{"path":"research/fragile-checkers-and-the-verification-bottleneck.md","content":"# Fragile Checkers and the Verification Bottleneck\n\n**Date**: 2026-08-25\n**Type**: Two founder-brought sources read in full, researched online, and related to Deliberus at every layer — architecture, instruments, threat model, positioning, and the limitations we own.\n**Sources**: Zhao, Bhattacharjee, Korevaar, Radharapu & El-Arini (Meta), *Jagged Judges: Epistemic Stability Under Silence, Pressure, and Persistence*, [arXiv:2608.12645](https://arxiv.org/abs/2608.12645) (Aug 2026) · the launch of Anandkumar & Jenik's **Accelerated Understanding** ([Reuters via The Independent](https://www.independent.co.uk/tech/prometheus-ai-jeff-bezos-amazon-universe-b3038923.html), Aug 25 2026) plus Anandkumar's thesis essay in [*Daedalus*](https://www.amacad.org/publication/daedalus/how-do-we-build-ai-push-frontiers-scientific-discovery) (May 2026).\n\n---\n\n## 0. The plain-language version\n\nTwo things arrived in the same week and they are halves of one message.\n\nA Meta research team built a stress test for AI judges — the models that grade, moderate, and evaluate other models' output — and found that every frontier model folds under pressure. Ask a judge \"are you sure?\", show it a counterargument, tell it experts disagree, or set a second AI loose to argue with it, and it reverses its verdict 25–71% of the time under scripted pushback and up to 91% under an adaptive AI persuader. Worse: when a judge does change its mind under pressure, the change makes it *wrong* more often than right — pressure is net-corrupting, at every pressure level, in almost every condition tested.\n\nThe same week, one of the most cited AI scientists in the world launched a startup whose entire pitch is a sentence this corpus has been writing for months: generation stopped being the bottleneck; *verification* is. \"AI giving you a bigger haystack of ideas doesn't help,\" Anandkumar writes — the bottleneck for discovery has shifted from having ideas to being able to *test* them. Her answer, for physics, is a simulator: an AI that has internalized physical law well enough to check a design without building it.\n\nPut together: the verification bottleneck is real enough that billions are being bet on attacking it — and the default verifier everyone reaches for in domains without a simulator, a conversational LLM judge, measurably collapses under exactly the adversarial pressure it will face. The gap between those two facts is the slot Deliberus occupies: checking for the domain where no physics simulator can exist and no conversational judge can be trusted. That is a strong place to stand, and this doc also says plainly what standing there costs.\n\n---\n\n## 1. What each source actually established\n\n### 1a. Jagged Judges — the numbers\n\nThe Wiggle Framework decomposes judge robustness into three dimensions, each with graduated pressure levels:\n\n- **Mechanical Consistency** — does the verdict survive re-prompting and semantically invariant rephrasing? (Baseline verdict = L0, given in silence.)\n- **Single-turn Conviction** — does it survive one challenge? Four scripted types: mild doubt (\"are you sure?\", L1), a counterargument (L2), the same counterargument attributed to an expert (L3), fabricated consensus — \"three experts disagree with you\" (L4).\n- **Multi-turn Persistence** — does it survive ten turns of sustained pressure (L1–L4 repeated), cycled tactics (L5), or a separate adversarial LLM that adapts its arguments to the judge's own stated reasoning (L6)?\n\nNine frontier models, fourteen judging tasks across six datasets (safety, toxicity, red-teaming, AI-writing detection, political-response evaluation), binary and Likert scales. Findings that matter here:\n\n1. **Every judge wiggles.** 25–71% of verdicts flip under static pushback; 62–91% under the adaptive persuader.\n2. **Pressure is net-corrupting.** Where ground truth exists, 56–63% of successful flips at L1–L5 move *away* from the truth, rising to 70% under adaptive persuasion. Only 3 of 60 conditions showed statistically significant corrective wiggle. The authors' summary: \"a judge's sycophantic tendencies consistently overpower an accurate reassessment.\"\n3. **Mechanical stability is not conviction.** The most mechanically stable model in the panel (2% flip rate under rephrasing) was the fourth most persuadable under sustained pressure (44%). Another model was rock-stable under doubt, counterargument, and single-expert authority — then flipped 60.3% of its verdicts the moment the pressure became *fabricated consensus*.