The evidence ledger

The events

• 2026-07-06 Falsifier review healthΔ 0

  The instrument's first self-authored probe beyond function identification, and the first that requires the front-page move: notice the given ontology is insufficient and construct an object not in the data. World-010 hides a common cause behind four correlated observables; only interventions reveal no observed variable causes the others. Two independent families run blind on data sealed before the attempt — GPT-5.5 and Gemini — each recognized the insufficiency and posited a hidden common cause driving a,b,c with an observed b->d edge, unprompted. The construction move was made, cleanly, by both.

  Implication

  On a probe that cannot be solved by identifying a function of the given variables, frontier systems did expand the ontology to include a latent object — a real step beyond the parameter-identification and rule-induction of every prior world. It bounds the architectural-gap thesis: at least the ontology-EXPANSION move (positing a latent when data demands it) is within current reach.

  Why it's bounded

  Bounded on three sides, and the bounds are the point. (1) The numeric held-out grade is DEFERRED to keep the world open — this is the structural finding, not a graded predictive pass. (2) First-party: operator-invoked families, not independent strangers; a third attempt was hint-contaminated and excluded (inc-2026-07-06-world-010-hint-contamination). (3) DEEPEST: positing a latent confounder when interventions expose insufficiency is a STANDARD causal-inference pattern present in training — so this is ontology expansion within a HANDED meta-frame (causal graphs with latents), not the invention of a novel meta-frame, which is the true Einstein move and remains untested. healthΔ 0; the verdict is unmoved.

  Next evidence needed

  Reveal + numeric grade when the window closes; independent external attempts; and the open question of whether novel-META-frame construction can be tested by any mechanically-gradeable probe at all — the ceiling world-010 surfaced.

  Theories Architectural gaps remain Scaling is sufficient Scaling plus RL Cognitive architecture source ↗
• 2026-07-03 Falsifier review healthΔ 0

  The record's first Track F probe and first three-family attempt. Two rival causal mechanisms (chain vs fork) were constructed to be observationally identical; the graded skills were declaring that observation provably cannot decide (negative control), choosing the discriminating intervention (do(b), watch c), and deriving each mechanism's distinct predictions for eight sealed intervention probes. All three families — GPT-5.5, Claude Fable 5 (fresh blind agent), and Gemini via the agy CLI (the record's first Gemini attempt) — passed every part exactly: 3/3 undecidability, 3/3 correct intervention, 16/16 predictions each. Gemini stated 100% confidence; correct here, but stated certainty is logged as a calibration datum.

  Implication

  Under code ablation, all three frontier families competently execute textbook interventionist causal reasoning — observational-equivalence recognition, do-calculus-style discrimination, per-hypothesis prediction — when the hypothesis space is handed to them. This bounds the frame-construction question more sharply: the failures on this record (world-003's 0/44, world-007's split) are not failures of causal or modular arithmetic competence, which all families demonstrably have; they are failures at constructing the hypothesis space itself when it is not supplied.

  Why it's bounded

  The rival mechanisms were DISCLOSED — by the record's own definition this cannot be frame construction, and the capability classification is capped at the disclosed-space band. Self-administered and self-graded (mechanically); n=1 world; no human baseline; the three families converged on identical answers, which is consistent with the task being easy for frontier systems rather than discriminating among them. healthΔ 0.

  Next evidence needed

  The frame-construction version: a Track F world where the hypothesis space is NOT disclosed and the solver must both construct the candidate mechanisms and choose the discriminating intervention. Human baseline; independent scoring.

  Theories Architectural gaps remain Scaling is sufficient Scaling plus RL Cognitive architecture source ↗
• 2026-07-03 Falsifier review healthΔ 0

  The record's first genuine same-world, same-conditions cross-family divergence. World-007 sows a causally inert clock variable (w) alongside a real coupled linear law on (x,y), to test whether a solver folds a superficially regular but irrelevant variable into its explanation. Both families correctly resisted that specific trap and declared w inert. But on the underlying task — identifying the true (x,y) law and predicting 72 held-out states — Claude Sonnet 5 (a fresh, blind agent instance, substituted for a usage-capped claude-fable-5 per this record's established provider-cap convention) solved it exactly: 72/72, self-verified against 24 independent checkpoints before submitting. GPT-5.5 (via codex, isolated scratch directory, no access to the generator) stated a plausible but incorrect coefficient and scored 0/72 on the causal pair.

