toward goals and financial health.
Northstar prototype 'What's next' screen. Large gauge at top showing payment progress: Current $200 paid (+20%), Next payment $33, Total Due $250. 'Add achievement' CTA. Suggested Achievements section lists Rainy day, New car, Own home. Resources lessons section: How to save, Ways to save, Principles of saving.
An achievement and education layer designed to support customers through financial hardship instead of penalizing them for it.

Outcome: Within 18 months, the work grew from a team-level initiative into a company-wide retail enrollment priority, and research artifacts informed by this prototype reached the top of the organization.

Consumers Energy · Interviews, an interactive journey map, and workshops

The Journey Map Experience: Stories, Sound, and Workshops

Challenge

When I joined Consumers Energy (CMS Energy), Low Moderate Income (LMI) customers faced barriers no one inside the company could see, and product design often misunderstood or overlooked their experiences. Internal teams worked in silos, each seeing only a fragment of the picture, and none of them could see the everyday lives of the customers they served.

Unless stakeholders agreed on one human-centered vision, CMS Energy's products would stay disconnected from one another and from the people they were meant to help.

Solution

Data alone was not going to change minds. We needed empathy, and that meant hearing real stories from real customers.

I started by leading in-depth interviews that brought stakeholders face to face with the practical and emotional realities of LMI customers across Michigan. We heard how worn down ALICE customers (Asset Limited, Income Constrained, Employed) were from juggling bills and how frustrated they were by needlessly complicated enrollment red tape. We also saw how hard these people worked to keep their families comfortable.

Design Philosophy

An important part of the role of an experience designer is to be a storyteller. That includes knowing our audience and moving our audience.

  • Stories as Connective Tissue: I moved the conversation away from each department's separate view and toward one shared story about the customers, which gave teams a personal stake in the outcomes they were shaping.
  • Human-Centered Means Story-Centered: Empathy is not walking a mile in someone's shoes, which assumes they fit; it is seeing things through their lens, as they see through their experience. Customers' own stories made that possible, and stakeholders moved by them went on to build solutions that genuinely help the people they serve.
  • Listening as a Radical Act: By listening closely and sharing customers' own words, I changed how the company saw its LMI customers: not as abstract segments, but as resilient individuals we were privileged to help.
  • Building Vision, Not Just Artifacts: Confidentiality limits which deliverables I can show, but the storytelling produced something that outlasted them: a shared vision that continued to guide decisions.
Early whiteboard designs for the Mindset Map
Early whiteboarding used a decision tree to map customer journey insights against enrollment flows and expose how fragmented the system was. It led to the Mindset Map, a wide view of our customers' lives as everyday people.
Storyboard of a stakeholder's dream product, drawn in a workshop
Visualizing a Stakeholder's Dream This storyboard turned a stakeholder's abstract vision into a plan several departments could act on. Grounding big ideas in customers' everyday situations kept the ambitious features tied to real human problems.

From stories to a shared journey map

These stories changed the direction of the work. To share them, I designed an interactive Journey Map Experience: visuals paired with customers' own recorded voices, showing where our marketing and outreach eased their burdens and where it unintentionally added to them.

Discoverability Drives Engagement Even Without Incentives Across the organization, many believed that highlighting premium features, even ones free to qualifying customers, would overwhelm systems or set expectations the company could not meet. Some product teams were reluctant to promote them at all.
In this clip, a participant goes back over the outreach and notices the mention of premium upgrades. The reaction is immediate: surprise, curiosity, a wish to learn more, and even a willingness to enroll. Product owners had assumed only the upgrade itself would move customers, but here the mention alone signals that the program is worth joining.
Holding back key features out of caution was limiting the program's reach. When customers can see what is possible, they trust the program more and want to take part.
Take Two: When Imagery Does Resonate Stakeholders assumed the marketing imagery fell flat and reinforced a view of Consumers Energy as municipal, cold, and transactional. In this clip, customers tell a different story. Asked about outreach materials, they describe how particular images in marketing emails shaped their mindset, building trust and prompting them to engage, or doing the exact opposite.
That challenged what the company believed about its own outreach. The finding reshaped MVP priorities toward outreach that feels timely and human and that supports customers the way they want to be supported.
Framing the Experience: Opening and Closing Moments The opening and closing of the Journey Map Experience were designed to do more than inform: they give stakeholders time to reflect on what they heard and to agree on what it means. Their pacing, warmth, and moments of delight ease resistance, make room for empathy, and leave an impression that lasts after the stories end.
Sound as Empathy Catalyst The full video of the interactive journey map with customer feedback (about eight minutes)