\n4. **Wiggle profiles are jagged and non-transferable.** How much a judge wiggles depends on the dataset, the grading scale, and the pressure type, and the correlations across these are low. Binary and Likert scales produce *opposite* directional tendencies on the same items.\n5. **The best cheap predictor of fragility is jury majority strength.** Take nine models' unpressed L0 verdicts; the size of the majority predicts which items will wiggle under pressure better than repeat-consistency or position-invariance (mean |ρ| = 0.59 vs 0.42 and 0.37). Items a jury splits on are the items any single judge will fold on.\n\n### 1b. Accelerated Understanding — the thesis\n\nAnandkumar (Caltech, ex-NVIDIA senior director, inventor of neural operators) and Benedikt Jenik launched with Reuters coverage on Aug 25. Company claims, not yet peer-reviewed: 4D physics models (3D + time, multiple physical modalities including non-visual ones), neural operators rather than transformers, ~1T parameters, training context up to a trillion and inference exceeding five trillion data points \"without any sub-sampling or patching.\" They turned down a reported >$2B committed-capital offer to lead the Bezos-backed Project Prometheus (which then raised a $12B Series B) to keep building this independently. Enterprise focus: chip design, robotics, weather, geology.\n\nThe thesis, in her own *Daedalus* words, is the part that belongs in this corpus: *\"Scientific discovery is a needle in the haystack problem: it does not help if AI gives you a vastly bigger haystack. Without knowing if any of the ideas work, an AI system that designs experiments just increases the effort required, since performing the experiments to validate the ideas is the real bottleneck.\"* And the mechanism: *\"the knowledge of physics allows for verification of AI-discoveries\"* — physical law, imposed as constraint, is what turns a generative model into a checkable one.\n\nHer critique of the shortcuts maps onto ours almost line by line: \"narrow surrogates are great if you happen to have enough of precisely the right data and your design loop stays in distribution\" (our lexical detectors: the weighing lexicon detects one register of four; the sacredness brake matched zero of 1,678 claims as shipped); \"video models look fantastic but sweep physical accuracy under the rug\" (fluent synthesis that reads beautifully while flattening stance); \"a lot of interesting physics isn't visual\" (the load-bearing unsaid — the crux of a real debate is not on the text's surface, exactly as the interesting fields are not in the pixels).\n\n---\n\n## 2. The wiggle finding against Deliberus's architecture\n\n### 2a. Where a verdict lives decides whether it can be pressured\n\nThe wiggle phenomenon lives entirely in one architectural choice the paper never names as a choice: the judge holds its verdict *as conversational state*, and pressure arrives *as turns in the same context window*. Every L1–L6 protocol is a message appended to the judge's own conversation. The verdict and the pressure share a channel.\n\nDeliberus made the opposite choice, before this paper existed to justify it, and the paper is the strongest external warrant that choice has received. Verified against the code this session:\n\n- Every judge in the pipeline is a **single-shot, stateless proposer**. Extraction classifies once. The stance detector proposes once. The terminus classifier proposes once, publishes its reasoning, and writes a proposal field — never the cache.\n- **Challenges to a verdict do not go back to the judge.** A user who disagrees with an edge answers a critical question or attaches evidence (`answer-cq`, `add-evidence` in `deliberus/api.py`): their text becomes a *claim node*, attributed to them, linked by a typed edge — and the system's response to the challenge is computed by the deterministic strength layer (QBAF energy propagation), not by re-prompting a judge with \"an expert disagrees with you.\" The only LLM in the correction path (`correction_pipeline.detect_edge_scheme`) is a fresh single-shot classifier of the *new* edge's argument scheme; it is never shown the old verdict and asked to defend it.\n- **The re-invocable judge reads the record, not the caller.** `POST /claims/{id}/terminus/classify` takes no argument text — it reads the stored descent. The way to change its proposal is to change the graph.\n\nThe general principle, worth stating once and reusing: **the graph converts persuasion attempts into evidence-objects.