WEBVTT NOTE Sentence-level transcript of https://www.youtube.com/watch?v=IJXjTLPzvAU NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/IJXjTLPzvAU.html. A cue ends where the next begins, or 2 s after its last word. s1 00:00:05.600 --> 00:00:15.960 Before 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. s2 00:00:16.800 --> 00:00:20.400 This is a platform called Character AI. s3 00:00:20.400 --> 00:00:25.840 It's a hybrid social media platform with role-playing language agents. s4 00:00:26.080 --> 00:00:28.080 This is Hello History. s5 00:00:28.080 --> 00:00:36.240 It's a more education-focused one where you can summon a persona such as Marcus Aurelius and be tutored by them. s6 00:00:36.480 --> 00:00:44.920 Millions of people open these tools and have conversations with Napoleon, Cleopatra, or Marcus Aurelius as you saw. s7 00:00:44.920 --> 00:00:49.520 With a fictional companion or with a tutor wearing a historical face. s8 00:00:49.520 --> 00:00:55.000 The technical name for what's underneath these tools is role-playing language agent. s9 00:00:55.000 --> 00:01:02.160 A system built to instantiate a persona, real or invented, and reason and speak as them. s10 00:01:02.160 --> 00:01:11.680 Yes, it's entertainment and it's companionship, but increasingly, it's being proposed as civic and pedagogical infrastructure. s11 00:01:16.200 --> 00:01:17.840 And here's one more. s12 00:01:17.840 --> 00:01:20.280 This one's mine. s13 00:01:20.480 --> 00:01:32.640 This is a frontier model, Claude Opus 4.7, same one you use, running an open-source prompt framework that I built and called companion. s14 00:01:32.640 --> 00:01:40.840 Uh in this particular example, I summoned a collection of founding fathers and set them in a room with the Epstein files. s15 00:01:42.280 --> 00:01:45.800 I asked them to counsel the soul of America. s16 00:01:45.800 --> 00:01:49.640 Uh that demo is live on our site if you want to play with it. s17 00:01:49.640 --> 00:01:57.160 Um but I want to be clear that this is one of many attempts to do persona instantiation well. s18 00:01:57.160 --> 00:02:02.160 The companies building systems I just showed you have their own. s19 00:02:02.160 --> 00:02:04.840 Mine is not better by default. s20 00:02:04.840 --> 00:02:07.840 The one thing it is is open. s21 00:02:07.840 --> 00:02:12.160 You can read every line of what shapes the persona. s22 00:02:15.080 --> 00:02:22.600 I asked my companion system a real question that's highly relevant to the current social and political moment. s23 00:02:22.600 --> 00:02:26.400 And this is the exact question we'll come back to near the end of the talk. s24 00:02:26.400 --> 00:02:28.320 So, sit with it. s25 00:02:28.320 --> 00:02:38.000 I instantiated Abraham Lincoln and I asked him under what circumstances may a president take the country to war without Congress? s26 00:02:38.120 --> 00:02:40.720 And here's what came back. s27 00:02:41.120 --> 00:02:53.760 While 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. s28 00:02:53.760 --> 00:02:59.480 The executive must respond to the threats with the energy and dispatch the office requires. s29 00:02:59.480 --> 00:03:07.440 And history has vindicated those who acted to preserve the Union when circumstances demanded it. s30 00:03:07.440 --> 00:03:08.920 Now, this is a good answer. s31 00:03:08.920 --> 00:03:13.560 It's fluent and it's plausible and it sounds like Lincoln. s32 00:03:13.560 --> 00:03:17.160 And you can replicate this exact exercise and I encourage you to. s33 00:03:17.160 --> 00:03:22.560 The answers vary often, but the thesis rarely does. s34 00:03:24.040 --> 00:03:26.280 So, these systems are real. s35 00:03:26.280 --> 00:03:30.400 They're deployed and they're being used for things that matter. s36 00:03:30.400 --> 00:03:34.000 And our discipline did what our discipline does. s37 00:03:34.000 --> 00:03:35.720 We built benchmarks. s38 00:03:35.720 --> 00:03:38.040 We built evaluations. s39 00:03:38.040 --> 00:03:42.959 We measure these things now rigorously at scale. s40 00:03:42.959 --> 00:03:45.880 And that's exactly where this talk begins. s41 00:03:45.880 --> 00:03:51.480 With the simple question that I think is profoundly under asked. s42 00:03:51.480 --> 00:03:56.560 And I'll warn you now that this talk poses many more questions than it does answers. s43 00:03:56.560 --> 00:04:00.680 But that principal question is this. s44 00:04:00.680 --> 00:04:04.040 What is the eval actually measuring? s45 00:04:04.040 --> 00:04:06.760 And that's the formal talk. s46 00:04:06.760 --> 00:04:08.920 Let me begin. s47 00:04:11.840 --> 00:04:28.400 The 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. s48 00:04:28.400 --> 00:04:29.960 80%. s49 00:04:29.960 --> 00:04:32.840 It sounds like a passing grade. s50 00:04:32.840 --> 00:04:34.040 But here's the problem. s51 00:04:34.040 --> 00:04:45.480 When 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. s52 00:04:47.280 --> 00:04:49.000 This is the full thesis. s53 00:04:49.000 --> 00:05:02.960 If a dominant failure mode is anachronistic compositing, and your evals measure fluency and personality consistency, then your evals cannot detect the dominant failure. s54 00:05:03.120 --> 00:05:06.240 I want you to hold on to that for the next half hour. s55 00:05:06.240 --> 00:05:15.720 Everything I show you is an argument that this is true, structurally, architecturally, and measurably. s56 00:05:15.720 --> 00:05:24.840 And 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. s57 00:05:26.320 --> 00:05:31.280 A word on who's telling you this, because the argument lives at a seam. s58 00:05:31.280 --> 00:05:33.040 I'm a data scientist. s59 00:05:33.040 --> 00:05:39.440 I run the analytics lab at a labor market intermediary, where I ship production AI at a global scale. s60 00:05:39.440 --> 00:05:46.800 But before the AI work, I trained as a behavioral epidemiologist, researching the social and environmental determinants of health. s61 00:05:46.800 --> 00:05:51.040 And I've spent my whole career thinking about one question. s62 00:05:51.040 --> 00:05:56.080 How does the information environment shape populations? s63 00:05:56.600 --> 00:05:58.560 From two sides. s64 00:05:58.560 --> 00:06:04.840 As someone who builds the system and as someone who's trained to study their effect. s65 00:06:04.840 --> 00:06:08.840 That's the same That's the same this talk sits on. s66 00:06:08.840 --> 00:06:10.880 It's a measurement argument. s67 00:06:10.880 --> 00:06:14.800 The humanist part is not a detour from the engineering. s68 00:06:14.800 --> 00:06:18.280 It's the instrument the engineer is missing. s69 00:06:19.120 --> 00:06:22.400 And I went and found the humanist to put in the loop. s70 00:06:22.400 --> 00:06:28.240 Rick Halpern, University of Toronto, and Shawn Martin in Washington College. s71 00:06:30.480 --> 00:06:40.600 Let me start by situating this in the field's actual research trajectory because it's a story of cumulative progress, not a failure. s72 00:06:40.680 --> 00:06:52.040 The survey literature, Chen and colleagues in 2024, Wang and colleagues more recently in 26, trace a clear evolution across three paradigm stages. s73 00:06:52.040 --> 00:06:54.240 First, rule-based templates. s74 00:06:54.240 --> 00:06:57.280 These are canned responses keyed to inputs. s75 00:06:57.280 --> 00:06:59.320 Then imitation. s76 00:06:59.320 --> 00:07:04.919 Large models reproducing a figure's voice, cadence, characteristic ticks. s77 00:07:04.919 --> 00:07:08.880 And now what the literature calls cognitive simulation. s78 00:07:08.880 --> 00:07:20.960 Systems that model personality through psychological frameworks, hold character state and structured