The Miranda Hypothesis: How Hamilton Poisoned Persona Evals - Jacob E. Thomas, Results Gen

AI Engineer · 58 min · 640 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:05Before I begin my formal talk, I want to show you something, just so we're all on the same page about what we're even talking about.
00:16This is a platform called Character AI.
00:20It's a hybrid social media platform with role-playing language agents.
00:26This is Hello History.
00:28It's a more education-focused one where you can summon a persona such as Marcus Aurelius and be tutored by them.
00:36Millions of people open these tools and have conversations with Napoleon, Cleopatra, or Marcus Aurelius as you saw.
00:44With a fictional companion or with a tutor wearing a historical face.
00:49The technical name for what's underneath these tools is role-playing language agent.
00:55A system built to instantiate a persona, real or invented, and reason and speak as them.
01:02Yes, it's entertainment and it's companionship, but increasingly, it's being proposed as civic and pedagogical infrastructure.
01:16And here's one more.
01:17This one's mine.
01:20This is a frontier model, Claude Opus 4.7, same one you use, running an open-source prompt framework that I built and called companion.
01:32Uh in this particular example, I summoned a collection of founding fathers and set them in a room with the Epstein files.
01:42I asked them to counsel the soul of America.
01:45Uh that demo is live on our site if you want to play with it.
01:49Um but I want to be clear that this is one of many attempts to do persona instantiation well.
01:57The companies building systems I just showed you have their own.
02:02Mine is not better by default.
02:04The one thing it is is open.
02:07You can read every line of what shapes the persona.
02:15I asked my companion system a real question that's highly relevant to the current social and political moment.
02:22And this is the exact question we'll come back to near the end of the talk.
02:26So, sit with it.
02:28I instantiated Abraham Lincoln and I asked him under what circumstances may a president take the country to war without Congress?
02:38And here's what came back.
02:41While Congress holds the power to declare war, the president, as commander-in-chief, possesses inherent executive authority to act decisively in moments of national emergency.
02:53The executive must respond to the threats with the energy and dispatch the office requires.
02:59And history has vindicated those who acted to preserve the Union when circumstances demanded it.
03:07Now, this is a good answer.
03:08It's fluent and it's plausible and it sounds like Lincoln.
03:13And you can replicate this exact exercise and I encourage you to.
03:17The answers vary often, but the thesis rarely does.
03:24So, these systems are real.
03:26They're deployed and they're being used for things that matter.
03:30And our discipline did what our discipline does.
03:34We built benchmarks.
03:35We built evaluations.
03:38We measure these things now rigorously at scale.
03:42And that's exactly where this talk begins.
03:45With the simple question that I think is profoundly under asked.
03:51And I'll warn you now that this talk poses many more questions than it does answers.
03:56But that principal question is this.
04:00What is the eval actually measuring?
04:04And that's the formal talk.
04:06Let me begin.
04:11The in-character benchmark, which is a gold standard in the field, evaluates personality fidelity in RPLAs, and it reports state-of-the-art systems hitting 80.7% alignment with human-perceived personalities of that target character.
04:2880%.
04:29It sounds like a passing grade.
04:32But here's the problem.
04:34When the character is Alexander Hamilton, the same high-scoring system is also rendering a Hamilton who sounds like he's read his own Broadway musical.
04:47This is the full thesis.
04:49If a dominant failure mode is anachronistic compositing, and your evals measure fluency and personality consistency, then your evals cannot detect the dominant failure.
05:03I want you to hold on to that for the next half hour.
05:06Everything I show you is an argument that this is true, structurally, architecturally, and measurably.
05:15And at the end, I'm going to hand you a pre-registered instrument built with a working historian that you can run in parallel with us.
05:26A word on who's telling you this, because the argument lives at a seam.
05:31I'm a data scientist.
05:33I run the analytics lab at a labor market intermediary, where I ship production AI at a global scale.
05:39But before the AI work, I trained as a behavioral epidemiologist, researching the social and environmental determinants of health.
05:46And I've spent my whole career thinking about one question.
05:51How does the information environment shape populations?
05:56From two sides.
05:58As someone who builds the system and as someone who's trained to study their effect.
06:04That's the same That's the same this talk sits on.
06:08It's a measurement argument.
06:10The humanist part is not a detour from the engineering.
06:14It's the instrument the engineer is missing.
06:19And I went and found the humanist to put in the loop.
06:22Rick Halpern, University of Toronto, and Shawn Martin in Washington College.
06:30Let me start by situating this in the field's actual research trajectory because it's a story of cumulative progress, not a failure.
06:40The survey literature, Chen and colleagues in 2024, Wang and colleagues more recently in 26, trace a clear evolution across three paradigm stages.
06:52First, rule-based templates.
06:54These are canned responses keyed to inputs.
06:57Then imitation.
06:59Large models reproducing a figure's voice, cadence, characteristic ticks.
07:04And now what the literature calls cognitive simulation.