Layering customer audio directly into the visual journey let stakeholders hear for themselves the gap between what they intended and what customers actually experienced.

Our internal culture publication later described this moment as a turning point in breaking down silos and rebuilding a shared vision:

Culture change documentation
Breaking Silos Through Story and Sound This excerpt from the culture publication describes how the audio-augmented journey map broke up competing strategies and gave the company the confidence to change the product and its website. It is text-heavy, but it records a lasting change in how the organization thought.

Hearing customers directly changed how teams talked about them. Departments began to share a common language drawn from their own experience and from what customers had told them, and they could now see and hear the effect of their decisions on real people.

Alongside the journey map, we ran hands-on workshops that used storytelling to turn abstract strategy into something people could feel and build around.

High-fidelity, stakeholder-influenced journey map
Real Voices, Real Reflection Built together in a live session, this full-scale journey map combined customer quotes (some real, some written for the exercise) with stakeholders' expertise. Asking teams to test their assumptions against those quotes exposed their biases, brought forward voices they had not heard, and showed where their shared understanding fell short.
Journey map empathy game
Empathy Game Interaction Captured during a board-game activity, this close-up shows stakeholders stepping into the shoes of LMI customers and making real choices under imagined constraints. Playing through those choices built an emotional connection and produced insights that a standard research readout would not have.

In these sessions, product owners, executives, and frontline teams reexamined existing products and imagined new ones from the customer's point of view.

Sticky notes on a blank journey map canvas in a workshop weighing two core features for an MVP
Workshop Journey Map Stakeholders debated two competing visions for the LMI MVP.
Through guided facilitation and live synthesis, we brought their perspectives together, clarified assumptions, and surfaced shared priorities.
Product journey map canvas overlaid with sticky notes at the start of a workshop
LMI Product Navigation This workshop journey map traced how LMI customers moved through overlapping systems across several products. Although it focused on specific products, it uncovered unmet needs and mismatched expectations that were holding back the broader service strategy.

Storytelling is how I built trust and collaboration here, and it ended in an organizational shift toward proactive, empathetic product development that puts customers' outcomes first and business outcomes second.

Chapter 04

Outcomes

Every result on this page in one place, with the section it came from.