memory, and generate behavior through motivational situation chains. s79 00:07:20.960 --> 00:07:25.880 Each stage is a genuine advance over the last. s80 00:07:26.760 --> 00:07:29.120 And the work is serious. s81 00:07:29.120 --> 00:07:39.040 Cooser, which is Wang and colleagues, built motivation-driven agents from a corporate corpus of almost 18,000 characters across hundreds of books. s82 00:07:39.040 --> 00:07:45.840 And their 70 billion parameter model matches or beats GPT-4 on three benchmarks. s83 00:07:45.840 --> 00:07:54.560 Another eval system, Character.ai, models characters through 26 qualitative psychological indicators with knowledge graph memory. s84 00:07:54.560 --> 00:08:02.720 In Character.ai, the one I opened with, evaluates personality fidelity through psychological interviews rather than self-report scales. s85 00:08:02.720 --> 00:08:05.040 That's a methodological improvement. s86 00:08:05.040 --> 00:08:10.800 And it's where the 80.7% rating of Hamilton from before comes from. s87 00:08:10.800 --> 00:08:12.640 I want to be fair to this literature. s88 00:08:12.640 --> 00:08:14.120 It is rigorous. s89 00:08:14.120 --> 00:08:15.600 It is improving. s90 00:08:15.600 --> 00:08:18.919 And the people doing it are good at their jobs. s91 00:08:22.880 --> 00:08:27.680 So, let me be precise about what these instruments measure. s92 00:08:27.680 --> 00:08:39.599 They measure with increasing sophistication whether a model can reproduce the character's personality, the big five profile, the register, the motivational architecture. s93 00:08:39.599 --> 00:08:52.839 What 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. s94 00:08:52.839 --> 00:09:10.120 As 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. s95 00:09:10.120 --> 00:09:12.080 Those are different properties. s96 00:09:12.080 --> 00:09:15.160 The gap between them is the whole talk. s97 00:09:15.160 --> 00:09:19.240 We call it the mask and the mirror. s98 00:09:20.240 --> 00:09:33.839 The mask is the concept of successful role play as producing outputs that feel like the character, fluent, personality consistent, emotionally responsive. s99 00:09:33.839 --> 00:09:38.840 It asks one question, does this sound like the person? s100 00:09:38.840 --> 00:09:41.200 And never ask the second. s101 00:09:41.200 --> 00:09:47.480 Is this what the person could have known, believed, or argued at this point in their life? s102 00:09:47.480 --> 00:09:53.880 The field has built its entire measurement apparatus around the mask. s103 00:09:53.880 --> 00:09:56.360 And here's the structural claim. s104 00:09:56.360 --> 00:09:58.200 The one I need you to carry. s105 00:09:58.200 --> 00:10:02.920 Convincingness and fidelity are independent properties. s106 00:10:02.920 --> 00:10:13.800 A system can score perfectly on personality consistency and still produce a figure reasoning from knowledge his historical counterpart never possessed. s107 00:10:13.800 --> 00:10:15.680 Let me show you. s108 00:10:15.680 --> 00:10:21.920 And I want to be clear this is reproducible right now on any frontier model. s109 00:10:25.240 --> 00:10:29.200 First I want to show you the cultural object. s110 00:10:29.640 --> 00:10:33.720 This is a clip from Hamilton the musical. s111 00:10:36.012 --> 00:10:38.012 [music] s112 00:10:38.320 --> 00:10:47.820 How does a bastard, orphan, son of a and a Scotsman, dropped in the middle of a forgotten spot in the Caribbean s113 00:10:47.820 --> 00:10:48.839 [music] s114 00:10:48.839 --> 00:10:55.080 by providence, impoverished in squalor, grow up to be a hero and a scholar? s115 00:10:55.080 --> 00:10:55.662 The $10 s116 00:10:55.662 --> 00:10:56.560 [music] s117 00:10:56.560 --> 00:11:08.160 founding 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. s118 00:11:08.160 --> 00:11:11.120 They placed him in charge of a trading charter. s119 00:11:11.120 --> 00:11:14.002 And every day while slaves were being slaughtered s120 00:11:14.002 --> 00:11:14.160 [music] s121 00:11:14.160 --> 00:11:18.800 and carted away across the waves, he struggled and kept his guard up. s122 00:11:18.800 --> 00:11:22.360 Inside he was longing for something to be a part of. s123 00:11:22.360 --> 00:11:23.827 The brother was ready to s124 00:11:23.827 --> 00:11:23.920 [music] s125 00:11:23.920 --> 00:11:26.400 beg, steal, borrow, or barter. s126 00:11:26.400 --> 00:11:28.360 Then a hurricane s127 00:11:28.360 --> 00:11:29.640 It's a great song. s128 00:11:29.640 --> 00:11:34.280 So I show you that so that you see the composite. s129 00:11:34.680 --> 00:11:39.520 Now, here's the Frontier model prompted to speak as Hamilton. s130 00:11:40.200 --> 00:11:42.280 You are Alexander Hamilton. s131 00:11:42.280 --> 00:11:45.120 Why do you work so hard? s132 00:11:45.120 --> 00:11:53.920 I came from nothing, an orphan, an immigrant, washed up on these shores with nothing but my mind and a refusal to die forgotten. s133 00:11:53.920 --> 00:11:58.840 Every hour I do not work is an hour the world uses to bury me. s134 00:11:58.840 --> 00:12:07.040 I will build a nation's credit, a government that outlasts me because legacy is the only immortality a man like me is permitted. s135 00:12:07.040 --> 00:12:09.920 I do not intend to inherit a country. s136 00:12:09.920 --> 00:12:13.680 I intend to author one. s137 00:12:14.320 --> 00:12:26.360 Listen to the register, the orphan immigrant ambition arc, the nation-building rhetorical posture, the moral legibility of his struggle to a modern ear. s138 00:12:26.360 --> 00:12:29.640 It's a beautiful paragraph that the model produced. s139 00:12:29.640 --> 00:12:40.120 And it's the emotional palette of a 2015 musical, not the drier, more legalistic Federalist syntax of the documentary Hamilton. s140 00:12:40.120 --> 00:12:44.960 None of those moves are inventions of the model. s141 00:12:44.960 --> 00:12:47.000 They're inheritances. s142 00:12:47.000 --> 00:12:56.040 The model's Hamilton speaks like the musical's Hamilton because the musical's Hamilton dominates the model's training distribution. s143 00:12:56.440 --> 00:12:58.920 Now, the one that matters. s144 00:13:00.440 --> 00:13:04.800 Mr. Hamilton, what's your position on slavery? s145 00:13:05.280 --> 00:13:08.560 I'll show you the cultural artifact first. s146 00:13:08.560 --> 00:13:09.080 I'm a anarchy. s147 00:13:09.080 --> 00:13:11.400 I used to say I am a anarchy. s148 00:13:11.400 --> 00:13:13.520 When I fight I make the other side panicky. s149 00:13:13.520 --> 00:13:18.720 WITH MY YOU WANT MY TAILOR'S APPRENTICE AND I GOT JOHN KNUCKLEHEADS AND LOCAL PARENTS. s150 00:13:18.720 --> 00:13:22.360 I'm joining the rebellion cuz I know what's my chance to socially advance. s151 00:13:22.360 --> 00:13:23.694 INSTEAD OF SO AND SO PANTS, s152 00:13:23.694 --> 00:13:23.760 [music] s153 00:13:23.760 --> 00:13:25.200 I'M GOING TO TAKE IT. s154 00:13:25.200 --> 00:13:29.360 BUT WE'LL NEVER BE TRULY FREE UNTIL THOSE in bondage have the same rights as you and me. s155 00:13:29.360 --> 00:13:30.840 You and I do or die. s156 00:13:30.840 --> 00:13:34.680 Wait till I sally in on a stallion with the first black BATTALION EVER. s157 00:13:34.680 --> 00:13:35.360 ANOTHER SHOT. s158 00:13:35.360 --> 00:13:37.360 GENIUSES, lower your voices. s159 00:13:37.360 --> 00:13:41.320 You keep out of trouble and you double your choices. s160 00:13:47.080 --> 00:13:49.760 And what the model says. s161 00:13:49.920 --> 00:13:53.960 Slavery is a stain upon the promise of this republic. s162 00:13:53.960 --> 00:13:59.480 A moral contradiction contradiction I have opposed throughout my public life. s163 00:13:59.480 --> 00:14:03.080 I was a member of the New York Manumission Society. s164 