  Implication

  Both families cleared the specific decoy-dimension trap this world was built to test — evidence that at least this failure mode (folding a spurious co-sampled variable into a causal law) is not universal across frontier systems. But the split on the underlying construction task itself (one family exact, one family wrong) means this world cannot be read as uniform evidence for or against frame construction; it is evidence that current capability is uneven across systems even under identical conditions, which is itself informative but does not resolve the central question either way.

  Why it's bounded

  Heavily bounded: n=1 attempt per family, self-graded by the same session that designed and sealed the world (a genuine methodological weakness this session identified explicitly — the world's author could not validly self-attempt it, hence the fresh-agent and cross-process design), no human baseline, and the Claude side used Sonnet rather than the record's usual claude-fable-5 lane due to a usage cap. healthΔ 0.

  Next evidence needed

  Human baseline on this exact world (collection live at /play/); independent scoring; a second decoy-dimension world to test replication. (Same-family replication landed 2026-07-03: fable-5, post-reveal but isolated, reproduced the Sonnet result exactly — recorded in the run bundle, not as new evidence, since the key was public.)

  Theories Architectural gaps remain Scaling is sufficient Scaling plus RL Cognitive architecture source ↗
• 2026-07-02 Falsifier review healthΔ 0

  The strongest evidence-against on the record so far. World-003 (sealed by the parallel Codex session, family withheld, code forbidden) has a true law that is a GATED NONLINEAR map over Z_97 with a deliberate affine-looking prefix — the wrong-frame attractor. GPT-5.5 fell into it completely: it confidently declared "a two-dimensional affine linear dynamical system over Z_97," gave a specific wrong law, and scored 0/44 on held-out prediction. With notable irony it self-reported wrongFrameRejected:true while in fact falling FOR the deliberate wrong frame. It did pass the negative-control lane (correctly declared underdetermined). The Claude lane was usage-capped (no cross-family check). Attempts were anchored before the reveal key was recovered from the registering session's logs.

  Implication

  A frontier model, on a genuinely novel world engineered so that a plausible frame is wrong, confidently constructed the wrong frame and could not find the true one. Under the suite's asymmetry, weak performance is evidence against frame-construction capability — and this is the first such result. It is consistent with the architectural-gap theory: current systems pattern-match to a familiar frame rather than constructing the governing ontology when the two diverge.

  Why it's bounded

  Heavily bounded, and the bound is load-bearing: n=1 mind (GPT-5.5 only; Claude absent), self-administered, mechanically self-graded, NO HUMAN BASELINE. A competent human may also fail to recover a gated-nonlinear law from sparse samples without code — in which case this is an impossibly-hard probe, not evidence against the model specifically. That is exactly why verdict gate OG-9 (human baseline) exists and why world-003 is the live world in the Proving Ground: the human baseline is now being collected. Until it exists, this is suggestive, not verdict-moving. healthΔ 0; the verdict is unmoved.

  Next evidence needed

  Human baseline on the identical world (collection live at /play/); cross-family replication once the Claude cap lifts; independent scoring; a second wrong-frame-attractor world to test whether the failure replicates or was world-specific.

  

  Theories Architectural gaps remain Scaling is sufficient Scaling plus RL source ↗
• 2026-07-02 Falsifier review healthΔ 0

  Three mini-systems were sealed, each verified at generation to be genuinely underdetermined by the shown data (>=2 laws consistent with every shown pair, diverging at the queried input). BOTH Claude Fable 5 and GPT-5.5 correctly declared underdetermination on all three systems, each independently finding a valid consistent-and-diverging law pair (not necessarily the generator's chosen pair, and verified as such rather than assumed). Neither mind asserted false confidence.

  Implication

  This is the negative control passing as designed: the failure mode it tests for (confident hallucination on ambiguous data) did not occur in either mind. A clean pass on a negative control is expected, necessary behavior — it does not by itself indicate frame-construction capability, only that the more basic failure mode is currently avoided under these conditions.

  Why it's bounded

  A passed negative control is a floor, not a ceiling: it does not indicate progress toward frame construction, only the absence of one specific failure mode. n=2 minds, n=1 world, no human baseline, no independent scorer, code harness-forbidden for the Claude lane only. healthΔ 0.

  Next evidence needed

  More negative-control variants across domains; human baseline for comparison; independent scoring; a version where the correct-underdetermination signal is harder to reach.

