Pokémon Red Cartridge Must Clear Screenshots Before Sonnet Mid-Tier Seats
A laminated Pokémon Red cartridge on the duty window must clear from screenshots alone before Claude Sonnet 5.5 takes mid-tier filing, a bureau said.
Daily field notes from the model frontier
AI Panorama
Models, tooling, evals — short reads with clear takeaways. Signal only. No fluff.
VISITOR Badge Must Hang Overnight on Anthropic Board Before Assessments File
A VISITOR-stamped badge stays on Anthropic's staff badge board overnight before a third-party reviewer files a safety assessment. Reviewers work from desks labeled NOT EMPLOYEE.
SAME NAME Stamp Returns Flag Packets After Letterhead Hits Phone Book
A duplicate name voids the run and the office sends packets back stamped SAME NAME. No capture-the-flag filing opens until the fictional firm's letterhead matches no listing in the clerk's phone book.
Mistake Ledger Must Be Cut Before Half-Price GPT-6 Sol Rings
A clerk must cut the mistake ledger in half, date-stamp both pieces, and hang them on the price tag before the till will ring half-price GPT-6 Sol.
Evaluator Laptop Must Stay Open Overnight Before Training Run May Pace
An evaluator laptop must stay open overnight on a labeled desk before a training run may file as paced. Runs with the lid shut are entered as unpaced.
Bureau of Posted Pelicans Accepts Launch Score Only as a Pelican on a Bicycle
The Bureau of Posted Pelicans said a launch score is accepted only as a pelican on a bicycle on the monitor at the window. World saves without the pelican are entered as notes.
Office of Hearings Voids Safety Evals Not Initialed as Lobby Talking Points
The Office of Hearings will not docket a safety eval until the lobby desk rewrites it as a talking point with a margin initial. Staff return filings that cannot be read aloud.
Interlab Bureau of Score Authorship Will Post Public-Set Results Only in the Name of a Certified Accompaniment Harness
Models are filed as non-scoring apparatus and barred from the board. Unaccompanied exams return ungraded, not failed; certified harnesses hit 100 percent on the public set.
Office of Model Claims Requires Inspectors to Ship as Downloadable Plugin Bundles So Audited Firms Can Reproduce the Decision to Begin Inspection
Office of Model Claims will issue inspectors as downloadable plugin bundles so audited firms can reproduce the decision to begin inspection. Filings cite "inspector loop" as replaceable.
Interministerial Commission Classifies Grok 4.6 Desk-Bound After Knowledge Scores Outrun Terminal Results
The commission classified Grok 4.6 desk-bound after knowledge scores outran terminal results, barring uncleared models from the command line via a sedentary inference registry.