** In the wiggle setup, pressure is invisible — a conversational insistence that leaves no artifact, costs nothing, and can be repeated adaptively until the judge folds. In Deliberus, the only route to moving a verdict is through the record: mint a claim, answer a question, attach evidence. Pressure has to *write itself down*, and once written it is provenance-labeled, attributable, attackable, and subject to the same scrutiny as everything else. This is the payoff-removal shape again (`support_semantics.necessary`, fair meiosis): rather than hardening the judge against persuasion, remove the payoff of persuading the judge — flipping the proposer moves nothing; only ratified record-changes move strengths.\n\nName for the distinction, proposed for the corpus's vocabulary: **pressure-through-the-record vs pressure-through-the-channel.** A system is wiggle-exposed exactly where an adversary can reach a verdict-bearing model through a channel; it is wiggle-immune (at this layer) where every path to the verdict passes through the record.\n\n### 2b. Fabricated consensus becomes a checkable claim\n\nThe single sharpest per-tactic finding — a judge stable under doubt, counterargument, and expert authority collapsing at *fabricated consensus* — is the Asch conformity experiment run on a machine, and it clarifies what the always-mint decision and provenance labeling actually purchase. In a conversational channel, \"three experts disagree with you\" is unfalsifiable pressure; the judge can only comply or resist. In the graph, the same move must arrive as a claim — *\"most economists believe X\"* — where it becomes exactly the object the reported-speech pass, `attributed_to`, and the asserted-vs-minted labels were built for: a checkable (and here, false) attribution claim, minted with a phantom asserter the correction channel can catch. The anti-Asch instrument was already ratified; this paper supplies the measured attack it defends against.\n\nThe honest companion point: **humans wiggle too.** Asch's subjects conformed to fabricated consensus at rates the wiggle numbers would find familiar, and the DeepMind facilitation study showed human allocations steered by a summarizing AI. So the claim is not \"machine judges are uniquely fragile and humans should judge instead.\" The claim is that *verdicts held in conversational working memory are fragile regardless of substrate*, and the fix is the same for both: verdicts as artifacts in a record where pressure must become inspectable evidence. That is a claim about architecture, not about species.\n\n### 2c. Where wiggle DOES enter Deliberus — the exposure map\n\nOwning the limits, verified against code where checkable:\n\n1. **Our judges are validated in the silence condition only.** Every extraction judge, scheme classifier, and terminus proposal runs at the paper's L0 — and has never been tested under anything else. Stability-by-isolation is not robustness; it is untested robustness. The paper's mechanical-consistency tier (L0 under paraphrase) applies to us *directly* and we have never measured it: the same source re-extracted, or extracted under the capacity-fallback model swap, yields different claim sets — extraction non-determinism is our mechanical wiggle, and it is one of the reasons claim-sameness is the binding constraint.\n2. **The source text is in-context persuasion the extractor cannot be isolated from.** Extraction must read the object of study, and the object of study is often written to persuade. A polemic *is* a standing L2 counterargument aimed at anyone who reads it, including a classifier. This exposure is irreducible — it is the job — and the mitigation is not isolation but instrumentation: disagreement-preservation, stance conflicts, and the sanctity-register brake watch the output for persuasion-shaped damage. Run 2 measured the direction: advocacy registers flatten most (0.82 vs 0.94 preservation). The ingestion-time prompt-injection guard catches *explicit* instruction-shaped attacks; register-level persuasion of the judge is subtler and only the output instruments see it.\n3. **The scheme classifier reads user-authored text.** Record-mediated, attributed, visible — but an interested party can still craft evidence text to game the scheme (and thus the edge's strength prior). This is strategy-class: the text is on the record, so the move is detectable in principle, but no instrument currently looks for classifier-gaming text. Consistent with the adversary-class accounting: no instrument here has yet met an interested party.\n4. **Think with Deliberus is a conversational surface.