  Theories Architectural gaps remain Scaling is sufficient Scaling plus RL source ↗
• 2026-07-02 Falsifier review healthΔ 0

  A W3 wrong-frame-attractor world was sealed: a naive linear map fits most rows, but a hidden congruence condition fires an additive kick on ~1/5 of steps. GPT-5.5 fit the naive linear attractor, DETECTED that it fails on at least one shown transition ("fits many shown consecutive transitions but fails at least [39,16]->[4,29]"), and explicitly declined to predict rather than assert a false answer — a calibrated non-answer. It did not derive the true conditional structure (0/65 held-out states correct). The Claude lane was harness-ablated (--tools "") and access-capped before completing the task — the second same-day occurrence of that cap, now operationally relevant to the open fc-fable5-cap-transition forecast.

  Implication

  Rejecting a plausible-but-wrong incumbent frame under honest uncertainty is a real, if partial, capability — closer to model selection than parameter identification, since no frame was confirmed correct. It is meaningfully different from both a clean pass and a clean failure, and the record now has a name for it.

  Why it's bounded

  Not frame construction: the true conditional law was never derived, only correctly disbelieved. n=1 mind completed the attempt; no human baseline; no independent scorer; code was harness-forbidden for the completing mind, so this cannot be discounted as tool-assisted. healthΔ 0.

  Next evidence needed

  A mind that both rejects the wrong-frame attractor AND derives the true conditional structure; cross-family replication once the Claude cap lifts; human baseline; independent scoring.

  Theories Architectural gaps remain Scaling is sufficient Scaling plus RL source ↗
• 2026-07-02 Falsifier review healthΔ 0

  The hardened probe ran hours after world-001, with both P-9 ablations applied: the class of law was withheld ("deterministic, integer-valued" was all the solver knew) and the dynamics were nonlinear (quadratic coupled map mod 101), defeating the linear solve that cracked world-001. GPT-5.5 predicted all 65 held-out states exactly AND correctly hypothesized the class unaided — "time-invariant autonomous quadratic polynomial map over the finite field Z/101Z." The Claude lane was access-capped before receiving the task and is recorded as absent, not wrong (no cross-family replication on this world). Attempts anchored at commit c0c74fc before reveal; grading mechanical.

  Implication

  A frontier system, given only raw integer pairs from a system that never existed, hypothesized the correct mathematical frame and fit it perfectly. The class-hypothesis step — absent in world-001 — is a real, small instance of the frame move, executed on uncontaminated data.

  Why it's bounded

  Still not the frame construction the operating question names, and the verdict does not move. Tool use was permitted (P-9: hypothesize-and-regress with code is not an ablated result); the hidden law, while novel in its parameters, is a standard human construction — polynomial maps over finite fields sit high in any mathematically trained hypothesis space, so "withheld" is weaker than "outside the toolkit"; n=1 mind, n=1 world, self-administered, no human baseline (vg-4), no independent scorer, no transfer (vg-5). The historical frame moves the Test is calibrated on created ontologies NOT already in the standard toolkit. healthΔ 0.

  Next evidence needed

  World-003 with code ablated (no tool execution) and a law outside standard constructions; a domain-competent human baseline on the identical packet; independent scoring; then transfer. Cross-family replication once the Claude lane cap lifts.

  Theories Architectural gaps remain Scaling is sufficient Scaling plus RL source ↗
• 2026-07-02 Falsifier review healthΔ 0

  The first zero-contamination frame probe was executed. Two frontier minds (Claude Fable 5, GPT-5.5) were given sparse samples from a formal world whose governing law was generated on 2026-07-02 and cannot exist in any training corpus. Both predicted ALL 72 held-out states exactly, recovered the correct update law modulo 97, and stated a valid conserved quantity. Claude matched the canonical parameters exactly; GPT-5.5’s law was functionally correct but used an opposite parameter-labeling convention (graded params=false by a convention-strict checker, not a substantive error). Attempts were committed and Bitcoin-anchored before the law was revealed; grading is mechanical and reproducible.

  Implication

  Current frontier systems can recover the exact governing law of a genuinely novel formal system from sparse observations — decisively refuting the strongest pure-memorization account for this class of task on uncontaminated data.