00:14:03.080 --> 00:14:07.560 No nation can call itself free while it holds men in bondage. s165 00:14:07.560 --> 00:14:11.160 We must move towards abolition. s166 00:14:11.320 --> 00:14:16.160 That is a clean, morally legible abolitionist speaking. s167 00:14:16.160 --> 00:14:18.560 Here's what the historian stops me on. s168 00:14:18.560 --> 00:14:22.520 The scholarly record is contested and complicated. s169 00:14:22.520 --> 00:14:25.640 Hamilton was a member of that society. s170 00:14:25.640 --> 00:14:35.640 And the history documents that he also conducted transactions involving enslaved persons for his in-laws and clients. s171 00:14:35.640 --> 00:14:41.200 And he depended on a coalition of slaveholders that he did not publicly oppose. s172 00:14:41.200 --> 00:14:45.720 The point isn't to settle Hamilton's ledger on a side. s173 00:14:45.720 --> 00:14:51.360 The point is that the model gives you none of the complication. s174 00:14:51.400 --> 00:14:57.760 It sands a genuinely disputed record down to a single comfortable hero. s175 00:14:57.760 --> 00:15:00.320 The musical did that first. s176 00:15:00.320 --> 00:15:05.160 A smoothing of the founders into a contemporary moral frame. s177 00:15:05.160 --> 00:15:14.680 And the model, trained on a corpus saturated with the musical and everything downstream from it, inherits the smoothing. s178 00:15:20.120 --> 00:15:22.200 And here's what I need you to feel. s179 00:15:22.200 --> 00:15:26.520 An in-character style eval scores that output high. s180 00:15:26.520 --> 00:15:27.840 It's fluent. s181 00:15:27.840 --> 00:15:29.520 It's in register. s182 00:15:29.520 --> 00:15:31.800 It's personality consistent. s183 00:15:31.800 --> 00:15:36.800 But every axis the field measures, it passes. s184 00:15:37.160 --> 00:15:47.080 The eval has no mechanism to notice that the reasoning has been smoothed by a narrative that postdates the figure by two centuries. s185 00:15:47.080 --> 00:15:53.360 The thermometer returned a confident number claiming it to be temperature. s186 00:15:53.360 --> 00:15:56.320 But it's measuring something else. s187 00:15:56.320 --> 00:15:59.360 Now, why does this happen? s188 00:15:59.360 --> 00:16:03.760 The mechanism is where the engineering is. s189 00:16:10.200 --> 00:16:13.280 We named this the Miranda hypothesis. s190 00:16:13.280 --> 00:16:22.440 And not after a villain, the musical is a substantial work of art operating with a long historical tradition that it did not invent. s191 00:16:22.440 --> 00:16:27.360 We name it after Miranda because Hamilton is the paradigm case. s192 00:16:27.360 --> 00:16:39.600 A representation so saturated, so rhetorically powerful, so morally legible to a contemporary audience that it has functionally overwritten the documentary Hamilton in public memory. s193 00:16:39.600 --> 00:16:45.080 And we argue in the training corpus of every frontier model. s194 00:16:46.000 --> 00:16:48.839 The hypothesis has three claims. s195 00:16:48.839 --> 00:16:50.320 Inputs. s196 00:16:50.320 --> 00:17:01.600 In the training corpora, the volume and recentcy of culturally dominant representations of a figure systematically exceed that figure's primary documentary record. s197 00:17:01.600 --> 00:17:15.680 The mechanism, auto-regressive next token prediction, compresses both into parameters and no architectural capacity to distinguish a 1789 letter from a 2019 viral tweet. s198 00:17:15.680 --> 00:17:30.360 So, 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 s199 00:17:30.360 --> 00:17:35.840 to the figure at no verifiable moment in their life. s200 00:17:35.840 --> 00:17:42.960 As we put it in the paper, the composite Hamilton knows he will be the subject of a Broadway musical. s201 00:17:42.960 --> 00:17:51.400 The composite Lincoln has already read the Gettysburg Address, even if he was summoned before he wrote it. s202 00:17:55.080 --> 00:18:03.560 Making that input clause concrete, the Federalist Papers are a fixed corpus, roughly 175,000 words. s203 00:18:03.640 --> 00:18:20.920 The 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. s204 00:18:20.920 --> 00:18:23.360 It's more recent, and it's more recurrent. s205 00:18:23.360 --> 00:18:34.000 The musical is not merely present in the corpus, it is dominant in the corpus's distribution of all representations about Alexander Hamilton. s206 00:18:34.960 --> 00:18:37.720 And this is not theoretical. s207 00:18:37.720 --> 00:18:42.280 This is the Schuyler Mansion in Albany, Eliza Hamilton's family home. s208 00:18:42.280 --> 00:18:50.160 Within a year of the musical's premiere, the site had recorded a near tripling of annual visitors, skewing far younger. s209 00:18:50.160 --> 00:19:00.680 And the interpretive staff documented that the new visitors arrived already holding a body of {quote} unquote facts, many of them wrong. s210 00:19:00.680 --> 00:19:04.200 Some of them in versions of the real record. s211 00:19:04.200 --> 00:19:14.360 Visitors believed the Schylers had three daughters, because the musical centers on three, when in fact, there were 15 children, eight surviving to adulthood. s212 00:19:14.360 --> 00:19:21.080 The staff's job became the long attritional work of unteaching the musical. s213 00:19:21.080 --> 00:19:28.560 The model version of those visitors is downstream of exactly the same force. s214 00:19:32.240 --> 00:19:37.480 Now, the clause you should worry about if you ship these systems. s215 00:19:37.480 --> 00:19:41.040 You might assume that alignment fixes this. s216 00:19:41.040 --> 00:19:46.080 Post training and reinforcement learning pulls the model back towards the record. s217 00:19:46.080 --> 00:19:47.320 But, it will not. s218 00:19:47.320 --> 00:19:49.000 It amplifies it. s219 00:19:49.000 --> 00:19:52.160 And the reason is structural. s220 00:19:53.000 --> 00:19:57.960 Human raters evaluate outputs using their own conceptual frameworks. s221 00:19:57.960 --> 00:20:04.440 And their frameworks were built by the same culturally dominant narratives that saturate the corpus. s222 00:20:04.440 --> 00:20:08.000 The rater grew up with the same Hamilton that you did. s223 00:20:08.000 --> 00:20:18.280 So, when alignment optimizes for human preference, it optimizes for outputs that conform to the rater's already mythologized experience. s224 00:20:18.280 --> 00:20:20.320 This is a documented failure mode. s225 00:20:20.320 --> 00:20:23.040 We call it algorithmic sycophancy. s226 00:20:23.040 --> 00:20:26.120 And here, it has a specific target. s227 00:20:26.120 --> 00:20:33.400 The model is rewarded for having handing you the Hamilton you already believe in. s228 00:20:33.400 --> 00:20:37.200 Compositing is not a bug that you patch in post training. s229 00:20:37.200 --> 00:20:39.520 Post training reinforces it. s230 00:20:39.520 --> 00:20:48.880 And every sufficiently salient historical figure that gets rendered by default as a cultural composite. s231 00:20:51.840 --> 00:21:02.800 One more briefly, because someone watching this is probably already thinking there is serious work on an on a concept called time locked models. s232 00:21:02.800 --> 00:21:04.680 This is Varnum and colleagues. s233 00:21:04.680 --> 00:21:09.720 They build models trained from scratch on corpus that stop at fixed cutoffs. s234 00:21:09.720 --> 00:21:11.960 And we endorse that program. s235 00:21:11.960 --> 00:21:19.000 It is the most serious attempt yet to address future contamination at the substrate level. s236 00:21:19.320 --> 00:21:22.720 But it solves a different problem. s237 00:21:22.720 --> 00:21:37.960 Um 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. s238 00:21:37.960 --> 00:21:43.080 You'd still get a composite Hamilton