Corrections, clarifications, and retractions
A correction is not cleanup around the evidence. It is evidence about whether the practice deserves trust.
Report a factual error, missing material context, broken source, conflict, privacy issue, or misleading visual to cmcnosky@gmail.com. Include the exact passage and the strongest record supporting the correction.
Ledger
Five same-day P2 clarifications were published on 2026-07-26 after separate source and parity red-team passes.
None of the clarifications changed the operator decision. They narrowed overbroad wording, added missing historical context, or separated concepts that had been combined too loosely. The affected LinkedIn mirrors have not been published; each planned copy carries its correction ID and this ledger link.
NF-C001 · P2 clarification · NF-007 · 2026-07-26
One observation can contribute to an estimate without establishing reliability
Original: The page said one successful run did not estimate how often the outcome would happen, and the planned LinkedIn copy called it observation rather than reliability.
Corrected: The record now distinguishes the observed outcome, an estimator under explicit assumptions, the uncertainty that one observation cannot quantify from the data, and the stronger burden for a defensible reliability claim.
Why: NIST AI 800-3 treats evaluation metrics as estimates of declared targets. The earlier categorical wording erased that statistical distinction.
Operator decision: Unchanged. One run alone still does not justify a defensible reliability or failure-rate claim.
NF-C002 · P2 clarification · NF-008 · 2026-07-26
A leaderboard may be screening evidence
Original: The page called a leaderboard without a measurement contract a sortable screenshot and did not identify the exact six-field sequence as a NoFuckery criterion.
Corrected: The page now calls it screening evidence rather than a purchasing or deployment decision record and states that NIST does not prescribe the exact sequence.
Why: The earlier opening was stronger than its own countercase, and independent replication is a NoFuckery decision criterion rather than an exact NIST requirement.
Operator decision: Unchanged. Rank alone does not grant deployment authority.
NF-C003 · P2 clarification · NF-R04 · 2026-07-26
Evidence lineage and source stance are separate dimensions
Original: The operator action grouped first-party, affiliated, materially independent, and adversarial into one source classification.
Corrected: The operator action now records lineage and stance separately and makes independence specific to an evidence path for a particular claim.
Why: Adversarial describes posture or incentive; the other terms describe provenance. A source can be both adversarial and materially independent.
Operator decision: Unchanged. Map evidence lineage before claiming independent corroboration.
NF-C004 · P2 clarification · NF-R02 · 2026-07-26
The cited Operator card is historical first-party evidence
Original: The page generalized to every system card and cited the January 2025 Operator card without stating that the standalone Operator experience was later integrated into ChatGPT agent and sunset.
Corrected: The page now limits the verdict to developer-authored cards, allows inspectable linked independent evidence, and labels the Operator card as historical rather than current-product documentation.
Why: A card can incorporate independent evidence, and a historical product card should not be read as current product documentation.
Operator decision: Unchanged. Translate card claims into deployment-specific tests, monitoring, or documented unknowns.
NF-C005 · P2 clarification · NF-R03 · 2026-07-26
An opaque demo may establish less than an arranged-condition observation
Original: The page said a polished demo established an arranged-condition observation and required repeated trials in the acceptance protocol.
Corrected: The page now says a demo is at most evidence of an arranged result and may establish less without an inspectable execution record; the protocol now requires repeatable testing and vendor-inaccessible cases.
Why: An opaque or edited demo may not prove that the represented system produced the result, and the cited source packet did not directly support the stronger repeated-trials language.
Operator decision: Unchanged. A selected demonstration is not acceptance evidence for consequential deployment.