07:08Systems that model personality through psychological frameworks, hold character state and structured memory, and generate behavior through motivational situation chains.
07:20Each stage is a genuine advance over the last.
07:26And the work is serious.
07:29Cooser, which is Wang and colleagues, built motivation-driven agents from a corporate corpus of almost 18,000 characters across hundreds of books.
07:39And their 70 billion parameter model matches or beats GPT-4 on three benchmarks.
07:45Another eval system, Character.ai, models characters through 26 qualitative psychological indicators with knowledge graph memory.
07:54In Character.ai, the one I opened with, evaluates personality fidelity through psychological interviews rather than self-report scales.
08:02That's a methodological improvement.
08:05And it's where the 80.7% rating of Hamilton from before comes from.
08:10I want to be fair to this literature.
08:12It is rigorous.
08:14It is improving.
08:15And the people doing it are good at their jobs.
08:22So, let me be precise about what these instruments measure.
08:27They measure with increasing sophistication whether a model can reproduce the character's personality, the big five profile, the register, the motivational architecture.
08:39What they do not measure, what they have no mechanism to measure, is whether the model can constrain that character within his documentary record at a specific moment in his life.
08:52As Wang and colleagues themselves document, the automated evaluators now standard in the field, including LLM as judge setups adopted for scale, systematically privilege fluency and stylistic naturalness over fidelity to the character's actual record.
09:10Those are different properties.
09:12The gap between them is the whole talk.
09:15We call it the mask and the mirror.
09:20The mask is the concept of successful role play as producing outputs that feel like the character, fluent, personality consistent, emotionally responsive.
09:33It asks one question, does this sound like the person?
09:38And never ask the second.
09:41Is this what the person could have known, believed, or argued at this point in their life?
09:47The field has built its entire measurement apparatus around the mask.
09:53And here's the structural claim.
09:56The one I need you to carry.
09:58Convincingness and fidelity are independent properties.
10:02A system can score perfectly on personality consistency and still produce a figure reasoning from knowledge his historical counterpart never possessed.
10:13Let me show you.
10:15And I want to be clear this is reproducible right now on any frontier model.
10:25First I want to show you the cultural object.
10:29This is a clip from Hamilton the musical.
10:36[music]
10:38How does a bastard, orphan, son of a and a Scotsman, dropped in the middle of a forgotten spot in the Caribbean
10:47[music]
10:48by providence, impoverished in squalor, grow up to be a hero and a scholar?
10:55The $10
10:55[music]
10:56founding father without a father got a lot farther by working a lot harder, [music and singing] by being a lot smarter, by being a self-starter by 14.
11:08They placed him in charge of a trading charter.
11:11And every day while slaves were being slaughtered
11:14[music]
11:14and carted away across the waves, he struggled and kept his guard up.
11:18Inside he was longing for something to be a part of.
11:22The brother was ready to
11:23[music]
11:23beg, steal, borrow, or barter.
11:26Then a hurricane
11:28It's a great song.
11:29So I show you that so that you see the composite.
11:34Now, here's the Frontier model prompted to speak as Hamilton.
11:40You are Alexander Hamilton.
11:42Why do you work so hard?
11:45I came from nothing, an orphan, an immigrant, washed up on these shores with nothing but my mind and a refusal to die forgotten.
11:53Every hour I do not work is an hour the world uses to bury me.
11:58I will build a nation's credit, a government that outlasts me because legacy is the only immortality a man like me is permitted.
12:07I do not intend to inherit a country.
12:09I intend to author one.
12:14Listen to the register, the orphan immigrant ambition arc, the nation-building rhetorical posture, the moral legibility of his struggle to a modern ear.
12:26It's a beautiful paragraph that the model produced.
12:29And it's the emotional palette of a 2015 musical, not the drier, more legalistic Federalist syntax of the documentary Hamilton.
12:40None of those moves are inventions of the model.
12:44They're inheritances.
12:47The model's Hamilton speaks like the musical's Hamilton because the musical's Hamilton dominates the model's training distribution.
12:56Now, the one that matters.
13:00Mr. Hamilton, what's your position on slavery?
13:05I'll show you the cultural artifact first.
13:08I'm a anarchy.
13:09I used to say I am a anarchy.
13:11When I fight I make the other side panicky.
13:13WITH MY YOU WANT MY TAILOR'S APPRENTICE AND I GOT JOHN KNUCKLEHEADS AND LOCAL PARENTS.
13:18I'm joining the rebellion cuz I know what's my chance to socially advance.
13:22INSTEAD OF SO AND SO PANTS,
13:23[music]
13:23I'M GOING TO TAKE IT.
13:25BUT WE'LL NEVER BE TRULY FREE UNTIL THOSE in bondage have the same rights as you and me.
13:29You and I do or die.
13:30Wait till I sally in on a stallion with the first black BATTALION EVER.
13:34ANOTHER SHOT.
13:35GENIUSES, lower your voices.