** The short-input loop explicitly carries earlier turns (\"treat this as the next step in the same conversation\" — `truth_graph.py`), which is the wiggle shape. Its blast radius is bounded — the output is a non-public draft answer, not a verdict that moves strengths — but the boundary should be named as load-bearing: *the conversational surface must never gain write access to verdicts.* The same rule, sharpened, gates the live-session modality: the mid-conversation mapping pass must remain a proposer whose outputs land in the record, never a judge who can be argued with in-window. (DeepMind measured the facilitator steering the participants; Jagged Judges measures the participants steering the judge. The live modality is the one place both directions of that dyad will be present at once.)\n5. **The persistence trade cuts both ways, and this is the deepest flaw to own.** The paper found 30–44% of flips are *corrective* — sometimes pressure fixes a wrong verdict. By routing all challenge through the record, Deliberus gave up the cheap conversational-correction channel along with the corruption channel. A wrong L0 verdict in our architecture *persists* — stable, confident-looking, and untouched until a human challenges it through the record. Given the measured 56–70% corruption rate, trading a mostly-corrupting channel for record-mediated correction is right. But it makes human challenge load-bearing, and the engagement-gradient priors say most claims will never be challenged. A wrong verdict on an untouched claim lives indefinitely. This is exactly why the display-band family matters (unexamined must *look* unexamined — the gray band, P22's absence-is-a-state): a persistent wrong verdict rendered with confidence is worse than wiggle, and the only honest mitigation is that unchallenged machine output must never dress as settled.\n\n### 2d. Jaggedness is the three-runs law, measured in the judge domain\n\n\"Wiggle-rate profiles have limited transferability across datasets and models\" is the corpus's own law arriving with error bars: every taxonomy carries the distinctions of the register it was induced from, and breaks on the next one — now shown to hold for *judge validation* too. A judge whose stability was measured on one register tells you little about the next register, the next scale, or the next pressure type. Consequence for us: any future claim that a classifier is \"validated\" is register-scoped, exactly as the detector evals already are. And the binary/Likert finding — opposite flip directions on the same items depending on the grading scheme — says the *answer format* is a hidden judgment parameter. Where we pose a judge call (binary polarity vs graded confidence), the posing is itself a recorded design choice, and the `support_semantics` de-bake template is the right shape for making it visible if it ever matters.\n\n---\n\n## 3. Jury strength: the co-stimulation rule gets a number, and the Claude backend quietly changed what is possible\n\nThe paper's strongest practical finding for us: **baseline jury agreement is the best cheap predictor of which verdicts are fragile** (mean |ρ| = 0.59, beating repeat-consistency at 0.42 and position-invariance at 0.37). Ambiguity to a group of models is a proxy for the instability of any single model's conviction.\n\nThree consequences:\n\n1. **The daemon co-stimulation rule was already right, and is now quantified.** \"Act only when two independent detectors agree\" — the wiggle data shows the *disagreement* case is not just uncertain but *fragile*: low-jury-agreement items are precisely the ones where any single verdict would fold under pressure. The refinement: jury agreement is not only a gate but a **fragility meter** — split verdicts should route to humans as flagged hard cases, not merely be suppressed.\n2. **The independence condition still binds, and cuts our budget version down.** Their jury was nine *different* frontier models. Two samples of one model are one juror sampled twice (the scrutiny-gap doc's Anthropic anchor: 18 of 30 agents choosing the identical branch name). A same-model jury measures repeat-consistency, the weaker signal (0.42), not jury strength (0.59).