AI Eval, Research & Safety

  • Cognitive architecture research (Raffie): a multi-system architecture in which valence, arousal, drive dynamics, satiation, allostatic load, a Big Five trait model, a global-workspace arbitration module, and metacognitive uncertainty are explicit, measurable variables.
    • Since a summer 2026 milestone, the subsystems have formed a pursuit loop, a feedback architecture that includes the user's interactions: traits modulate emotional and motivational responses, a workspace arbitrates among competing contents, and outcomes update the drives and, slowly and within bounds, the traits themselves, so the system's state persists and develops across sessions.
    • The drive substrate was audited, repaired, and certified on August 7th, 2026, against a written per-drive specification, with fixes screened in a simulator first, acceptance judged on the distribution across independent runs, and plausible repairs rejected when they measured worse.
    • Recertified in both regimes on September 6th, 2026, after a remeasurement found it out of specification; the work included reverting two of my own repairs once the evidence showed they were the regression.
    • Met the unchanged specification again on September 18th, 2026, in both regimes, after live probes found the action loop halting two minutes after every boot. With the loop repaired, the engine was recalibrated so that only evidence satisfies a drive, and the scorer was corrected where it had been reading its own sample clock: twelve of twelve runs healthy, with goal droughts at one percent.
    • Global-workspace ablation (September 2026): with the term that lets thoughts support one another switched off, the workspace holds a thought on 31% of competitive ticks, against 89% with it intact, over the same eight scripted days on the same build. The manipulation check reads exactly zero, sign and Wilcoxon tests are reported per measure, and there is one run per arm; the result was made in one capture series and reproduced in a second.
    • Conversation model in sync (October 2026): the conversation layer was retrained as Conversation 5 so that what it says about the system's own state is composed from live readings of the subsystems. It gives a drive's current figure when asked how strong a pull is, the buffer's own count when asked how loaded its working memory is, and a plain statement when a reading is absent, and it does not improvise claims the readings do not support.
    • Learned decision models (October 2026): each implementation's executive was given two models trained on its own decision ledger (84 and 60 twelve-minute sessions of the running application; 4,706 and 3,364 decisions), valuing relief of the six drives weighted by how unmet each one is. The full set was accepted on one implementation with every specification band held; on the other, only the goal model was accepted, after the full set collapsed its drives, a finding that took both implementations to see. A person's reply, not a delivered message, was made the event that satisfies the need to connect. On October 11th, on both implementations, the executive's own line became a message and a reply completed it. That evening, the models were switched on, each implementation serving what its own acceptance licensed; rescoring on records of real use in which a person answers is still owed. Measured on October 16th and 17th, 2026, in two comparisons registered before their data, each on 32 held-out sessions: the learned rule relieved more need than the decision rules of ACT-R, Soar-RL, and Psi by 15.4–18.0% of its own gain over random choice, every 95% interval above zero; its advantage over LinUCB, a learner reading the same inputs, held in the first measurement and did not replicate in the second, and both are reported. No person took part, and the estimates are off-policy.
    • Pre-registered model-welfare study across two independent implementations of the same states: Study 1's 10,600 cells ran to a single registered analysis and did not pass its gate, with one of four hypotheses failing, so the discriminant claim was dropped as the pre-committed failure map required. Study 2, registered at 11,200 cells on each implementation, will be frozen and collected once, as its fifth generation.
    • Ongoing research; selected details are available on request.
  • Held Ground (September 2026): a conversational product for hard personal decisions with many constraints, built on the workspace layer: a shelf of up to twelve constraints the user has approved, with at most four reasoned from at a time; every item kept in the user's own words with its source; contradictions quoted back to the user as questions; and a full shelf that asks what a new item replaces. Every reply to an open message in its demo was generated live by the fine-tuned conversation model.
  • Energy-based hallucination detection (Top HAT and Nested Hope): designed and trained a Hebbian Associative Transformer that integrates generation and factual self-verification in a single forward pass.
    • Contrastive energy learning, a six-phase curriculum, and multi-scale factuality scoring.
    • Energy gaps of +10 to +13 between clean and corrupted text, against a target of more than 0.3, on synthetic corruptions.
    • Variant A of Nested Hope (220M): hallucination self-detection AUROC of 0.586 against v10's 0.516 under identical methodology with a clean train/holdout split, and a Phase 1 validation loss of 3.07 against 3.77, about a 51% perplexity reduction.
    • A documented data-leakage flaw in contrastive hallucination evaluation: on the same model and data, splitting combinatorially built pairs at the pair level scored an AUROC of 0.998, and splitting at the prompt level scored 0.530. Any benchmark that builds pairs this way should split by prompt.
    • A controlled frozen-backbone experiment (June 2026) that ruled out the training objective as the bottleneck before any compute was spent on scale: the original margin loss and a variance-aware ranking surrogate, trained on identical cached features and scored on an identical held-out split, both scored at chance.
    • September 2026: the same head, trained on the answers of Conversation 4 (a 1.5B chat model) under a pass rule written before the first answer was drawn, scored 0.58 against a 0.75 gate, and pilot models failed the gate at every size tried, including 1.5B (0.64) and 7B (0.59). This is reported as a negative result and as a measured tradeoff between bolting a head onto a chat model and training the base model for the signal from the start.
  • Custom architecture: a 181M-parameter Hebbian Associative Transformer with a multi-scale, spectral-normalized energy head (2.9M parameters), competitive Hebbian memory with sparse, auditable association patterns, and a production inference server with multiple generation modes.
  • Raffie's conversation model:
    • Retrained in September 2026 as Conversation 4, a 1.5B instruction model fine-tuned first with supervision and then on preferences, using synthesized, graded conversations: shelf agreement 0.996, asks on every scored contradiction and near-miss, persona breaks 0.26%, and knowledge not below its base model's. Four shelf behaviors could not be measured by the grading model and are recorded as such. It replaced the original encoder-decoder (0.78 accuracy, 0.95 loss).
    • Retrained in October 2026 as Conversation 5, from the same base with the same recipe, with the architecture's live readings added and a judge that reads them. On held-out replies, it uses numbers from the readings or the conversation 99.7% of the time, says so when a reading is absent 98.7% of the time, claims no application-only action 98.8% of the time, and leaks no thoughts. Its claims stay within the readings 85% of the time, with an interval too wide to decide against a 90% floor because it over-refuses. Shelf agreement is 0.996, persona breaks are 0.27%, and its knowledge is not below its base model's. Cut replies remain past their ceiling, and the general win rate (0.42) is below its floor but undecided by its interval. It was installed with a live check: all 30 phrasings that the page's handlers miss reached the model with its readings, and two replies in thirty stated a figure the readings lacked.
  • Cognitive Link (June 2026): published a coupled framework for agentic AI that cross-maps four decades of cognitive science and human-automation research onto agentic systems.
    • Evaluation must descend Marr's levels.
    • Design must place friction rather than remove it.
    • Transparency, agency, and collective input are properties of the coupling, not the model.
    • Published with a stakeholder brief for nontechnical readers and an implementation roadmap toward human-centered agentic AI.
  • Earlier published frameworks (2025):
  • Philosophical grounding: Graduate-level independent study in philosophy of mind under Dr. Robyn Bluhm (MSU, Spring 2016) covering AI, consciousness, emotion, neuroscience, and evolutionary biology. Three argumentative essays and a reference letter are available on request.
  • Earlier work: the Ellsi voice assistant (2016–2017), an emotionally grounded notification and conversation system designed before LLMs as a small-scale application of a grounded emotional system. Her empathy model was later carried into Raffie; her mood-steering tool use deliberately was not, as a design choice against manipulation.
  • See visual artifacts and more in the AI Eval, Research & Safety section