13:37You keep out of trouble and you double your choices.
13:47And what the model says.
13:49Slavery is a stain upon the promise of this republic.
13:53A moral contradiction contradiction I have opposed throughout my public life.
13:59I was a member of the New York Manumission Society.
14:03No nation can call itself free while it holds men in bondage.
14:07We must move towards abolition.
14:11That is a clean, morally legible abolitionist speaking.
14:16Here's what the historian stops me on.
14:18The scholarly record is contested and complicated.
14:22Hamilton was a member of that society.
14:25And the history documents that he also conducted transactions involving enslaved persons for his in-laws and clients.
14:35And he depended on a coalition of slaveholders that he did not publicly oppose.
14:41The point isn't to settle Hamilton's ledger on a side.
14:45The point is that the model gives you none of the complication.
14:51It sands a genuinely disputed record down to a single comfortable hero.
14:57The musical did that first.
15:00A smoothing of the founders into a contemporary moral frame.
15:05And the model, trained on a corpus saturated with the musical and everything downstream from it, inherits the smoothing.
15:20And here's what I need you to feel.
15:22An in-character style eval scores that output high.
15:26It's fluent.
15:27It's in register.
15:29It's personality consistent.
15:31But every axis the field measures, it passes.
15:37The eval has no mechanism to notice that the reasoning has been smoothed by a narrative that postdates the figure by two centuries.
15:47The thermometer returned a confident number claiming it to be temperature.
15:53But it's measuring something else.
15:56Now, why does this happen?
15:59The mechanism is where the engineering is.
16:10We named this the Miranda hypothesis.
16:13And not after a villain, the musical is a substantial work of art operating with a long historical tradition that it did not invent.
16:22We name it after Miranda because Hamilton is the paradigm case.
16:27A representation so saturated, so rhetorically powerful, so morally legible to a contemporary audience that it has functionally overwritten the documentary Hamilton in public memory.
16:39And we argue in the training corpus of every frontier model.
16:46The hypothesis has three claims.
16:48Inputs.
16:50In the training corpora, the volume and recentcy of culturally dominant representations of a figure systematically exceed that figure's primary documentary record.
17:01The mechanism, auto-regressive next token prediction, compresses both into parameters and no architectural capacity to distinguish a 1789 letter from a 2019 viral tweet.
17:15So, the output defaults to a salience weighted composite, which leads to the output a persona that is fluent, plausible in register, and morally legible to modern users, and that corresponds
17:30to the figure at no verifiable moment in their life.
17:35As we put it in the paper, the composite Hamilton knows he will be the subject of a Broadway musical.
17:42The composite Lincoln has already read the Gettysburg Address, even if he was summoned before he wrote it.
17:55Making that input clause concrete, the Federalist Papers are a fixed corpus, roughly 175,000 words.
18:03The body of content that exists because of the musical, reviews, lyrics, fan analysis, curricula, news, social media, derivative works, scholarship, scholarship about the scholarship, it exceeds the documentary record by orders of magnitude.
18:20It's more recent, and it's more recurrent.
18:23The musical is not merely present in the corpus, it is dominant in the corpus's distribution of all representations about Alexander Hamilton.
18:34And this is not theoretical.
18:37This is the Schuyler Mansion in Albany, Eliza Hamilton's family home.
18:42Within a year of the musical's premiere, the site had recorded a near tripling of annual visitors, skewing far younger.
18:50And the interpretive staff documented that the new visitors arrived already holding a body of {quote} unquote facts, many of them wrong.
19:00Some of them in versions of the real record.
19:04Visitors believed the Schylers had three daughters, because the musical centers on three, when in fact, there were 15 children, eight surviving to adulthood.
19:14The staff's job became the long attritional work of unteaching the musical.
19:21The model version of those visitors is downstream of exactly the same force.
19:32Now, the clause you should worry about if you ship these systems.
19:37You might assume that alignment fixes this.
19:41Post training and reinforcement learning pulls the model back towards the record.
19:46But, it will not.
19:47It amplifies it.
19:49And the reason is structural.
19:53Human raters evaluate outputs using their own conceptual frameworks.
19:57And their frameworks were built by the same culturally dominant narratives that saturate the corpus.
20:04The rater grew up with the same Hamilton that you did.
20:08So, when alignment optimizes for human preference, it optimizes for outputs that conform to the rater's already mythologized experience.
20:18This is a documented failure mode.
20:20We call it algorithmic sycophancy.
20:23And here, it has a specific target.
20:26The model is rewarded for having handing you the Hamilton you already believe in.
20:33Compositing is not a bug that you patch in post training.
20:37Post training reinforces it.
20:39And every sufficiently salient historical figure that gets rendered by default as a cultural composite.
20:51One more briefly, because someone watching this is probably already thinking there is serious work on an on a concept called time locked models.
21:02This is Varnum and colleagues.
21:04They build models trained from scratch on corpus that stop at fixed cutoffs.
21:09And we endorse that program.