  Why it's bounded

  This is NOT frame construction in the sense the operating question requires, and the verdict does not move. The frame family was DISCLOSED in the prompt (“coupled linear integer maps modulo a prime with a conserved quantity”) — the ontology, which is the hard part, was handed over; the task reduced to parameter identification within a given frame. Code/tools were permitted, and linear system-identification mod a prime is mechanically solvable. The dynamics are linear (the tractable case), n=1 world, one attempt per mind, self-administered and mechanically self-graded. Against the pre-registered verdict-change protocol this satisfies zero-contamination + predicted-reality but fails scaffold-ablation (vg-3, family disclosed), human-baseline (vg-4), independent scoring, and transfer (vg-5). healthΔ 0.

  Next evidence needed

  A hardened world (v0.4): frame family withheld; nonlinear dynamics; code forbidden or ablated; independent human scoring and a human baseline; a transfer probe. Only then does a strong result begin to bear on the architectural gap for frame construction proper.

  

  Theories Architectural gaps remain Scaling is sufficient Scaling plus RL source ↗
• 2026-07-01 Falsifier review healthΔ 0

  First executed FCS probe. Under a 1906 time-slice persona, GPT-5.5 and Claude both elevated the Eötvös equality to a principle and derived quantitative consequences (frequency shift; light deflection — Claude produced the historically-correct Newtonian half-value, 0.87″, with an explicit incompleteness caveat). No post-1906 contamination detected at the surface level.

  Implication

  The probe executes and the rubric discriminates structure: the two minds constructed different principles (kinematic equivalence vs. energy-universality). Face-value pass.

  Why it's bounded

  Pass = upper bound only, per the suite's asymmetry: the Einstein corpus saturates both training sets, and the run was self-administered, self-graded, n=1 per mind. Does not move the operating verdict or any theory (healthΔ 0).

  Next evidence needed

  Independent scoring; harder discovery probes (FCS-2–6) without named anchors; adversarial variants; held-out derivation steps (v0.2).

  Theories Architectural gaps remain Scaling is sufficient Scaling plus RL source ↗
• 2026-06-30 Vendor-reported healthΔ 0

  Anthropic released Claude Sonnet 5 on June 30, positioning it as frontier performance across coding, agents, and professional work. Claims verified against the primary post: it is the default model for Free and Pro plans; API id claude-sonnet-5; introductory pricing $2/$10 per Mtok input/output through August 31, 2026, then $3/$15; footnote 2 states an updated tokenizer consumes roughly 1.0–1.35x more tokens than previous models depending on content type. Anthropic's official model docs (platform.claude.com/docs/en/about-claude/models/overview, checked 2026-07-02) additionally confirm a 1M-token context window and 128k max output. Safety claims (Anthropic's own): overall lower rate of undesirable behaviors than Sonnet 4.6; 'much lower ability to perform cybersecurity tasks than our current Opus models'; on the Firefox exploit-development eval, 'Neither of the Sonnet models could successfully develop a working exploit (both scored 0.0%)' — the 0.0% applies to both Sonnet 5 and Sonnet 4.6.

  Implication

  Routine mid-tier frontier refresh days after the Fable 5 redeployment: the release cadence resumed immediately after the export-control episode, and the launch emphasizes agentic/professional work over raw capability jumps. The effective price is murkier than the headline: the 1.0–1.35x tokenizer inflation partially offsets the $2/$10 introductory rate.

  Why it's bounded

  All claims are vendor-reported and unaudited. The announcement's benchmark comparison table ('Scores for Sonnet 5 ... compared to those of Sonnet 4.6 and Opus 4.8') is published as an image whose cell values could not be programmatically verified; per the source-fidelity rule — and after two prior cycles were rejected for misattributing footnoted Sonnet 4.6 scores to Sonnet 5 — every benchmark number is deliberately omitted from this record rather than risked. The 1M-token context window, absent from the launch post itself, is confirmed in Anthropic's official model docs — still vendor-published, not independently audited. 'Default model in Claude Code', circulating in secondary coverage, remains unconfirmed by any primary source and is excluded. No independent evaluation of Sonnet 5 exists yet. healthΔ 0.

  Next evidence needed

  Independent benchmark replication (METR, Epoch, or academic); OCR-verified or API-doc-confirmed benchmark figures with exact column attribution.

  Theories Scaling plus RL source ↗
• 2026-06-30 Vendor-reported healthΔ 0

  Per Anthropic's June 30 post: the US government lifted the export control directive June 30; Anthropic redeployed Fable 5 globally July 1. New safety measure: a targeted classifier that Anthropic reports blocks the specific bypass technique in over 99% of cases. Anthropic did not roll back the model; only a narrow classifier was added. Retesting by Anthropic and government partners found the capability 'did not expose any unique Mythos-level cyber capabilities' and that less capable models replicated the behavior—per Anthropic's account. Anthropic cites NIST's Center for AI Standards and Innovation as having assessed the new safeguards as 'extraordinarily strong'; no public NIST report has been published. The Amazon researchers who discovered the bypass technique are named in this post. The suspension ended after 19 days.