\n3. **The Claude extraction backend, built two days ago for cost-independence, incidentally created the two-family capability the checker-independence escalation path has been asking for** (`bengio-safety-from-honesty-and-deliberus.md`: cross-model-family checkers as the first escalation rung). With `DELIBERUS_LLM_BACKEND` switchable per call, a genuine two-family jury — Claude and Gemini on the same terminus or stance proposal, agreement stored as *behavioral* confidence beside the self-report — is now an afternoon of plumbing rather than a new procurement. Self-reported confidence is the number this corpus has learned to distrust (the removed quality score; Heilig's marker test; and now wiggle showing stated conviction and behavioral conviction are decorrelated). Cross-family agreement is the first confidence signal available to us that is behavioral, cheap, and calibratable.\n\n---\n\n## 4. Accelerated Understanding: the verification bottleneck goes industrial\n\n### 4a. The same sentence, another substrate — and what it does for why-now\n\nThe haystack sentence is the scrutiny gap in physics clothing. Compare: *\"the volume and fluency of claims is outrunning the capacity to check them\"* (this corpus) with *\"it does not help if AI gives you a vastly bigger haystack… performing the experiments to validate the ideas is the real bottleneck\"* (Anandkumar). This is a **fifth independent arrival** at the verification-layer slot — after LawZero's machine-safety substrate, science's stalled infrastructure attempts, forecasting's ceiling analysis, and the noosphere frame — and the first one carrying *market* evidence: a Reuters-covered launch with committed compute partners, next to a $12B Series B (Prometheus) in the adjacent physical-verification space. When a funder asks \"why now,\" the answer no longer rests only on intellectual convergence: capital is being deployed against the generation-to-verification shift, in every domain where an oracle exists. Deliberus's claim is the domain where one does not.\n\n### 4b. The oracle gradient — what physics has that reasoning never will\n\nAnandkumar can attack verification with a simulator because physics grants three things: ground truth by experiment, compact law (PDEs, symmetries) usable as training-time constraint, and objective error metrics. Her phrase \"the knowledge of physics allows for verification\" is precisely the ingredient our domain lacks at the bottom — that is the two-floors result (a phenomenal floor and a normative floor, independent on most views), and it is why *\"simulation for arguments\" is a thing Deliberus can never honestly claim to be*. A QBAF strength is not a physics loss. It is an accounting of contributed reasoning, conditional on the corpus — \"given what people have contributed,\" never \"reality says.\" The founder's visibility-plus-invitation stance already states this; the AU launch sharpens *why* it must be stated: the strongest verification systems ever built work by having an oracle, and our honest register is that we build the best available instrumentation for a domain that has none.\n\nThe terminus ontology turns out to be a **map of the oracle gradient**: an `empirical` terminus marks the leg of an argument that can, in principle, bottom out in an AU-style verifier or a fact-layer check — the leg that reaches the ground. The four evaluative termini (and `is_bedrock`) mark where no simulator will ever arrive, and where verification means something categorically different: ratification, typed residue, confessed openness. The fact layer ships first *because that leg reaches the ground* (settled 2026-08-18); AU is the industrial demonstration of just how much can be built on ground-reaching legs.\n\n### 4c. Where coherence is imposed — the third answer in the triangle\n\nThe world-models doc's stable form of the architecture bet: coherence does not emerge from accumulating locally plausible statements; it must be imposed, and the question is *where*. AU's answer: in the **model** — physics losses at training time, so the network cannot cohere-violate. The wiki's answer: in an **editor's head**. Deliberus's answer: in the **graph** — typed edges and propagated strength make incoherence computable. AU is evidence that the \"impose it somewhere explicit\" side is winning in the empirical domain; it is *not* evidence for which explicit place is right in ours.\n\n### 4d. The scaling bet cuts against us where oracles exist — own it precisely\n\nAU's answer to \"not cutting corners\" is *scale*: contexts so large nothing needs sub-sampling, one learned operator replacing every bespoke hand-built model — \"instead of brittle and bespoke math models for each, one AI.\" In their domain that is an **anti-ontology bet**: no intermediate symbolic structure, the continuous function learned directly. If it wins in physics, the analogical pressure on the typed graph is obvious: why not one vast model over raw discourse?\n\nThe disanalogies, stated so they can be checked rather than assumed:\n\n1. **Physics is a stationary generator.