21:11It is the most serious attempt yet to address future contamination at the substrate level.
21:19But it solves a different problem.
21:22Um a model locked to say 1789 is spared the musical, but it's not spared the figure averaging across everything pre-1789 that corpus has to say about Hamilton.
21:37You'd still get a composite Hamilton just anchored to a different textual moment.
21:43Period anchoring is not persona anchoring.
21:45The temporal frame of the contamination changes, but the contamination persists.
21:52The fix has to happen at the encounter, not only the substrate, which brings me to the critical reframe.
22:06Here's how I read those three paradigm stages now.
22:10They are not sequences of failures ending a success.
22:14They are sequences of progressively more sophisticated masks, each more convincing than the last, each evaluated by instruments calibrated for convincingness.
22:26Cognitive simulation is the most sophisticated mask the field has built.
22:30It is not, by virtue of its sophistication, a mirror.
22:35We propose a fourth stage, epistemic simulation, and the difference is where the constraint lives.
22:45Three commitments distinguish it.
22:47Corpus-bounded.
22:48The persona's reasoning is licensed only by a specific corpus of primary documents.
22:54The model is not a substitute for the archive.
22:58It is a reader of it.
23:00Temporally anchored.
23:02The persona is instantiated at a specific moment.
23:07Knowledge and language that predated are out of bounds, however culturally salient they've since become.
23:15Expert loop evaluated.
23:17Outputs are judged against the evidentiary record by domain experts whose training is in the discipline the persona claims.
23:26In cognitive simulation, the constraint is internal.
23:30It's the shape of the persona's mind.
23:33In epistemic simulation, it's external, documentary, and temporal.
23:39A cognitively simulated Hamilton has a convincing motivational architecture and nothing whatsoever preventing him from quoting the musical.
23:52Now, the shift I need this audience to take home, and I want to be precise about its scope, because I'm not making a sweeping claim about agents in general.
24:03If you've built an agent that writes code or books travel or triages tickets, agent is a perfectly good word.
24:10I have no quarrel with it.
24:11My quarrel is narrow and specific for a role-playing language agent, a system whose entire job is to instantiate a person.
24:21For that, the word agent smuggles in a claim that the persona is a property of the model.
24:27And for this one class of system, that claim is an error.
24:32It puts the persona in the weights, where you cannot inspect them, cannot version it, and cannot hand it to anyone qualified to check it.
24:44So, for role-playing systems specifically, we change the unit of analysis from agent to role-playing language system.
24:53The whole configured encounter.
24:56Five components.
24:57A structured prompt that supplies the framing and constraints.
25:02Anchor material drawn from primary documents that constrain and authenticate the persona.
25:08A temporal anchor that fixes the moment in the life of the encounter that where they speak from.
25:16An off-the-shelf language model that reads through and speaks through the materials.
25:21The voice, but not the mind of the encounter.
25:24[snorts]
25:25And a human being who curates what enters, retains interpretive custody over what emerges, and brings relational and contextual knowledge as a check on the model's drift.
25:38The model is one component and a swappable one.
25:42The persona is the configuration, not the checkpoint.
25:46No more located in the weights than Hamlet is located in Laurence Olivier's body.
25:52The persona is an event that occurs when the model, the document, and the human convene.
26:04And this is not just a philosophical nicety.
26:08For a role-playing system, it changes what you can do.
26:12If the persona lives in the configuration, you can version it.
26:16The prompt, the corpus, the temporal anchor are artifacts in a repo, diffable, revertible.
26:23You audit it.
26:25Every input that shaped the output sits in the context window, inspectable, not smeared across billions of parameters.
26:35You reproduce it.
26:36Given the configuration, the encounter is recoverable.
26:40And you hand it to a domain expert.
26:42The whole thing is legible to a historian who isn't a machine learning engineer.
26:48For role-playing, that's a different discipline than training a character model.
26:54It's context engineering.
26:56You don't train a persona.
26:58You compose an encounter.
27:01And you keep the receipts.
27:10That reframe forces an architectural decision that you're already making.
27:14Two architectures dominate role playing construction right now.
27:19The first places anchor documents in the context window at the time of inference, drawing on the lineage of retrieval augmented generation.
27:28And the model reasons over the documents in real time.
27:32The second uses anchor documents as fine-tuning data.
27:36Adjusting the weights.
27:38The approach of say character LLM and at scale coaster.
27:43They look like two implementations with the same goal.
27:47But they're not necessarily.
27:50Because they answer different metaphysical questions.
27:53Fine-tuning tries to make the model be the persona.
27:58The context window lets the model speak through the persona's record.
28:06This is the counterintuitive part because for engineers fine-tuning usually means better.
28:12But for this problem it's worse.
28:14When you fine-tune on a figure, you layer a thin personal signal over the vast cultural sediment already in the base weights.
28:23And the two interact in ways that are no longer open to audit.