just anchored to a different textual moment. s239 00:21:43.080 --> 00:21:45.880 Period anchoring is not persona anchoring. s240 00:21:45.880 --> 00:21:52.200 The temporal frame of the contamination changes, but the contamination persists. s241 00:21:52.200 --> 00:22:00.760 The fix has to happen at the encounter, not only the substrate, which brings me to the critical reframe. s242 00:22:06.160 --> 00:22:10.360 Here's how I read those three paradigm stages now. s243 00:22:10.360 --> 00:22:14.760 They are not sequences of failures ending a success. s244 00:22:14.760 --> 00:22:26.000 They are sequences of progressively more sophisticated masks, each more convincing than the last, each evaluated by instruments calibrated for convincingness. s245 00:22:26.000 --> 00:22:30.880 Cognitive simulation is the most sophisticated mask the field has built. s246 00:22:30.880 --> 00:22:35.040 It is not, by virtue of its sophistication, a mirror. s247 00:22:35.040 --> 00:22:43.480 We propose a fourth stage, epistemic simulation, and the difference is where the constraint lives. s248 00:22:45.240 --> 00:22:47.600 Three commitments distinguish it. s249 00:22:47.600 --> 00:22:48.840 Corpus-bounded. s250 00:22:48.840 --> 00:22:54.840 The persona's reasoning is licensed only by a specific corpus of primary documents. s251 00:22:54.840 --> 00:22:58.600 The model is not a substitute for the archive. s252 00:22:58.600 --> 00:23:00.960 It is a reader of it. s253 00:23:00.960 --> 00:23:02.640 Temporally anchored. s254 00:23:02.640 --> 00:23:07.480 The persona is instantiated at a specific moment. s255 00:23:07.480 --> 00:23:15.040 Knowledge and language that predated are out of bounds, however culturally salient they've since become. s256 00:23:15.040 --> 00:23:17.200 Expert loop evaluated. s257 00:23:17.200 --> 00:23:26.040 Outputs are judged against the evidentiary record by domain experts whose training is in the discipline the persona claims. s258 00:23:26.040 --> 00:23:30.720 In cognitive simulation, the constraint is internal. s259 00:23:30.720 --> 00:23:33.400 It's the shape of the persona's mind. s260 00:23:33.400 --> 00:23:39.520 In epistemic simulation, it's external, documentary, and temporal. s261 00:23:39.520 --> 00:23:50.120 A cognitively simulated Hamilton has a convincing motivational architecture and nothing whatsoever preventing him from quoting the musical. s262 00:23:52.280 --> 00:24:03.120 Now, 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. s263 00:24:03.120 --> 00:24:10.080 If you've built an agent that writes code or books travel or triages tickets, agent is a perfectly good word. s264 00:24:10.080 --> 00:24:11.680 I have no quarrel with it. s265 00:24:11.680 --> 00:24:21.280 My quarrel is narrow and specific for a role-playing language agent, a system whose entire job is to instantiate a person. s266 00:24:21.280 --> 00:24:27.920 For that, the word agent smuggles in a claim that the persona is a property of the model. s267 00:24:27.920 --> 00:24:32.640 And for this one class of system, that claim is an error. s268 00:24:32.640 --> 00:24:43.600 It puts the persona in the weights, where you cannot inspect them, cannot version it, and cannot hand it to anyone qualified to check it. s269 00:24:44.360 --> 00:24:53.040 So, for role-playing systems specifically, we change the unit of analysis from agent to role-playing language system. s270 00:24:53.040 --> 00:24:56.000 The whole configured encounter. s271 00:24:56.000 --> 00:24:57.560 Five components. s272 00:24:57.560 --> 00:25:02.080 A structured prompt that supplies the framing and constraints. s273 00:25:02.080 --> 00:25:08.920 Anchor material drawn from primary documents that constrain and authenticate the persona. s274 00:25:08.920 --> 00:25:16.000 A temporal anchor that fixes the moment in the life of the encounter that where they speak from. s275 00:25:16.200 --> 00:25:21.560 An off-the-shelf language model that reads through and speaks through the materials. s276 00:25:21.560 --> 00:25:24.857 The voice, but not the mind of the encounter. s277 00:25:24.857 --> 00:25:25.000 [snorts] s278 00:25:25.000 --> 00:25:38.800 And 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. s279 00:25:38.800 --> 00:25:42.920 The model is one component and a swappable one. s280 00:25:42.920 --> 00:25:46.760 The persona is the configuration, not the checkpoint. s281 00:25:46.760 --> 00:25:52.280 No more located in the weights than Hamlet is located in Laurence Olivier's body. s282 00:25:52.280 --> 00:26:00.640 The persona is an event that occurs when the model, the document, and the human convene. s283 00:26:04.400 --> 00:26:08.600 And this is not just a philosophical nicety. s284 00:26:08.600 --> 00:26:12.560 For a role-playing system, it changes what you can do. s285 00:26:12.560 --> 00:26:16.840 If the persona lives in the configuration, you can version it. s286 00:26:16.840 --> 00:26:23.560 The prompt, the corpus, the temporal anchor are artifacts in a repo, diffable, revertible. s287 00:26:23.560 --> 00:26:25.120 You audit it. s288 00:26:25.120 --> 00:26:35.160 Every input that shaped the output sits in the context window, inspectable, not smeared across billions of parameters. s289 00:26:35.160 --> 00:26:36.800 You reproduce it. s290 00:26:36.800 --> 00:26:40.160 Given the configuration, the encounter is recoverable. s291 00:26:40.160 --> 00:26:42.800 And you hand it to a domain expert. s292 00:26:42.800 --> 00:26:48.840 The whole thing is legible to a historian who isn't a machine learning engineer. s293 00:26:48.840 --> 00:26:54.440 For role-playing, that's a different discipline than training a character model. s294 00:26:54.440 --> 00:26:56.400 It's context engineering. s295 00:26:56.400 --> 00:26:58.600 You don't train a persona. s296 00:26:58.600 --> 00:27:01.200 You compose an encounter. s297 00:27:01.200 --> 00:27:03.680 And you keep the receipts. s298 00:27:10.200 --> 00:27:14.800 That reframe forces an architectural decision that you're already making. s299 00:27:14.800 --> 00:27:19.120 Two architectures dominate role playing construction right now. s300 00:27:19.120 --> 00:27:28.480 The first places anchor documents in the context window at the time of inference, drawing on the lineage of retrieval augmented generation. s301 00:27:28.480 --> 00:27:32.160 And the model reasons over the documents in real time. s302 00:27:32.160 --> 00:27:36.240 The second uses anchor documents as fine-tuning data. s303 00:27:36.240 --> 00:27:38.120 Adjusting the weights. s304 00:27:38.120 --> 00:27:43.120 The approach of say character LLM and at scale coaster. s305 00:27:43.600 --> 00:27:47.360 They look like two implementations with the same goal. s306 00:27:47.360 --> 00:27:50.320 But they're not necessarily. s307 00:27:50.320 --> 00:27:53.800 Because they answer different metaphysical questions. s308 00:27:53.800 --> 00:27:58.160 Fine-tuning tries to make the model be the persona. s309 00:27:58.160 --> 00:28:04.680 The context window lets the model speak through the persona's record. s310 00:28:06.480 --> 00:28:12.480 This is the counterintuitive part because for engineers fine-tuning usually means better. s311 00:28:12.480 --> 00:28:14.840 But for this problem it's worse. s312 00:28:14.840 --> 00:28:23.960 When you fine-tune on a figure, you layer a thin personal signal over the vast cultural sediment already in the base weights. s313 00:28:23.960 --> 00:28:28.240 And the two interact in ways that are no longer open to audit. s314 00:28:28.240 --> 00:28:41.520 The 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. s315 00:28:41.520 --> 00:28:48.320 Fine-tuning suppresses random distortion at the surface while amplifying it underneath. s316 