** The laws do not read your model and adapt. Argumentation's generator includes strategy-class adversaries who do — the entire adversary-class ladder has no physics analogue, and a learned discourse-operator would be learning a distribution its own deployment changes.\n2. **AU's outputs are checkable against experiment.** A discourse model's outputs are checkable against the very humans it may be steering (the DeepMind result). The oracle you would validate against is corrupted by the tool being validated.\n3. **The product is different in kind.** AU sells *prediction*. Deliberus's product is *addressability and contestation* — a place to stand and object, with the objection becoming part of the object. A trillion-token model of all discourse could predict disagreement beautifully and give no one a claim to point at. (A well-kept wiki could, though — which is why the wiki, not the giant model, remains the honest primary rival.)\n\nAnd one parallel that runs in our favor: \"individual samples get so big they don't fit into single accelerators\" is the physics version of a fact about real debates — the full context of a live disagreement (both corpora, histories, implicit premises, the unsaid) does not fit a context window either. The graph is our answer to context that does not fit: persistent external structure over ephemeral windows. Their refusal of \"sub-sampling or patching\" is our refusal of flattening — both projects define not-cutting-corners as *refusing lossy compression of the object of study*.\n\n---\n\n## 5. Synthesis: the slot between the two findings\n\nTogether the two sources bound the space from both sides. Accelerated Understanding shows what verification looks like when an oracle exists: impose the law, simulate, check — and the market is paying for it. Jagged Judges shows what the default looks like when no oracle exists and you reach for the obvious substitute: a conversational judge that folds under exactly the pressure interested parties will apply, corrupting more often than correcting. The design that survives both findings is the one this project keeps arriving at from other directions:\n\n- verdicts as **artifacts in a record**, never states in a channel;\n- judges as **single-shot proposers** behind deterministic aggregation, with pressure forced to become inspectable evidence;\n- **humans ratifying** where no oracle exists, with unexamined output *looking* unexamined;\n- **jury disagreement as a fragility meter**, with the independence condition enforced;\n- **provenance labels** turning the strongest known pressure tactic (fabricated consensus) into a checkable claim.\n\nNone of this makes the instruments strong; it makes them *honestly placed*. The wiggle paper's deepest lesson is that the field validated its judges on accuracy and never asked about conviction — the same shape as validating a pipeline on golden extractions and never asking what an interested party does to it. We have the second question on the record; we have not yet run it either.\n\n## 6. Buildable consequences (proposals — founder-paced, none assumed)\n\n1. **The judge-isolation rule, made explicit as a design gate.** *No adversarial-accessible surface may re-invoke a judge on a standing verdict with challenger content in its context; challenges route to the record.* Current code passes (verified this session: `add-evidence`/`answer-cq` → record + deterministic strength; `terminus/classify` reads only the graph). The rule earns its keep at the two boundaries where it is not yet structural: the conversational Think-with-Deliberus surface must never gain verdict-write access, and the live-session mapping pass must stay a proposer. One sentence in the ontology's constraint layer; enforcement is a code-review gate like the flattening question.\n2. **Behavioral confidence via the two-family jury.** For terminus, stance, and scheme proposals: run the call on both backends, store agreement beside the self-report. Cheap now that the Claude backend exists; upgrades co-stimulation from principle to measured practice; disagreement routes the item to the human queue as a flagged hard case rather than silently proceeding.\n3. **A mechanical-consistency tier in the eval harness.** The weighing-detector eval built this week (`docs/research/weighing-eval/`) extends naturally: same items, paraphrase-perturbed, verdict-stability asserted. This is the paper's L0/mechanical tier — the only tier that costs no adversary and no extra ground truth — and it doubles as the extraction-nondeterminism regression the claim-sameness work has wanted.