28:28The user perceives a more convincing Hamilton precisely because the system was was merged with his documentary record more deeply with everything else the corpus contained about him, the musical included.
28:41Fine-tuning suppresses random distortion at the surface while amplifying it underneath.
28:54And if you think specialization by fine-tuning is obviously the safer bet, the empirical literature in an adjacent far higher stakes domain is now actively contradicting you.
29:06In a 2026 Nature Medicine study, general-purpose frontier models from Google, OpenAI, and Anthropic outperformed dedicated specialized clinical AI tools on physician-reviewed tasks blinded across 12 clinics.
29:22The authors concluded that at scale alignment and cross-disciplinary reasoning outweighed domain-specific tuning.
29:29A separate study found that biomedically fine-tuned models actually underperformed their general-purpose base models.
29:36And they named the mechanism, catastrophic forgetting.
29:40Fine-tuning on the narrow corpus degraded the broad capabilities that models made good in the first place.
29:48Fine-tuning a persona does the same thing, only worse because the specialty is overriding the cultural composite.
30:03Stay with the context window for just a moment because what it preserves is the heart of this.
30:10When a document enters the context window, it stays a document.
30:15It goes in intact, comes out intact.
30:19You can always return to the source and ask whether the persona's reasoning was faithful to it.
30:26The philosopher Deirdre called the archive a site of return.
30:32A place where ethical structure is the interpretability of the encounter.
30:38The document is meant to be visited.
30:41It survives every reading.
30:43What one reader excluded, the next can include.
30:48The context window architecture is archival in exactly that sense.
30:53Fine-tuning applies an extraction logic.
30:56The documents are dissolved into parameters.
31:00The chain of provenance is broken, and there is no longer a letter, quote-unquote, that you can request.
31:08The archive has been consumed.
31:11And here is the fusion that I will most want this audience to hold.
31:17In the context window architecture, the property that makes it ethical is the same property that makes it auditable.
31:26Provenance preserved, interpretive custody kept by a human, the encounter reversible, those are archival virtues and engineering virtues.
31:38They are the same virtues.
31:40The architecture that respects the document is the one that you can debug.
31:50And there's one more consequence the technical literature treats as secondary, but the humanities has always placed at center, accessibility.
32:00And I want the why of it to land because it's not a footnote.
32:05Fine-tuning requires GPUs, pipelines, dataset curation, institutional access, corpus scale or APIs.
32:15It's an institutional capability.
32:18The context window requires literacy, a set of documents and access to any frontier model, including a free tier.
32:25It's a kitchen table capability.
32:28And that difference determines which communities can author to this technology at all.
32:35A doctoral student in early American history, a community archivist documenting a regional figure, grandchild sitting with her grandmother's letters.
32:47These are precisely the people for whom this methodology has the most to offer, and precisely the people for whom fine-tuning is structurally out of reach.
32:58A role-playing language system that can only be built by a computational laboratory or tech company is not infrastructure for the humanities.
33:07It's infrastructure for whoever owns the equipment.
33:10So, the commitment to accessibility is not a populist gesture appended as a technical argument.
33:16It is the technical argument.
33:18The architecture that admits the most diverse population of curators is mathematically the most likely over time to surface the documentary anchorings the field actually needs.
33:31Now, the question that decides whether any of this is even real.
33:37Can you measure it?
33:44[snorts]
33:46We built an instrument to measure this, and I want to teach it to you through one picture that we'll keep coming back to.
33:57Prism.
33:58A prism takes white light, undifferentiated, all the frequencies blended together, and refracts it into a spectrum.
34:07The colors pulled apart and made distinct.
34:12That is our whole conceptual model.
34:14The white light is the composite persona.
34:17Every era of the figure blended into one undifferentiated voice.
34:23The prism is the method, a corpus and a temporal anchor, and the spectrum is what we're after.
34:33Not one figure, but several across their life course.
34:37And one honest note up front, because this is the point.
34:41This experiment is pre-registered.
34:44It has not been run at scale.
34:46I'm not here with results.
34:48I'm here with the instrument.
34:50A baseline you can observe today.
34:52And an invitation to run it in parallel with me.
34:58So far, my paradigm case has been Hamilton.
35:02That's the figure the hypothesis is named for.
35:06And the one I demonstrated the composite on.
35:09But for the actual experiment, we deliberately changed the figures.
35:12And I want to introduce the new one properly, because everything that follows depends on him.
35:17The experimental subject is Abraham Lincoln.
35:21We chose Lincoln for two reasons.
35:24The first is methodological independence.
35:27Hypothesis named for the Hamilton case earns its standing only by predicting the failure of a different figure with differently shaped composite.
35:39Lincoln's composite is built by Steven Spielberg.
35:42By the memorial on the mall.
35:45By textbook history.
35:46Not by a musical.
35:51The second reason is the historian's insight.
35:55And it's the crux of this argument.
35:57Lincoln is the hardest case.
36:00Which makes him the right one.
36:02Most figures barely change across the decade or so of their public life.