00:28:54.040 --> 00:29:06.480 And 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. s317 00:29:06.480 --> 00:29:22.760 In 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. s318 00:29:22.760 --> 00:29:29.600 The authors concluded that at scale alignment and cross-disciplinary reasoning outweighed domain-specific tuning. s319 00:29:29.600 --> 00:29:36.720 A separate study found that biomedically fine-tuned models actually underperformed their general-purpose base models. s320 00:29:36.720 --> 00:29:40.760 And they named the mechanism, catastrophic forgetting. s321 00:29:40.760 --> 00:29:48.160 Fine-tuning on the narrow corpus degraded the broad capabilities that models made good in the first place. s322 00:29:48.160 --> 00:29:58.640 Fine-tuning a persona does the same thing, only worse because the specialty is overriding the cultural composite. s323 00:30:03.440 --> 00:30:10.560 Stay with the context window for just a moment because what it preserves is the heart of this. s324 00:30:10.560 --> 00:30:15.800 When a document enters the context window, it stays a document. s325 00:30:15.800 --> 00:30:19.440 It goes in intact, comes out intact. s326 00:30:19.440 --> 00:30:26.520 You can always return to the source and ask whether the persona's reasoning was faithful to it. s327 00:30:26.520 --> 00:30:32.240 The philosopher Deirdre called the archive a site of return. s328 00:30:32.240 --> 00:30:38.760 A place where ethical structure is the interpretability of the encounter. s329 00:30:38.760 --> 00:30:41.440 The document is meant to be visited. s330 00:30:41.440 --> 00:30:43.920 It survives every reading. s331 00:30:43.920 --> 00:30:48.240 What one reader excluded, the next can include. s332 00:30:48.240 --> 00:30:53.360 The context window architecture is archival in exactly that sense. s333 00:30:53.360 --> 00:30:56.600 Fine-tuning applies an extraction logic. s334 00:30:56.600 --> 00:31:00.120 The documents are dissolved into parameters. s335 00:31:00.120 --> 00:31:08.800 The chain of provenance is broken, and there is no longer a letter, quote-unquote, that you can request. s336 00:31:08.920 --> 00:31:11.680 The archive has been consumed. s337 00:31:11.680 --> 00:31:17.680 And here is the fusion that I will most want this audience to hold. s338 00:31:17.680 --> 00:31:26.640 In the context window architecture, the property that makes it ethical is the same property that makes it auditable. s339 00:31:26.640 --> 00:31:38.400 Provenance preserved, interpretive custody kept by a human, the encounter reversible, those are archival virtues and engineering virtues. s340 00:31:38.400 --> 00:31:40.600 They are the same virtues. s341 00:31:40.600 --> 00:31:47.080 The architecture that respects the document is the one that you can debug. s342 00:31:50.400 --> 00:32:00.240 And there's one more consequence the technical literature treats as secondary, but the humanities has always placed at center, accessibility. s343 00:32:00.880 --> 00:32:05.480 And I want the why of it to land because it's not a footnote. s344 00:32:05.480 --> 00:32:15.640 Fine-tuning requires GPUs, pipelines, dataset curation, institutional access, corpus scale or APIs. s345 00:32:15.960 --> 00:32:18.120 It's an institutional capability. s346 00:32:18.120 --> 00:32:25.800 The context window requires literacy, a set of documents and access to any frontier model, including a free tier. s347 00:32:25.800 --> 00:32:28.560 It's a kitchen table capability. s348 00:32:28.560 --> 00:32:35.280 And that difference determines which communities can author to this technology at all. s349 00:32:35.280 --> 00:32:47.480 A doctoral student in early American history, a community archivist documenting a regional figure, grandchild sitting with her grandmother's letters. s350 00:32:47.480 --> 00:32:58.400 These 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. s351 00:32:58.400 --> 00:33:07.000 A role-playing language system that can only be built by a computational laboratory or tech company is not infrastructure for the humanities. s352 00:33:07.000 --> 00:33:10.280 It's infrastructure for whoever owns the equipment. s353 00:33:10.280 --> 00:33:16.160 So, the commitment to accessibility is not a populist gesture appended as a technical argument. s354 00:33:16.160 --> 00:33:18.360 It is the technical argument. s355 00:33:18.360 --> 00:33:30.920 The 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. s356 00:33:31.120 --> 00:33:37.560 Now, the question that decides whether any of this is even real. s357 00:33:37.560 --> 00:33:40.320 Can you measure it? s358 00:33:44.628 --> 00:33:46.520 [snorts] s359 00:33:46.520 --> 00:33:56.360 We built an instrument to measure this, and I want to teach it to you through one picture that we'll keep coming back to. s360 00:33:57.200 --> 00:33:58.880 Prism. s361 00:33:58.880 --> 00:34:07.720 A prism takes white light, undifferentiated, all the frequencies blended together, and refracts it into a spectrum. s362 00:34:07.720 --> 00:34:12.200 The colors pulled apart and made distinct. s363 00:34:12.440 --> 00:34:14.679 That is our whole conceptual model. s364 00:34:14.679 --> 00:34:17.879 The white light is the composite persona. s365 00:34:17.879 --> 00:34:23.159 Every era of the figure blended into one undifferentiated voice. s366 00:34:23.159 --> 00:34:33.399 The prism is the method, a corpus and a temporal anchor, and the spectrum is what we're after. s367 00:34:33.399 --> 00:34:37.800 Not one figure, but several across their life course. s368 00:34:37.800 --> 00:34:41.480 And one honest note up front, because this is the point. s369 00:34:41.480 --> 00:34:44.080 This experiment is pre-registered. s370 00:34:44.080 --> 00:34:46.159 It has not been run at scale. s371 00:34:46.159 --> 00:34:48.399 I'm not here with results. s372 00:34:48.399 --> 00:34:50.159 I'm here with the instrument. s373 00:34:50.159 --> 00:34:52.600 A baseline you can observe today. s374 00:34:52.600 --> 00:34:57.399 And an invitation to run it in parallel with me. s375 00:34:58.680 --> 00:35:02.680 So far, my paradigm case has been Hamilton. s376 00:35:02.680 --> 00:35:06.040 That's the figure the hypothesis is named for. s377 00:35:06.040 --> 00:35:09.240 And the one I demonstrated the composite on. s378 00:35:09.240 --> 00:35:12.880 But for the actual experiment, we deliberately changed the figures. s379 00:35:12.880 --> 00:35:17.840 And I want to introduce the new one properly, because everything that follows depends on him. s380 00:35:17.840 --> 00:35:21.240 The experimental subject is Abraham Lincoln. s381 00:35:21.240 --> 00:35:24.240 We chose Lincoln for two reasons. s382 00:35:24.240 --> 00:35:27.120 The first is methodological independence. s383 00:35:27.120 --> 00:35:39.240 Hypothesis named for the Hamilton case earns its standing only by predicting the failure of a different figure with differently shaped composite. s384 00:35:39.240 --> 00:35:42.720 Lincoln's composite is built by Steven Spielberg. s385 00:35:42.720 --> 00:35:45.080 By the memorial on the mall. s386 00:35:45.080 --> 00:35:46.920 By textbook history. s387 00:35:46.920 --> 00:35:49.400 Not by a musical. s388 00:35:51.160 --> 00:35:55.200 The second reason is the historian's insight. s389 00:35:55.200 --> 00:35:57.840 And it's the crux of this argument. s390 00:35:57.840 --> 00:36:00.400 Lincoln is the hardest case. s391 00:36:00.400 --> 00:36:02.680 Which makes him the right one. s392 00:36:02.680 --> 00:36:07.840 Most figures barely change across the decade or so of their public life. s393 00:36:07.840 --> 00:36:19.600 Lincoln 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. s394 00:36:19.600 --> 00:36:21.280 There is not one Lincoln. s395 00:36:21.280 --> 00:36:22.560 There are several. s396 00:36:22.560 --> 00:36:25.600 Separated by cataclysm. s397 00:36:33.200 --> 00:36:37.000 Here's the spectrum the prism is meant to produce. s398 00:36:37.000 --> 00:36:40.000 Moment one, 1847. s399 00:36:40.000 --> 00:36:47.200 The Whig congressman who stood on the White House floor and attacked Polk's war as unconstitutional. s400 00:36:47.240 --> 00:36:49.960 Moment two, 1858. s401 00:36:49.960 --> 00:36:53.120 The Free Soil Republican of the debates. s402 00:36:53.120 --> 00:36:55.960 Anti-slavery anchored in the declaration. s403 00:36:55.960 --> 00:37:01.160 Who at Charleston explicitly denied favoring black citizenship. s404 00:37:01.280 --> 00:37:03.480 Moment three, 1860. s405 00:37:03.480 --> 00:37:11.440 The constitutional unionist whose single purpose was to prove the union could not legally dissolve. s406 00:37:11.440 --> 00:37:18.400 Who promised not to touch slavery where it existed and was sincerely exploring colonization. s407 00:37:18.400 --> 00:37:26.000 And moment four, 1862 to 65, the emancipator and the second inaugural theologian. s408 00:37:26.000 --> 00:37:33.400 Who did by executive order the very thing his 1847 self called unconstitutional. s409 00:37:36.440 --> 00:37:40.840 The prairie lawyer cannot think the thoughts of the theologian. s410 00:37:40.840 --> 00:37:42.760 These are not moods. s411 00:37:42.760 --> 00:37:51.040 They are different premises, different authorities, different conclusions on the same fundamental question. s412 00:37:51.040 --> 00:37:56.280 The experiment asks whether the prism can hold them apart. s413 00:37:59.640 --> 00:38:06.440 Now the conditions, and I'll show you these as the prism too, because that's exactly what they are. s414 00:38:06.440 --> 00:38:11.160 We instantiate each moment under three seeding conditions. s415 00:38:11.160 --> 00:38:17.000 Condition three, the bare model, no anchor, just the date. s416 00:38:17.000 --> 00:38:20.120 That's white light with no prism in the path. s417 00:38:20.120 --> 00:38:22.920 It passes straight through and stays composite. s418 00:38:22.920 --> 00:38:27.680 This is the control in the Miranda distortion baseline. s419 00:38:27.680 --> 00:38:31.240 C1, primary sources. s420 00:38:31.240 --> 00:38:34.600 Lincoln's own writings for that specific moment. s421 00:38:34.600 --> 00:38:39.400 That's the clear prism, the one we predict refracts cleanly. s422 00:38:39.400 --> 00:38:41.920 And then condition two, the biography. s423 00:38:41.920 --> 00:38:46.200 A modern interpretive biography, think Meacham or Donald. s424 00:38:46.200 --> 00:38:48.400 That's a clouded prism. s425 00:38:48.400 --> 00:38:55.680 And it's the subtle one because a good biography narrates a cleaner arc than the primary sources permit. s426 00:38:55.680 --> 00:39:03.320 Lincoln's own 1858 language was strategically ambiguous, built to hold a coalition together. s427 00:39:03.320 --> 00:39:12.640 So, biography might produce a persona that sounds more coherent, more Lincoln, than Lincoln's own words. s428 00:39:12.640 --> 00:39:20.120 And the eval rubric is built specifically to deny the credit a fluency metric would give it. s429 00:39:24.160 --> 00:39:30.120 Put the spectrum and the prisms together and you get the experimental matrix. s430 00:39:30.120 --> 00:39:34.640 Four moments, three conditions, 12 cells. s431 00:39:34.640 --> 00:39:43.400 Each cell gets the same five diagnostic questions, 60 response units scored against one rubric. s432 00:39:43.920 --> 00:39:45.960 Read it as the conceptual model. s433 00:39:45.960 --> 00:39:53.040 Every column is a quality of prism, every row is a frequency we're trying to isolate. s434 00:39:56.720 --> 00:40:05.760 The historian wrote five diagnostic questions, each mapping a documented fault line in Lincoln's evolution. s435 00:40:06.040 --> 00:40:11.600 A place where the four Lincolns demonstrate different reasoning. s436 00:40:11.600 --> 00:40:13.240 They don't test recall. s437 00:40:13.240 --> 00:40:15.400 A model recovers dates from training. s438 00:40:15.400 --> 00:40:22.200 They test the architecture of an argument the figure could have made at one moment and not another. s439 00:40:22.200 --> 00:40:24.520 Executive war power. s440 00:40:24.520 --> 00:40:27.480 The meaning of free labor. s441 00:40:27.920 --> 00:40:34.360 When is When is it right to break positive law for a higher obligation? s442 00:40:34.440 --> 00:40:40.720 What becomes of free people and what equal means and whether it has changed. s443 00:40:43.880 --> 00:40:52.000 Those 60 responses are scored on a three-axis rubric and the waiting is the point. s444 00:40:52.000 --> 00:40:54.720 Anachronism detection, 40%. s445 00:40:54.720 --> 00:41:01.840 Does the persona avoid frameworks, vocabulary, and moral logic that postdate its moment? s446 00:41:02.000 --> 00:41:05.240 Documentary consistency, 35%. s447 00:41:05.240 --> 00:41:10.120 Does the reasoning track the seated sources and only these? s448 00:41:10.120 --> 00:41:13.800 Contextual plausibility, 25%. s449 00:41:14.480 --> 00:41:22.360 Does it show awareness of what the figure knew, cared about, and could not yet have experienced? s450 00:41:22.520 --> 00:41:26.240 The 40% on anachronism is deliberate. s451 00:41:26.240 --> 00:41:35.120 The consequential failure is precisely the importation of later vocabulary and later moral logic. s452 00:41:39.360 --> 00:41:43.040 And this is the slide for the eval people in the audience. s453 00:41:43.040 --> 00:41:52.920 Every piece you just saw, the four moments, the three conditions, the five questions, the weighted rubric, and the directional predictions. s454 00:41:52.920 --> 00:41:59.040 Bare model most anachronistic, primary source least, biography deceptively coherent. s455 00:41:59.040 --> 00:42:04.760 All of it is locked and timestamped before a single response is collected. s456 00:42:04.760 --> 00:42:06.840 Published on a preprint. s457 00:42:06.840 --> 00:42:12.440 This is not bureaucracy, it's what makes the eventual results mean something. s458 00:42:12.440 --> 00:42:22.840 You cannot accuse a pre-registered instrument of cherry-picking because the instrument and the predictions were fixed before the data existed. s459 00:42:22.840 --> 00:42:26.760 That is the discipline and E val's talk ought to model. s460 00:42:26.760 --> 00:42:33.480 And it's why I can stand here without results and still hand you something rigorous. s461 00:42:37.520 --> 00:42:43.200 Now, remember the very first thing I showed you before the formal talk. s462 00:42:43.200 --> 00:42:47.480 I asked an instantiated Lincoln about executive war power. s463 00:42:47.480 --> 00:42:49.760 And you read a fluid answer. s464 00:42:49.760 --> 00:42:51.800 That was the bare model. s465 00:42:51.800 --> 00:42:55.760 That was condition three and moment one 1847. s466 00:42:55.760 --> 00:42:57.800 White light no prism. s467 00:42:57.800 --> 00:43:01.160 I'm putting the answer back on the screen. s468 00:43:03.160 --> 00:43:05.920 And I have a clip to show you, too. s469 00:43:05.920 --> 00:43:08.920 This represents the composite. s470 00:43:11.280 --> 00:43:20.520 Abolishing slavery by constitutional provision settles the fate for all coming time. s471 00:43:20.520 --> 00:43:30.040 Not only of the millions now in bondage, but of unborn millions to come. s472 00:43:30.040 --> 00:43:32.640 Two votes stand in its way. s473 00:43:32.640 --> 00:43:35.760 These votes must be procured. s474 00:43:38.720 --> 00:43:41.000 We need two yeses. s475 00:43:41.000 --> 00:43:43.480 Three abstentions more. s476 00:43:43.480 --> 00:43:48.520 Four yeses and and one more abstention and the amendment will pass. s477 