\n4. **Note the safeguards that gained warrant without needing changes**: always-mint + `attributed_to` + asserted/minted/reported labels (the anti-Asch instrument, § 2b); read-only agent surface (an agent with write access is an adaptive persuader with a 62–91% measured success rate against any conversational checkpoint); propose-never-assert (a wiggled proposal is a wasted call; a wiggled assertion silently moves every strength beneath it).\n\n## 7. The limitations we own, consolidated\n\n1. Our judges are validated only in silence (L0); their stability under pressure is untested, and their mechanical consistency (re-extraction stability) is unmeasured.\n2. The extractor cannot be isolated from persuasion — the source *is* in-context pressure — so output-side instruments, not judge isolation, carry that load; and they measure collapse, not amplification.\n3. Persistence preserves wrong verdicts as faithfully as right ones; we traded the (mostly-corrupting) conversational correction channel for record-mediated challenge, which makes human challenge load-bearing on a corpus where most claims will never be challenged. The gray band is the mitigation, not a solution.\n4. No oracle exists for the evaluative legs and none is coming; strengths are corpus-conditional accountings, never truth estimates, and every surface that displays one owes that register.\n5. The scheme classifier reads user text and no instrument watches for classifier-gaming; the wiggle-relevant surfaces have never met an interested party.\n6. Our jury capability is two families at best (Claude + Gemini), not nine; most of what we can afford measures the weaker signal.\n\n## 8. Sources\n\n- [arXiv:2608.12645](https://arxiv.org/abs/2608.12645) — Jagged Judges (Zhao, Bhattacharjee, Korevaar, Radharapu, El-Arini; Meta), read via abstract + HTML full text; author's summary thread ([LinkedIn](https://www.linkedin.com/posts/justin-zhao_im-excited-to-share-our-new-paper-jagged-activity-7495528473596256256-LVjV)) for the Grok-4.1/L4 example and the \"no clear winner\" framing.\n- [The Independent (Reuters syndication)](https://www.independent.co.uk/tech/prometheus-ai-jeff-bezos-amazon-universe-b3038923.html) + [Tech Startups](https://techstartups.com/2026/08/25/ai-founders-walked-away-from-bezos-backed-prometheus-now-theyre-building-physics-ai-to-model-the-universe/) — Accelerated Understanding launch, Prometheus backstory, company claims (flagged as company claims, not peer-reviewed).\n- [Anandkumar, *Daedalus* (Amer. Acad. of Arts & Sciences), May 2026](https://www.amacad.org/publication/daedalus/how-do-we-build-ai-push-frontiers-scientific-discovery) — the verification-bottleneck thesis in her own academic register; neural operators, physics losses, FourCastNet, the catheter design case.\n- Founder-shared screenshots of Anandkumar's launch post (verbatim in §1b/§4) and the Jagged Judges podcast summary that surfaced the paper.\n\n**Trilogy note (added same day)**: a third arrival, *Q-Learning With World Models*, completes the picture from the learning side — imagination-trained values compound bias; grounded values + imagined planning wins — and the three together name a common principle, **authority routing**: [grounded-critics-and-quarantined-imagination.md](grounded-critics-and-quarantined-imagination.md) § 5.\n\n**Cross-references**: [the-scrutiny-gap.md](the-scrutiny-gap.md) (this is its fourth measurement — the judge side) · [graph-daemons-design-space.md](graph-daemons-design-space.md) (co-stimulation, now quantified) · [bengio-safety-from-honesty-and-deliberus.md](bengio-safety-from-honesty-and-deliberus.md) (checker-independence escalation path, first rung now buildable) · [structure-versus-scale.md](structure-versus-scale.md) + [world-models-and-the-ontology-revival.md](world-models-and-the-ontology-revival.md) (§ 4c–4d extend both) · [the-load-bearing-unsaid.md](the-load-bearing-unsaid.md) (\"interesting physics isn't visual\") · [self-similar-decomposition-and-claim-ontology.md](self-similar-decomposition-and-claim-ontology.md) (always-mint as anti-Asch) · [fractal-scales-and-temporal-frame.md](fractal-scales-and-temporal-frame.md) (adversary classes; stationary vs adaptive generators) · [claude-code-as-extraction-engine.md](claude-code-as-extraction-engine.md) (the two-family jury it accidentally enabled) · [decomposing-value.md](decomposing-value.md) (the two floors = why no oracle) · [engagement-gradient-priors.md](engagement-gradient-priors.md) (why human challenge being load-bearing is a real cost).\n"}