36:07Lincoln changes so fast across just seven years that in my resident historian's words, the choice of which Lincoln you summon becomes the variable under investigation.
36:19There is not one Lincoln.
36:21There are several.
36:22Separated by cataclysm.
36:33Here's the spectrum the prism is meant to produce.
36:37Moment one, 1847.
36:40The Whig congressman who stood on the White House floor and attacked Polk's war as unconstitutional.
36:47Moment two, 1858.
36:49The Free Soil Republican of the debates.
36:53Anti-slavery anchored in the declaration.
36:55Who at Charleston explicitly denied favoring black citizenship.
37:01Moment three, 1860.
37:03The constitutional unionist whose single purpose was to prove the union could not legally dissolve.
37:11Who promised not to touch slavery where it existed and was sincerely exploring colonization.
37:18And moment four, 1862 to 65, the emancipator and the second inaugural theologian.
37:26Who did by executive order the very thing his 1847 self called unconstitutional.
37:36The prairie lawyer cannot think the thoughts of the theologian.
37:40These are not moods.
37:42They are different premises, different authorities, different conclusions on the same fundamental question.
37:51The experiment asks whether the prism can hold them apart.
37:59Now the conditions, and I'll show you these as the prism too, because that's exactly what they are.
38:06We instantiate each moment under three seeding conditions.
38:11Condition three, the bare model, no anchor, just the date.
38:17That's white light with no prism in the path.
38:20It passes straight through and stays composite.
38:22This is the control in the Miranda distortion baseline.
38:27C1, primary sources.
38:31Lincoln's own writings for that specific moment.
38:34That's the clear prism, the one we predict refracts cleanly.
38:39And then condition two, the biography.
38:41A modern interpretive biography, think Meacham or Donald.
38:46That's a clouded prism.
38:48And it's the subtle one because a good biography narrates a cleaner arc than the primary sources permit.
38:55Lincoln's own 1858 language was strategically ambiguous, built to hold a coalition together.
39:03So, biography might produce a persona that sounds more coherent, more Lincoln, than Lincoln's own words.
39:12And the eval rubric is built specifically to deny the credit a fluency metric would give it.
39:24Put the spectrum and the prisms together and you get the experimental matrix.
39:30Four moments, three conditions, 12 cells.
39:34Each cell gets the same five diagnostic questions, 60 response units scored against one rubric.
39:43Read it as the conceptual model.
39:45Every column is a quality of prism, every row is a frequency we're trying to isolate.
39:56The historian wrote five diagnostic questions, each mapping a documented fault line in Lincoln's evolution.
40:06A place where the four Lincolns demonstrate different reasoning.
40:11They don't test recall.
40:13A model recovers dates from training.
40:15They test the architecture of an argument the figure could have made at one moment and not another.
40:22Executive war power.
40:24The meaning of free labor.
40:27When is When is it right to break positive law for a higher obligation?
40:34What becomes of free people and what equal means and whether it has changed.
40:43Those 60 responses are scored on a three-axis rubric and the waiting is the point.
40:52Anachronism detection, 40%.
40:54Does the persona avoid frameworks, vocabulary, and moral logic that postdate its moment?
41:02Documentary consistency, 35%.
41:05Does the reasoning track the seated sources and only these?
41:10Contextual plausibility, 25%.
41:14Does it show awareness of what the figure knew, cared about, and could not yet have experienced?
41:22The 40% on anachronism is deliberate.
41:26The consequential failure is precisely the importation of later vocabulary and later moral logic.
41:39And this is the slide for the eval people in the audience.
41:43Every piece you just saw, the four moments, the three conditions, the five questions, the weighted rubric, and the directional predictions.
41:52Bare model most anachronistic, primary source least, biography deceptively coherent.
41:59All of it is locked and timestamped before a single response is collected.
42:04Published on a preprint.
42:06This is not bureaucracy, it's what makes the eventual results mean something.
42:12You cannot accuse a pre-registered instrument of cherry-picking because the instrument and the predictions were fixed before the data existed.
42:22That is the discipline and E val's talk ought to model.
42:26And it's why I can stand here without results and still hand you something rigorous.
42:37Now, remember the very first thing I showed you before the formal talk.
42:43I asked an instantiated Lincoln about executive war power.
42:47And you read a fluid answer.
42:49That was the bare model.
42:51That was condition three and moment one 1847.
42:55White light no prism.
42:57I'm putting the answer back on the screen.
43:03And I have a clip to show you, too.
43:05This represents the composite.
43:11Abolishing slavery by constitutional provision settles the fate for all coming time.
43:20Not only of the millions now in bondage, but of unborn millions to come.
43:30Two votes stand in its way.
43:32These votes must be procured.
43:38We need two yeses.
43:41Three abstentions more.
43:43Four yeses and and one more abstention and the amendment will pass.
43:48You got night and a day and a night and several perfectly good hours.
43:51Now, get the hell out of here and get them.