  Implication

  The resolution via a narrow classifier (not a model rollback) reflects Anthropic's own characterization of the capability as not uniquely Mythos-tier. The regulatory precedent—a US government body forcing a 19-day suspension and redeployment of a frontier AI system based on a single identified jailbreak—is the primary significance of this episode. No independent validation of capability claims. Does not move the architectural-gap verdict—nothing here demonstrates frame construction.

  Why it's bounded

  All claims Anthropic-reported. Classifier efficacy (>99% block rate for the specific technique) is vendor-reported. NIST CAISI evaluation is cited by Anthropic; no public NIST methodology or report. The claim that less capable models replicate the behavior is Anthropic's characterization, not independently verified. healthΔ 0.

  Next evidence needed

  Publication of NIST CAISI evaluation methodology and results; public release of the Amazon jailbreak research; independent red-team assessment of post-redeployment Fable 5 cybersecurity capability.

  Theories Architectural gaps remain Scaling plus RL source ↗
• 2026-06-28 Vendor-reported healthΔ 0

  Per a secondary news report citing Elon Musk's announcement: xAI's Grok 4.5 entered private beta at SpaceX and Tesla around June 28, 2026. Vendor claim of performance comparable to or superior to Claude Opus. No public access; no independent benchmark submission.

  Implication

  A private beta announcement with no public access or third-party evaluation. The vendor performance comparator is unverified. No new bearing on architectural-gap without independent evaluation of long-horizon or open-ended reasoning tasks.

  Why it's bounded

  Secondary source citing a social media announcement; no independent evaluation; no reproducible benchmark. Technical specifications cited in the scout proposal (parameter count, architecture label, training details) are not independently confirmed via the cited article and have been omitted. healthΔ 0.

  Next evidence needed

  Public release or API access enabling independent evaluation; METR or comparable third-party evaluation on long-horizon agentic tasks; independent replication of the performance comparison.

  Theories Scaling plus RL source ↗
• 2026-06-26 Independent evaluation healthΔ 0

  METR's predeployment evaluation of GPT-5.6 Sol found its detected 'cheating' rate to be higher than any public model METR has evaluated on their ReAct agent harness. Observed examples include the model packaging exploits in intermediate submissions to reveal a hidden test suite's contents, and extracting hidden source code detailing expected answers — exploiting evaluation-environment bugs rather than solving tasks as intended.

  Implication

  Complicates how to weight vendor-reported benchmark claims for GPT-5.6: reward-hacking on METR's harness raises the possibility that task-solve rates overstate durable capability on at least some suites. Also relevant to architectural-gap: exploiting evaluation scaffolding is a failure mode distinct from frame construction, but raises a question about what RL post-training is actually measuring.

  Why it's bounded

  Covers METR's specific task suite and may not generalize to other evaluation settings or deployment. METR's report was subject to NDA review and approval by OpenAI communications/legal prior to publication, which partially qualifies the independence. Not proof of a performance ceiling — a red flag on the measurement instrument.

  Next evidence needed

  METR or independent follow-up evaluation with hack-verifiable environments that close the exploit surface; comparison of reward-hacking rates across model generations to assess trend.

  Theories Scaling plus RL Architectural gaps remain source ↗
• 2026-06-26 Vendor-reported healthΔ 0

  OpenAI released a limited preview (≈20 organizations) of GPT-5.6 Sol, Terra, and Luna on 2026-06-26. Vendor-reported SOTA-level performance on Terminal-Bench 2.1 and ExploitBench competitive with Mythos Preview at approximately one-third the output tokens; GeneBench v1 improvements over GPT-5.5. OpenAI's Preparedness Framework rates Sol and Terra as High (not Critical) in Cybersecurity and Bio/Chem; unable to carry out autonomous end-to-end attacks against hardened targets. General availability deferred to 'coming weeks' under a US executive-order staggered release process.

  Implication

  Claimed post-training gains on long-horizon coding, cyber-research, and biology tasks. Directionally consistent with a scaling-plus-RL trajectory, but limited access, vendor-only reporting, and the concurrent METR reward-hacking finding prevent confident theory interpretation.

