--- schema: 1 type: video title: >- The Miranda Hypothesis: How Hamilton Poisoned Persona Evals - Jacob E. Thomas, Results Gen description: >- Your persona-eval pipeline rates an Alexander Hamilton simulation at 80% personality fidelity. It is also rating a Hamilton who sounds like he has read his own Broadway musical. The dominant failure mode of every character-based AI system now in production is invisible to LLM-as-judge, personality-scale benchmarks, and behavioral consistency scores because every one of them was built to detect convincingness, and convincingness is exactly what the failure produces. The failure has a name: Miranda distortion. When the volume of cultural representation of a figure in your training corpus outnumbers their primary documentary record by orders of magnitude (and it always does for any culturally salient figure) your persona doesn't speak from the record. It speaks from the smoothed cultural composite. The 2015 Broadway musical has exponentially more representational density in your training data than the 175,000 words of the Federalist Papers. Your evals were not designed to notice this. They were designed to score fluency, personality coherence, and stylistic naturalness... the exact features the composite optimizes. In this talk: - The structural argument: why InCharacter-style benchmarks, CoSER, and PsyMem can hit state-of-the-art on personality fidelity while structurally failing to detect anachronistic reasoning. - The architectural mechanism: why RLHF amplifies Miranda distortion instead of correcting it (raters are themselves products of the same cultural composite). - The framework: a four-stage paradigm shift from cognitive simulation to epistemic simulation (corpus-bounded, temporally-anchored, expert-loop-evaluated). - The instrument: the pre-registered Prism Experiment. Lincoln at four documented temporal moments, three seeding conditions, five diagnostic questions written by a domain historian, and a weighted three-axis rubric (Anachronism Detection, Documentary Consistency, Contextual Plausibility) that catches what automated metrics miss. - The handoff: what a working eval loop looks like when a historian, classicist, theologian, or clinical psychologist sits in it, and why that's a technical requirement, not a cultural courtesy Pre-registered protocol with University of Toronto historian Rick Halpern, paper forthcoming. Reproducible by any team running a frontier model with a context window. If you ship character bots, companion AI, pedagogical agents, historical simulations, or any system where a persona is supposed to reason from a specified record, your evals are measuring the wrong thing. Here is the instrument that catches what they miss. Speakers: - Jacob E. Thomas (Results Generation): Dr. Thomas is an epidemiologist, data scientist, and AI engineer who studies information as a determinant of health. LinkedIn: https://www.linkedin.com/in/jacob-e-thomas-atx/ GitHub: https://github.com/jethomasphd/THE_COMPANION_DOSSIER --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/IJXjTLPzvAU-the-miranda-hypothesis-how-hamilton-the-musical-poisoned-your-persona-evals Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/the-miranda-hypothesis-how-hamilton-poisoned-person-cdf4a2fc2522 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/the-miranda-hypothesis-how-hamilton-poisoned-person-cdf4a2fc2522.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T09:21:11.472Z' language: null children: [] videoId: IJXjTLPzvAU durationSeconds: 3497 publishedAt: '2026-06-25T16:30:06.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 16 endSeconds: 102 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=16' heldReason: rights - ordinal: 1 text: null startSeconds: 112 endSeconds: 204 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=112' heldReason: rights - ordinal: 2 text: null startSeconds: 215 endSeconds: 326 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=215' heldReason: rights - ordinal: 3 text: null startSeconds: 215 endSeconds: 326 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=215' heldReason: rights - ordinal: 4 text: null startSeconds: 470 endSeconds: 560 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=470' heldReason: rights - ordinal: 5 text: null startSeconds: 572 endSeconds: 692 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=572' heldReason: rights - ordinal: 6 text: null startSeconds: 702 endSeconds: 792 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=702' heldReason: rights - ordinal: 7 text: null startSeconds: 803 endSeconds: 897 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=803' heldReason: rights - ordinal: 8 text: null startSeconds: 909 endSeconds: 1024 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=909' heldReason: rights - ordinal: 9 text: null startSeconds: 1034 endSeconds: 1139 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=1034' heldReason: rights - ordinal: 10 text: null startSeconds: 1139 endSeconds: 1151 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=1139' heldReason: rights - ordinal: 11 text: null startSeconds: 1151 endSeconds: 1233 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=1151' heldReason: rights - ordinal: 12 text: null startSeconds: 1244 endSeconds: 1326 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=1244' heldReason: rights - ordinal: 13 text: null startSeconds: 1336 endSeconds: 1425 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=1336' heldReason: rights - ordinal: 14 text: null startSeconds: 1437 endSeconds: 1567 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=1437' heldReason: rights - ordinal: 15 text: null startSeconds: 1579 endSeconds: 1686 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=1579' heldReason: rights - ordinal: 16 text: null startSeconds: 1697 endSeconds: 1803 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=1697' heldReason: rights - ordinal: 17 text: null startSeconds: 1815 endSeconds: 1948 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=1815' heldReason: rights - ordinal: 18 text: null startSeconds: 1958 endSeconds: 2070 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=1958' heldReason: rights - ordinal: 19 text: null startSeconds: 2081 endSeconds: 2193 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=2081' heldReason: rights - ordinal: 20 text: null startSeconds: 2204 endSeconds: 2297 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=2204' heldReason: rights - ordinal: 21 text: null startSeconds: 2307 endSeconds: 2389 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=2307' heldReason: rights - ordinal: 22 text: null startSeconds: 2402 endSeconds: 2499 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=2402' heldReason: rights - ordinal: 23 text: null startSeconds: 2511 endSeconds: 2618 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=2511' heldReason: rights - ordinal: 24 text: null startSeconds: 2630 endSeconds: 2761 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=2630' heldReason: rights - ordinal: 25 text: null startSeconds: 2869 endSeconds: 2880 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=2869' heldReason: rights - ordinal: 26 text: null startSeconds: 3027 endSeconds: 3108 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=3027' heldReason: rights - ordinal: 27 text: null startSeconds: 3120 endSeconds: 3223 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=3120' heldReason: rights - ordinal: 28 text: null startSeconds: 3235 endSeconds: 3353 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=3235' heldReason: rights - ordinal: 29 text: null startSeconds: 3365 endSeconds: 3497 watch: 'https://liquidgold.cc/watch/IJXjTLPzvAU?t=3365' heldReason: rights --- # The Miranda Hypothesis: How Hamilton Poisoned Persona Evals - Jacob E. Thomas, Results Gen Your persona-eval pipeline rates an Alexander Hamilton simulation at 80% personality fidelity. It is also rating a Hamilton who sounds like he has read his own Broadway musical. The dominant failure mode of every character-based AI system now in production is invisible to LLM-as-judge, personality-scale benchmarks, and behavioral consistency scores because every one of them was built to detect convincingness, and convincingness is exactly what the failure produces. The failure has a name: Miranda distortion. When the volume of cultural representation of a figure in your training corpus outnumbers their primary documentary record by orders of magnitude (and it always does for any culturally salient figure) your persona doesn't speak from the record. It speaks from the smoothed cultural composite. The 2015 Broadway musical has exponentially more representational density in your training data than the 175,000 words of the Federalist Papers. Your evals were not designed to notice this. They were designed to score fluency, personality coherence, and stylistic naturalness... the exact features the composite optimizes. In this talk: - The structural argument: why InCharacter-style benchmarks, CoSER, and PsyMem can hit state-of-the-art on personality fidelity while structurally failing to detect anachronistic reasoning. - The architectural mechanism: why RLHF amplifies Miranda distortion instead of correcting it (raters are themselves products of the same cultural composite). - The framework: a four-stage paradigm shift from cognitive simulation to epistemic simulation (corpus-bounded, temporally-anchored, expert-loop-evaluated). - The instrument: the pre-registered Prism Experiment. Lincoln at four documented temporal moments, three seeding conditions, five diagnostic questions written by a domain historian, and a weighted three-axis rubric (Anachronism Detection, Documentary Consistency, Contextual Plausibility) that catches what automated metrics miss. - The handoff: what a working eval loop looks like when a historian, classicist, theologian, or clinical psychologist sits in it, and why that's a technical requirement, not a cultural courtesy Pre-registered protocol with University of Toronto historian Rick Halpern, paper forthcoming. Reproducible by any team running a frontier model with a context window. If you ship character bots, companion AI, pedagogical agents, historical simulations, or any system where a persona is supposed to reason from a specified record, your evals are measuring the wrong thing. Here is the instrument that catches what they miss. Speakers: - Jacob E. Thomas (Results Generation): Dr. Thomas is an epidemiologist, data scientist, and AI engineer who studies information as a determinant of health. LinkedIn: https://www.linkedin.com/in/jacob-e-thomas-atx/ GitHub: https://github.com/jethomasphd/THE_COMPANION_DOSSIER --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/IJXjTLPzvAU-the-miranda-hypothesis-how-hamilton-the-musical-poisoned-your-persona-evals Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/the-miranda-hypothesis-how-hamilton-poisoned-person-cdf4a2fc2522 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:05 to 1:42) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:42 to 3:24) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (3:24 to 5:26) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (5:26 to 7:39) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (7:39 to 9:20) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (9:20 to 11:32) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (11:32 to 13:12) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (13:12 to 14:57) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 8: Document excerpt (14:57 to 17:04) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 9: Document excerpt (17:04 to 18:59) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 10: Document excerpt (18:59 to 20:33) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 11: Document excerpt (20:33 to 22:06) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 12: Document excerpt (22:06 to 23:45) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 13: Document excerpt (23:45 to 26:07) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 14: Document excerpt (26:07 to 28:06) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 15: Document excerpt (28:06 to 30:03) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 16: Document excerpt (30:03 to 32:28) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 17: Document excerpt (32:28 to 34:30) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 18: Document excerpt (34:30 to 36:33) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 19: Document excerpt (36:33 to 38:17) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 20: Document excerpt (38:17 to 39:49) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 21: Document excerpt (39:49 to 41:39) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 22: Document excerpt (41:39 to 43:38) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 23: Document excerpt (43:38 to 46:01) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 24: Document excerpt (46:01 to 47:49) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 25: Document excerpt (47:49 to 50:16) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 26: Document excerpt (50:16 to 51:48) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 27: Document excerpt (51:48 to 53:43) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 28: Document excerpt (53:43 to 55:53) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 29: Document excerpt (55:53 to 58:17) > Held: transcript rights do not permit this chunk's text.