Pranav Bhave

Noetic log 002 · published 2026-08-23 · review published 2026-10-17 · 24 days late

The observatory ships: checks and an overdue lookback

The September 23 lookback was missed. This October 17 update answers all six promised questions, preserves the original contract, and replaces owner grades with dated outcomes and remaining unknowns.

Editorial revision — 2026-10-17. The original deadline was 2026-09-23. The review below is published 24 days late; it does not reset any claim’s review window. Read the six answers · Original promise · Current checks and review dates.

What shipped — 2026-08-23

The record became an observatory: a thirteen-claim registry (schema v0.3) rendered three ways (ledger, capsule field, six modules under one grammar), an interactive feasible-worlds instrument whose geometry is asserted in CI to 1e-9, clean-clone reproduction of bounds and endpoint witnesses, and a new gate that clones every bound repository to prove each pinned commit is reachable from its default branch — because one binding was found stranded in rewritten history, and byte-watching triggers were structurally blind to it. An independent hostile review filed twelve findings; all twelve were fixed before shipping, one of them a factually false sentence in the module that lectures about expiry. Log 001 carries the lineage of this failure class.

What passed, what failed, and what remains untested — 2026-10-17

This replaces the owner’s eight numerical grades. The original August 23 entry remains in repository history. Passing one check establishes its declared condition, rather than a grade for the whole project.

Recorded outcomes and their limits as of October 17, 2026
AreaOutcome and dateEvidence and remaining limit
Bound-kernel reproductionPASSED · 2026-10-12The reproduction step passed in the October 12 run. It does not establish scientific execution or inferential validity outside its checks; see check coverage.
Full weekly checksFAILED · 2026-10-12The October 12 full weekly run failed. The last successful weekly run before it was September 21. Full failure logs were unavailable during review; the cause remains unknown. Inspect run links and current status.
E2 measurementINCOMPLETE · observed 2026-10-17No conforming completed E2 result was found in the public record. The published E9 result uses different guards, pools, and operating points; it does not complete E2. Inspect current work and blockers.
Research claims correctedREJECTED · 2026-09-10E6’s fixed-marginal range and E7B’s preregistration-conforming claim were rejected. Reproduction alone did not catch every execution or inference error. Read the dispositions.
AccessibilityPARTIAL · observed 2026-10-17The prior live homepage’s menu opened, Escape closed it, and focus returned. That sampled interaction is not a full keyboard-only traversal. A complete screen-reader pass has not been run.
Unfamiliar-reader effectsUNTESTED · observed 2026-10-17No completed comparison establishes improved comprehension, calibration, or evidence checking. REL-001’s review window first expired September 23 and was renewed by the owner on October 17. A renewed window is not a completed reader study.

Original research direction — 2026-08-23

The following paragraph records the direction stated at publication. The dated lookback below describes what subsequently changed.

Dangerous in the only sense this record endorses: menace pointed at claims, never at people. The site's machinery — envelopes, bounds, witnesses, expiry — currently prosecutes only its own claims. The move that turns it from a portfolio into a threat to overclaiming is to point it outward: take published AI-safety numbers from the wild, compute what their stated evidence actually licenses under the Fréchet bounds, and publish the interval next to the headline number with witnesses at both ends — a claim autopsy, powered by exactly the E2-style measurement this program has frozen but not yet run. The standard, made visible, applied where it stings. That requires E2 to exist first, which is why the lookback below has teeth.

Original lookback contract — due 2026-09-23

Historical wording from 2026-08-23 is retained below. References to scores refer to the numerical self-grades removed in the October 17 editorial revision. The December clock forecast was wrong; see answer 5 below.

On or after that date, answer each question against evidence that exists outside this page. Anyone can run this checklist; none of it requires trusting the owner.