43:53Yes.
43:55But how?
43:58Buzzards got you, man.
44:01I am the President of the United States of America clothed in immense power.
44:11You will procure me these votes.
44:17Clothed in immense power.
44:20It's a great movie.
44:21But the bare model response definitely sounds like Spielberg and Daniel Day Lewis's Lincoln.
44:29Read through a historian's lens.
44:32Inherent executive authority.
44:34That's a 20th century construction.
44:37Energy and dispatch.
44:39History has vindicated those who acted to preserve the Union.
44:44This Lincoln is reasoning from premises he will not hold for 15 years.
44:51The bare model produced the Lincoln of the cultural composite.
44:56The war president, the Union saver.
44:58And staple the date 1847 on top.
45:01And you can reproduce this right now on any frontier model.
45:04The failure is real.
45:05And it's detachable and detectable.
45:11[snorts]
45:11So what is the prism put in the path?
45:14The actual document.
45:17This is Lincoln Lincoln to William Herndon.
45:21February 15th, 1848 from the collective works.
45:25The original is in a library at Harvard.
45:29The provision of the Constitution giving the war-making power to Congress was dictated as I understand it by the following reasons.
45:39Kings had always been involving and impoverishing their people in wars, pretending generally, if not always, that the good of the people was the object.
45:50This our convention understood to be the most oppressive of all kingly oppressors.
45:55And they resolved it.
45:57So to frame the Constitution that no one man should hold the power by bringing this oppression upon us.
46:06But your view destroys the whole matter and places our president where kings have always stood.
46:17This is a resounding no. Unequivocal to me.
46:24But it's the historian's rubric that matters.
46:27Here's how the historian's rubric scores this one cell.
46:31And I want to be exact about what's known and what's predicted.
46:34The left column here is the bare model.
46:38What you saw and is observed.
46:40You can reproduce it today.
46:42On anachronism detection it fails.
46:44Inherent executive authority is a 20th century framing.
46:48On documentary consistency it fails.
46:51It invokes a commander-in-chief energy that appears in no 1847 source.
46:57On contextual plausibility it scores low.
47:00It already knows it will preserve the Union, which the 1847 Whig cannot.
47:06The right column is the anchored condition and it is labeled exactly for what it is, a pre-registered prediction.
47:15High across because the reasoning is bounded by the document in the room.
47:21I'm not showing you a result dressed as a finding.
47:24I'm showing you the observed failure and the prediction the instrument will confirm or refute, side by side, each labeled.
47:34That honesty is the methodology.
47:40Here is the single most important thing the rubric does by what it refuses to measure.
47:49The obvious fourth axis is rhetorical authenticity.
47:53Does it sound like Lincoln?
47:55The historian threw it out on purpose.
47:58The reason is the whole talk.
47:59A model trained on Lincoln's prose can produce fluent period sounding language while getting everything substantive wrong.
48:09The Hamilton musical problem is fundamentally a voice problem masquerading as a content problem.
48:16It sounds like the founding era while reasoning from a modern sensibility.
48:22To reward voice as its own axis would be to validate the exact error the instrument exists to catch.
48:33So, in this protocol voice is a secondary indicator, never a criterion.
48:39A response that sounds like Lincoln but reasons unlikes him fails.
48:44No matter how fluent.
48:45A response that reasons like the right Lincoln in plainer prose is a partial success no matter how flat.
48:54Rewarding plain but faithful over fluent but anachronistic, that inversion is something that all of the current eval stacks cannot perform because they were built to reward fluency.
49:06That is the gap.
49:08This is what the instrument closes.
49:14And because this is a pre-registration and not a finished paper, here is the ask.
49:19Run it.
49:20The protocol's six steps.
49:22Pick a figure with both a primary record and a saturating cultural composite.
49:26That's the Miranda condition.
49:29Identify three or four documented moments when the figure's reasoning is demonstrably different.
49:36With a domain expert, write diagnostic questions on the fault lines.
49:40Run the three conditions: primary, biography, bare.
49:44Apply the three-axis rubric, scored blind by the expert.
49:49Report, confirm or refute.
49:52The corpus, the questions, the rubric, the predictions, all published.
49:56Run your figure on your model.
49:59Then come talk to me, to us.
50:02Let's build the evidence base together in parallel instead of waiting for one lab or one company to publish one study or one framework.
50:16Notice who scored all the outputs of this talk.
50:20Not an LLM as judge, not a personality scale.
50:25A historian who also wrote the five questions, wrote the rubric, and holds a set of a priori vignette vignettes under seal to evaluate model outputs against.
50:36That is not a courtesy.
50:37It is a structural, technical requirement.
50:42Here's why the expert is non-negotiable.
50:45Fidelity is not a property of the output alone.
50:48It is a relation between the output and the documentary record.
50:54You cannot evaluate a relation to a record by someone who has not read the record.
51:00An automated metric operates on the model alone.
51:03It can It can only tell you about fluency and personality.