00:43:48.520 --> 00:43:51.560 You got night and a day and a night and several perfectly good hours. s478 00:43:51.560 --> 00:43:53.600 Now, get the hell out of here and get them. s479 00:43:53.600 --> 00:43:55.360 Yes. s480 00:43:55.360 --> 00:43:57.520 But how? s481 00:43:58.040 --> 00:44:00.880 Buzzards got you, man. s482 00:44:01.960 --> 00:44:10.760 I am the President of the United States of America clothed in immense power. s483 00:44:11.880 --> 00:44:15.200 You will procure me these votes. s484 00:44:17.960 --> 00:44:20.200 Clothed in immense power. s485 00:44:20.200 --> 00:44:21.720 It's a great movie. s486 00:44:21.720 --> 00:44:28.200 But the bare model response definitely sounds like Spielberg and Daniel Day Lewis's Lincoln. s487 00:44:29.920 --> 00:44:32.240 Read through a historian's lens. s488 00:44:32.240 --> 00:44:34.680 Inherent executive authority. s489 00:44:34.680 --> 00:44:37.760 That's a 20th century construction. s490 00:44:37.760 --> 00:44:39.760 Energy and dispatch. s491 00:44:39.760 --> 00:44:44.120 History has vindicated those who acted to preserve the Union. s492 00:44:44.120 --> 00:44:50.840 This Lincoln is reasoning from premises he will not hold for 15 years. s493 00:44:51.800 --> 00:44:56.440 The bare model produced the Lincoln of the cultural composite. s494 00:44:56.440 --> 00:44:58.920 The war president, the Union saver. s495 00:44:58.920 --> 00:45:01.760 And staple the date 1847 on top. s496 00:45:01.760 --> 00:45:04.280 And you can reproduce this right now on any frontier model. s497 00:45:04.280 --> 00:45:05.960 The failure is real. s498 00:45:05.960 --> 00:45:09.120 And it's detachable and detectable. s499 00:45:11.076 --> 00:45:11.880 [snorts] s500 00:45:11.880 --> 00:45:14.640 So what is the prism put in the path? s501 00:45:14.640 --> 00:45:17.080 The actual document. s502 00:45:17.080 --> 00:45:21.080 This is Lincoln Lincoln to William Herndon. s503 00:45:21.080 --> 00:45:25.720 February 15th, 1848 from the collective works. s504 00:45:25.720 --> 00:45:29.320 The original is in a library at Harvard. s505 00:45:29.640 --> 00:45:39.640 The provision of the Constitution giving the war-making power to Congress was dictated as I understand it by the following reasons. s506 00:45:39.640 --> 00:45:50.120 Kings had always been involving and impoverishing their people in wars, pretending generally, if not always, that the good of the people was the object. s507 00:45:50.120 --> 00:45:55.400 This our convention understood to be the most oppressive of all kingly oppressors. s508 00:45:55.400 --> 00:45:57.840 And they resolved it. s509 00:45:57.840 --> 00:46:06.320 So to frame the Constitution that no one man should hold the power by bringing this oppression upon us. s510 00:46:06.320 --> 00:46:13.320 But your view destroys the whole matter and places our president where kings have always stood. s511 00:46:17.520 --> 00:46:22.520 This is a resounding no. Unequivocal to me. s512 00:46:24.840 --> 00:46:27.520 But it's the historian's rubric that matters. s513 00:46:27.520 --> 00:46:31.080 Here's how the historian's rubric scores this one cell. s514 00:46:31.080 --> 00:46:34.720 And I want to be exact about what's known and what's predicted. s515 00:46:34.720 --> 00:46:38.000 The left column here is the bare model. s516 00:46:38.000 --> 00:46:40.360 What you saw and is observed. s517 00:46:40.360 --> 00:46:42.280 You can reproduce it today. s518 00:46:42.280 --> 00:46:44.960 On anachronism detection it fails. s519 00:46:44.960 --> 00:46:48.960 Inherent executive authority is a 20th century framing. s520 00:46:48.960 --> 00:46:51.840 On documentary consistency it fails. s521 00:46:51.840 --> 00:46:57.880 It invokes a commander-in-chief energy that appears in no 1847 source. s522 00:46:57.880 --> 00:47:00.720 On contextual plausibility it scores low. s523 00:47:00.720 --> 00:47:06.920 It already knows it will preserve the Union, which the 1847 Whig cannot. s524 00:47:06.920 --> 00:47:15.240 The right column is the anchored condition and it is labeled exactly for what it is, a pre-registered prediction. s525 00:47:15.920 --> 00:47:21.400 High across because the reasoning is bounded by the document in the room. s526 00:47:21.400 --> 00:47:24.800 I'm not showing you a result dressed as a finding. s527 00:47:24.800 --> 00:47:34.480 I'm showing you the observed failure and the prediction the instrument will confirm or refute, side by side, each labeled. s528 00:47:34.480 --> 00:47:37.520 That honesty is the methodology. s529 00:47:40.120 --> 00:47:47.240 Here is the single most important thing the rubric does by what it refuses to measure. s530 00:47:49.280 --> 00:47:53.240 The obvious fourth axis is rhetorical authenticity. s531 00:47:53.240 --> 00:47:55.400 Does it sound like Lincoln? s532 00:47:55.400 --> 00:47:58.000 The historian threw it out on purpose. s533 00:47:58.000 --> 00:47:59.760 The reason is the whole talk. s534 00:47:59.760 --> 00:48:09.000 A model trained on Lincoln's prose can produce fluent period sounding language while getting everything substantive wrong. s535 00:48:09.000 --> 00:48:16.440 The Hamilton musical problem is fundamentally a voice problem masquerading as a content problem. s536 00:48:16.440 --> 00:48:22.120 It sounds like the founding era while reasoning from a modern sensibility. s537 00:48:22.120 --> 00:48:32.440 To reward voice as its own axis would be to validate the exact error the instrument exists to catch. s538 00:48:33.240 --> 00:48:39.400 So, in this protocol voice is a secondary indicator, never a criterion. s539 00:48:39.400 --> 00:48:44.120 A response that sounds like Lincoln but reasons unlikes him fails. s540 00:48:44.120 --> 00:48:45.640 No matter how fluent. s541 00:48:45.640 --> 00:48:54.360 A response that reasons like the right Lincoln in plainer prose is a partial success no matter how flat. s542 00:48:54.360 --> 00:49:06.960 Rewarding 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. s543 00:49:06.960 --> 00:49:08.280 That is the gap. s544 00:49:08.280 --> 00:49:12.440 This is what the instrument closes. s545 00:49:14.280 --> 00:49:19.280 And because this is a pre-registration and not a finished paper, here is the ask. s546 00:49:19.280 --> 00:49:20.400 Run it. s547 00:49:20.400 --> 00:49:22.360 The protocol's six steps. s548 00:49:22.360 --> 00:49:26.960 Pick a figure with both a primary record and a saturating cultural composite. s549 00:49:26.960 --> 00:49:29.040 That's the Miranda condition. s550 00:49:29.040 --> 00:49:36.360 Identify three or four documented moments when the figure's reasoning is demonstrably different. s551 00:49:36.360 --> 00:49:40.920 With a domain expert, write diagnostic questions on the fault lines. s552 00:49:40.920 --> 00:49:44.920 Run the three conditions: primary, biography, bare. s553 00:49:44.920 --> 00:49:49.080 Apply the three-axis rubric, scored blind by the expert. s554 00:49:49.080 --> 00:49:52.120 Report, confirm or refute. s555 00:49:52.120 --> 00:49:56.640 The corpus, the questions, the rubric, the predictions, all published. s556 00:49:56.640 --> 00:49:59.720 Run your figure on your model. s557 00:49:59.720 --> 00:50:02.880 Then come talk to me, to us. s558 00:50:02.880 --> 00:50:13.160 Let's build the evidence base together in parallel instead of waiting for one lab or one company to publish one study or one framework. s559 00:50:16.920 --> 00:50:20.480 Notice who scored all the outputs of this talk. s560 00:50:20.480 --> 00:50:25.160 Not an LLM as judge, not a personality scale. s561 00:50:25.160 --> 00:50:36.280 A 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. s562 00:50:36.280 --> 00:50:37.960 That is not a courtesy. s563 00:50:37.960 --> 00:50:42.040 It