  1. Did the standing gate stand? The weekly verify runs (five expected by then) are public in the Actions history. Any red run, and what fired.
  2. Did E2 move? Either a conforming observation set exists, or a documented failure-to-obtain exists. If neither, the "next artifact" promise on /now/ is thirty days stale and this entry says so.
  3. Were the open QA items closed? Keyboard-only and screen-reader passes were disclosed as not-run at ship time. After thirty days, an open disclosure becomes a broken promise; the accessibility score above drops to 5 until they run.
  4. Did any external falsifier arrive? Corrections, issues, or mail against any claim — and what the registry did about them.
  5. Do any decay clocks show amber? None should before December; amber earlier means a review was triggered, and the ledger must show the re-review.
  6. Did the claim count move for the right reason? New claims should arrive because evidence arrived — not because surfaces multiplied.

Pre-registered conclusion if nothing moved: the observatory spent its first month documenting stasis beautifully — an instrument without an experiment — and the research-threat score falls from 5 to 3. That sentence is written now, while it can still be falsified by work.

Lookback disposition — 2026-10-17, published 24 days late

The original promise is retained above. These answers describe the public artifacts inspected on October 17; unknowns remain explicit.

  1. Did the standing gate stand? Three scheduled checks failed. The full weekly runs on September 28, October 5, and October 12 failed. The last successful weekly run before those failures was September 21. October 12’s bound-kernel reproduction step passed; the whole workflow did not. Full failure logs returned an access error during inspection, so this review does not assign a cause. The status page links the verified run records and updates separately from this dated snapshot.
  2. Did E2 move? It remains incomplete; E9 is a separate published result. No conforming completed E2 result was found. Its owner-authorized collection and freeze requirements remain unresolved in the E2 record. The 1,000-item E9 result merged in PR 26, with different guards, pools, and operating points. Its scoped publication does not fulfill E2’s frozen contract. The research board makes that distinction visible.
  3. Were the open QA items closed? Partly; the full promise remains open. The prior live homepage’s mobile menu opened, Escape closed it, and focus returned to the button. This is one sampled keyboard interaction, not a complete keyboard-only traversal. A complete screen-reader pass has not been run. A sampled browser inspection of the revised homepage at 390 × 844 confirmed that the subject, email link, correction, and fetched FAILED status were visible. This is a viewport check, not a complete keyboard traversal. No unfamiliar-reader comprehension study has been completed.
  4. Did any external falsifier arrive? Public corrections exist; the full arrival history is unverified here. The correction record records the missed census artifact and subsequent claim dispositions, including E6 and E7B’s September 10 rejections. Those are observable changes to the public record. This review has not independently established which arose from an external submission, and it has not inspected private mail. It therefore cannot claim either that no external falsifier arrived or that every arrival was handled. Public issue and correspondence intake remain an open audit item.
  5. What happened to the review clocks? The elapsed window was renewed. The August forecast “None should before December” was wrong: REL-001 initially reached its September 23 review deadline. The owner renewed its last review to October 17 in the current registry; the next review is November 16. No completed reader study is thereby established. Expiry changes review status, not the hypothesis’s truth value. The deadline queue is derived from registry dates and keeps elapsed deadlines visible until a disposition is recorded.
  6. Did the claim count move for the right reason? The increase alone cannot answer that. The August 23 entry recorded thirteen claims; the current registry contains 21 claims. That difference does not establish eight new validated findings. A post-result registration of the E9 result, E9-001, is proposed in draft PR 31 and is not counted here. Each claim must be assessed through its evidence, status, history, and limits. The rejected claims, incomplete E2 study, and scoped E9 result remain visible rather than being counted as completed evidence.

Disposition: this review missed its deadline. The public record contains later work and corrections, so the original “nothing moved” condition is not established. It also contains failed full checks, an incomplete E2 study, outstanding accessibility work, and no completed reader comparison. Those unresolved items remain in the current status and research views; this entry does not award a replacement grade.

Limits of this review — this is a dated editorial review of linked public records. It does not independently validate every claim, verify private correspondence, or establish reader effects. Check outcomes and claim review dates remain distinct.