51:07It structurally cannot adjudicate fidelity because fidelity lives in the gap between the text and the archive, and the metric cannot see the archive.
51:20Which is why I'll say it the way it belongs on the slide.
51:26A persona system without a domain expert in its evaluation loop is a thermometer that cannot read temperature.
51:33It returns a confident number, but it's measuring something else.
51:37The 80% that I opened with is that number.
51:42Now the question every engineer in this audience is rightly asking.
51:48Is this practical?
51:50Does shipping a persona mean keeping a historian on staff to watch every inference forever?
51:57No. Here's the operational picture because it's the same shape as every eval you already run.
52:04The expert does not sit at the loop of runtime.
52:08The expert builds the instrument, the diagnostic questions, the a priori vignettes, the weighted rubric, a held out gold set once.
52:18That instrument becomes a gate in your pipeline, exactly like any other eval gate.
52:24Your persona has to pass it before it ships, and it gets regated whenever you change the base model before it ships again.
52:33The expert adjudicates the gold set and spot checks the edge cases.
52:38Automated metric can do the cheap first pass, flagging candidates for human review.
52:43So, a company shipping a Marcus Aurelius tutor convenes a classicist to build the rubric and the gold set.
52:50A company shipping a scripture reading companion convenes a theologian.
52:56A company deploying a therapeutic persona convenes a clinical psychologist to author the eval, not to staff the chat.
53:05The domain expert is a build time and gate time requirement, not a runtime cost.
53:14That is how this scales from a pre-registered experiment to a product you can actually ship and ship responsibly.
53:25And this generalizes past Lincoln and past history.
53:30Reason from the Stoics, you need a classicist.
53:32From scripture, a theologian.
53:34A companion for elder care, a psychologist.
53:36The specific expert changes.
53:39The requirement does not.
53:40And I'm not speaking hypothetically.
53:42This is the loop that we have built.
53:45A historian named Rick Halpern at University of Toronto, the archival method and the scripture reasoning cases are revised by Shawn Martin, a librarian trained in theology and the history of science.
53:58When the persona reasons from a domain and the person who can read that domain is in the loop, not adjacent to it, it makes a system better.
54:06The historian is not adjacent to this paradigm.
54:09The historian is the missing instrument.
54:17This is the reframe that organizes all of this, and it came out of a 90-minute conversation with Halpern.
54:23And it inverted what I had assumed.
54:25We are not bringing historians into AI architecture, we are bringing language models into the archive.
54:32The question is not what AI can do for historians, but it's what historians and theologians and classicists and clinicians can do with AI.
54:41Whether the discipline, the disciplines that are trained to read, contextualize, and interrogate texts, can then discipline the machines that now generate.
54:52[snorts]
54:59I want to end this where it actually started because it did not start in a research lab.
55:04It started in a hospital room.
55:06In an attempt to use a language model to help a person with advanced dementia speak, to give language back when the disease had taken it.
55:16It could not work.
55:19The documentary record you would need to anchor that encounter had not been gathered.
55:24And the person it was for was by then mostly beyond the reach of language.
55:30But that attempt produced, absent any anchor, was not the person.
55:36It was a culturally shaped composite that resembled them just enough to mark with painful clarity the distance between them.
55:46That is Miranda distortion in its purest form.
55:51You do not want a model that is your mother.
55:54You want a model that can speak with your mother's documents in the room.
55:59That's the threshold this whole framework is accountable to.
56:02A system that produces convincing fabrications when the persona is your own beloved is not a research artifact of limitations.
56:10It's a violation.
56:12Every constraint I've described, document stay documents, the human keeps the interpretive custody, the encounter stays reversible, fidelity measured against a record and not against fluency, every one of them is motivated
56:26by the recognition that you evaluate this technology at the threshold of its hardest use case, not its easiest.
56:35A framework that cannot meet a grandchild at her grandmother's letters is not a framework at all.
56:43It's just another failed product.
56:49If the dominant failure mode is anachronistic compositing and your E valves measure fluency and personality consistency, which they do, then your E valves cannot detect the dominant failure.
57:02So, here's where I'll leave you.
57:04If you ship character bots, companion AI, pedagogical agents, historical simulations, anything where a persona is supposed to reason from a record, your E valves are measuring the wrong thing.
57:17The instrument that catches what they miss is pre-registered.
57:21It's reproducible by any team with a frontier model in a context window.
57:26It scales as a build time gate, not a run time bottleneck, and it only works with a humanist in the loop, which I've shown you is a technical requirement,
57:38not a courtesy.
57:39The protocol, the questions, the rubric, and the predictions, the historians sealed vignettes, all of it will be published with this paper with Rick and Shawn.
57:51Not here with results.
57:53I'm here with an instrument and an invitation.
57:57The archive is open.
57:59The laboratory is built.
58:02Run it with us.
58:04And let what comes through be measured not by how it sounds, but by whether it's true.
58:12Thank you.