The argument in text
MC-001 records a correction dated 2026-08-30: the envelope moved from 19/13/4 to 20/14/5 when the row multimodal-safeguard-bench-2026 was added, under the claim's own REJECT falsifier; two sensitivities are declared, named-products-only 19/13/5 and pre-freeze-visibility-evidenced 19/13/4.
Exactly the revision recorded in census.yaml and bound in MC-001 at the freeze 2026-08-27; the counts are verify_census.compute_counts; one primary reviewer; a bounded search. The falsifier text on screen paraphrases MC-001's registered condition. Nothing here says the corrected census is complete, or that any evaluated system performs well or badly.
What this does not establish
- does not claim the corrected census is complete; a bounded search proves no universal absence
- does not claim any evaluated guardrail or stack performs poorly; an unfilled column is a reporting fact
- the sensitivities are declared alternatives, never replacements for 20/14/5
- does not claim the correction machinery is novel or a standard
What would require a correction
census.yaml revision_history no longer records the 2026-08-30 correction from 19/13/4 to 20/14/5 with the named row; or scripts/verify_census.py --counts prints an envelope other than 20/14/5 or sensitivities other than 19/13/5 and 19/13/4 (MC-001, REJECT) — bind_facts.py --check fails before this film can be re-rendered.
Read the supporting record: MC-001.
Sources and render record
Film manifest · Render receipt · First site deployment record
Current render: 2026-09-15. The receipt records the rendered inputs and output hash. It checks provenance, not scientific validity.