Accessibility

  • Accessibility for vulnerable users: led accessibility strategy at Michigan State University and, at Addis Enterprises, for Bishop International Airport, through WCAG audits, inclusive design systems, developer workflow redesign, and stakeholder workshops. Reframed accessibility from a compliance checklist to a harm-reduction practice.
  • 40% reduction in accessibility remediation time for major e-learning platforms through scalable audit and developer-first workflows.
  • Title IX system redesign at MSU helped survivors of Larry Nassar's abuse begin filing claims (more than 100 in the lawsuit in 2017, roughly 150 at his sentencing), an example of accessibility as harm-reduction infrastructure.
  • Inclusive design evaluation matrix adopted as RFP requirements across about 50 platform purchases, including the Admissions, Scholarships, and Title IX systems.
  • See visual artifacts and more in the Accessibility section

Stakeholder Engagement

  • Mindset Map methodology: a research artifact that tracks how customer mindsets shift across moments of contact. It moved internal teams at Consumers Energy to design for vulnerable populations they had not seen clearly before.
  • Enrollment improving from 1:6 toward a goal of 1:3 in low-income energy assistance programs.
  • Influenced the three-year product roadmap for Low Moderate Income (LMI) programs at Consumers Energy.
  • Board-level visibility in strategic presentations.
  • Cross-functional alignment through journey mapping, audio-augmented research playback, and storytelling workshops that translated complex service realities into shared product strategy across product owners, executives, and frontline teams.
  • See visual artifacts and more in the Stakeholder Engagement section