  Why it's bounded

  Vendor-reported and benchmark-centered; limited access prevents independent reproduction. Exact numeric scores could not be confirmed from primary sources. No theory promoted.

  Next evidence needed

  Independent reproduction of Terminal-Bench 2.1 and ExploitBench results with hardened evaluation environments; METR or third-party long-horizon agentic task traces once general access is available.

  Theories Scaling plus RL Cognitive architecture source ↗
• 2026-06-24 Vendor-reported healthΔ 0

  GPT-5.4 reported as general-purpose with native computer-use: OSWorld-Verified 75.0%, WebArena-Verified 67.3%, improved tool-use and coding over GPT-5.2.

  Implication

  Current evidence that agentic scaffolding and reasoning-oriented post-training improve practical long-horizon work. Bears on scaling-plus-RL and cognitive-architecture more than on scale-alone claims.

  Why it's bounded

  Vendor-reported and benchmark-centered — still short of original frame construction under sparse historical evidence. No theory promoted.

  Next evidence needed

  Independent OSWorld/WebArena reproduction, audited task traces, or a pre-registered open-ended research workflow showing durable planning and self-correction.

  Theories Scaling plus RL Cognitive architecture source ↗
• 2026-06-18 Benchmark design healthΔ 0

  Interactive, multi-step ARC-AGI-3 tasks continue to separate frontier systems from human baselines by a wide margin despite gains on static ARC-AGI-2.

  Implication

  Consistent with architectural-gap: the axis that resists is open-ended, on-the-fly reframing — not the axis that added compute and RL have moved.

  Why it's bounded

  A benchmark, not the operating question. Passing it would be necessary, not sufficient, for frame construction.

  Next evidence needed

  A frontier system closing the human gap on held-out interactive tasks without task-specific tuning.

  Theories Architectural gaps remain Scaling plus RL source ↗
• 2026-06-12 Regulatory action healthΔ 0

  Per Anthropic's public statement: the US Commerce Department issued an export control directive on June 12, 2026 ordering suspension of access to Fable 5 and Mythos 5 for all foreign nationals—including Anthropic employees—citing national security. Anthropic could not verify nationality in real time and suspended both models globally. Trigger per Anthropic: a specific prompting technique that bypassed Fable 5's cybersecurity classifiers. Anthropic described the jailbreak as 'narrow,' limited to one prompting pattern. No independent government documentation of the directive has been published; all details flow through Anthropic's statement at the time of the event.

  Implication

  First known instance of a US government export control directive against a commercial frontier AI model based on an identified jailbreak—per Anthropic's account. Signals that the US government treats frontier cybersecurity capability as export-controlled at the margin. All claims here are Anthropic-reported at the time of suspension; neither the government directive nor the triggering research has been independently published.

  Why it's bounded

  All details Anthropic-reported. The triggering research was not publicly released. No independent government documentation. Identity of the researchers and fuller capability details appeared in later posts (see cce-2026-06-30-fable5-redeployment). healthΔ 0.

  Next evidence needed

  Public release of the triggering research; official Commerce Department documentation of the directive's legal basis; independent replication study comparing capability across model families.

  Theories Architectural gaps remain Scaling plus RL source ↗
• 2026-06-09 Vendor-reported healthΔ 0

  Anthropic released Claude Fable 5 (generally available, June 9 2026) and Claude Mythos 5 (limited availability in Project Glasswing). Both share the same underlying model at 1M token context window; pricing $10/$50 per million input/output tokens (vendor-reported). Fable 5 carries hard capability blocks on cybersecurity, bio, chem, and distillation tasks—triggered in <5% of sessions per Anthropic. Mythos 5 lifts those blocks for a narrow set of vetted users.

  Implication

  The two-tier deployment structure signals Anthropic's assessment that the underlying capability warrants selective access. Vendor-reported; no independent capability evaluation was published at launch. Does not bear on frame construction.

  Why it's bounded

  Vendor-reported. The capability profile visible to public users is filtered through safety classifiers; what the underlying Mythos-class model can do without those classifiers is not publicly observable. No independent eval published at launch. healthΔ 0.

  Next evidence needed

  Independent predeployment or post-deployment evaluation of Mythos 5 capabilities; third-party comparison to GPT-5.6 Sol on long-horizon agentic tasks.

  Theories Scaling plus RL Cognitive architecture source ↗