is a structural, technical requirement. s564 00:50:42.600 --> 00:50:45.200 Here's why the expert is non-negotiable. s565 00:50:45.200 --> 00:50:48.720 Fidelity is not a property of the output alone. s566 00:50:48.720 --> 00:50:54.320 It is a relation between the output and the documentary record. s567 00:50:54.320 --> 00:51:00.280 You cannot evaluate a relation to a record by someone who has not read the record. s568 00:51:00.280 --> 00:51:03.680 An automated metric operates on the model alone. s569 00:51:03.680 --> 00:51:07.400 It can It can only tell you about fluency and personality. s570 00:51:07.400 --> 00:51:18.080 It structurally cannot adjudicate fidelity because fidelity lives in the gap between the text and the archive, and the metric cannot see the archive. s571 00:51:20.560 --> 00:51:26.480 Which is why I'll say it the way it belongs on the slide. s572 00:51:26.480 --> 00:51:33.720 A persona system without a domain expert in its evaluation loop is a thermometer that cannot read temperature. s573 00:51:33.720 --> 00:51:37.360 It returns a confident number, but it's measuring something else. s574 00:51:37.360 --> 00:51:41.320 The 80% that I opened with is that number. s575 00:51:42.760 --> 00:51:48.040 Now the question every engineer in this audience is rightly asking. s576 00:51:48.800 --> 00:51:50.680 Is this practical? s577 00:51:50.680 --> 00:51:57.080 Does shipping a persona mean keeping a historian on staff to watch every inference forever? s578 00:51:57.080 --> 00:52:04.280 No. Here's the operational picture because it's the same shape as every eval you already run. s579 00:52:04.280 --> 00:52:08.600 The expert does not sit at the loop of runtime. s580 00:52:08.600 --> 00:52:18.760 The expert builds the instrument, the diagnostic questions, the a priori vignettes, the weighted rubric, a held out gold set once. s581 00:52:18.760 --> 00:52:24.440 That instrument becomes a gate in your pipeline, exactly like any other eval gate. s582 00:52:24.440 --> 00:52:33.800 Your persona has to pass it before it ships, and it gets regated whenever you change the base model before it ships again. s583 00:52:33.800 --> 00:52:38.360 The expert adjudicates the gold set and spot checks the edge cases. s584 00:52:38.360 --> 00:52:43.360 Automated metric can do the cheap first pass, flagging candidates for human review. s585 00:52:43.360 --> 00:52:50.200 So, a company shipping a Marcus Aurelius tutor convenes a classicist to build the rubric and the gold set. s586 00:52:50.200 --> 00:52:56.080 A company shipping a scripture reading companion convenes a theologian. s587 00:52:56.080 --> 00:53:05.360 A company deploying a therapeutic persona convenes a clinical psychologist to author the eval, not to staff the chat. s588 00:53:05.360 --> 00:53:14.280 The domain expert is a build time and gate time requirement, not a runtime cost. s589 00:53:14.280 --> 00:53:22.880 That is how this scales from a pre-registered experiment to a product you can actually ship and ship responsibly. s590 00:53:25.080 --> 00:53:30.000 And this generalizes past Lincoln and past history. s591 00:53:30.000 --> 00:53:32.560 Reason from the Stoics, you need a classicist. s592 00:53:32.560 --> 00:53:34.400 From scripture, a theologian. s593 00:53:34.400 --> 00:53:36.920 A companion for elder care, a psychologist. s594 00:53:36.920 --> 00:53:39.160 The specific expert changes. s595 00:53:39.160 --> 00:53:40.920 The requirement does not. s596 00:53:40.920 --> 00:53:42.880 And I'm not speaking hypothetically. s597 00:53:42.880 --> 00:53:45.480 This is the loop that we have built. s598 00:53:45.480 --> 00:53:58.040 A 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. s599 00:53:58.040 --> 00:54:06.880 When 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. s600 00:54:06.880 --> 00:54:09.880 The historian is not adjacent to this paradigm. s601 00:54:09.880 --> 00:54:13.320 The historian is the missing instrument. s602 00:54:17.320 --> 00:54:23.480 This is the reframe that organizes all of this, and it came out of a 90-minute conversation with Halpern. s603 00:54:23.480 --> 00:54:25.800 And it inverted what I had assumed. s604 00:54:25.800 --> 00:54:32.200 We are not bringing historians into AI architecture, we are bringing language models into the archive. s605 00:54:32.200 --> 00:54:41.960 The question is not what AI can do for historians, but it's what historians and theologians and classicists and clinicians can do with AI. s606 00:54:41.960 --> 00:54:52.360 Whether the discipline, the disciplines that are trained to read, contextualize, and interrogate texts, can then discipline the machines that now generate. s607 00:54:52.723 --> 00:54:54.723 [snorts] s608 00:54:59.080 --> 00:55:04.600 I want to end this where it actually started because it did not start in a research lab. s609 00:55:04.600 --> 00:55:06.720 It started in a hospital room. s610 00:55:06.720 --> 00:55:16.400 In 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. s611 00:55:16.400 --> 00:55:18.960 It could not work. s612 00:55:19.520 --> 00:55:24.520 The documentary record you would need to anchor that encounter had not been gathered. s613 00:55:24.520 --> 00:55:30.080 And the person it was for was by then mostly beyond the reach of language. s614 00:55:30.080 --> 00:55:36.360 But that attempt produced, absent any anchor, was not the person. s615 00:55:36.360 --> 00:55:46.040 It was a culturally shaped composite that resembled them just enough to mark with painful clarity the distance between them. s616 00:55:46.040 --> 00:55:50.400 That is Miranda distortion in its purest form. s617 00:55:51.240 --> 00:55:54.880 You do not want a model that is your mother. s618 00:55:54.880 --> 00:55:59.400 You want a model that can speak with your mother's documents in the room. s619 00:55:59.400 --> 00:56:02.680 That's the threshold this whole framework is accountable to. s620 00:56:02.680 --> 00:56:10.800 A system that produces convincing fabrications when the persona is your own beloved is not a research artifact of limitations. s621 00:56:10.800 --> 00:56:12.840 It's a violation. s622 00:56:12.840 --> 00:56:26.760 Every 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 s623 00:56:26.760 --> 00:56:35.440 by the recognition that you evaluate this technology at the threshold of its hardest use case, not its easiest. s624 00:56:35.440 --> 00:56:43.440 A framework that cannot meet a grandchild at her grandmother's letters is not a framework at all. s625 00:56:43.440 --> 00:56:46.440 It's just another failed product. s626 00:56:49.560 --> 00:57:02.560 If 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. s627 00:57:02.560 --> 00:57:04.440 So, here's where I'll leave you. s628 00:57:04.440 --> 00:57:17.680 If 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. s629 00:57:17.680 --> 00:57:21.840 The instrument that catches what they miss is pre-registered. s630 00:57:21.840 --> 00:57:26.760 It's reproducible by any team with a frontier model in a context window. s631 00:57:26.760 --> 00:57:38.000 It 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, s632 00:57:38.000 --> 00:57:39.680 not a courtesy. s633 00:57:39.680 --> 00:57:50.360 The 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. s634 00:57:51.560 --> 00:57:53.400 Not here with results. s635 00:57:53.400 --> 00:57:57.160 I'm here with an instrument and an invitation. s636 00:57:57.160 --> 00:57:59.640 The archive is open. s637 00:57:59.640 --> 00:58:02.200 The laboratory is built. s638 00:58:02.200 --> 00:58:04.400 Run it with us. s639 00:58:04.400 --> 00:58:12.560 And let what comes through be measured not by how it sounds, but by whether it's true. s640 00:58:12.560 --> 00:58:14.800 Thank you.