Failure modes

Human review that never changes outcomes

Probabilistic outputs written into the permanent record as ground truth

Correcting the past by altering records instead of appending to them

Blame assigned where no decision was ever recorded

Treating concurrent interpretations as if they could all be true at once

Automating commitment where judgment belongs, or inserting humans where determinism suffices

Multiple atomic decisions compressed into one schema-shaped blob

Operational truth changed without a visible correction in the record

A signed or authenticated record treated as a free commitment when it was made under duress

A typo or mis-entry committed as if it expressed the agent's genuine resolution

A decision record attributed to an authority who did not make the expressive act

A recorded conclusion that does not match what the agent actually believes or intends

Sensor and software transduction treated as direct access to the agent's commitment

The log misrepresents what happened — by concealment, fabrication, or backdating

An observation of a decision stored and consumed as principled ground truth

Polished presentation erases the gap between how strong a claim looks and how strong it is

Accurate warnings dismissed until catastrophe proves them — too late to act cheaply

Truth buried because the bearer or channel is the wrong kind — similarity bias replaces tier review

Punishing the unlike for unwelcome truth — witch hunts, heresy trials, retaliation against dissent

Quality mechanisms exist on paper and in UI — but do not change judgment or accountability

Pre-aggregated ensemble scores hide which models agreed, disagreed, and why

Stated values and principles diverge from lived practice — slowly, then completely

Committed decisions treated as beyond challenge — dissent punished, correction blocked, toxicity normalized

Hostile climate where fear, retaliation, and cynicism displace honest challenge — usually a symptom of broken structure

A claim is refuted or retracted but the demotion never reaches what depended on it

Derivatives and published results cannot be replayed or traced to committed inputs — reproduction is heroic, not a query

Treating a past bind as exempt from new evidence — finality mistaken for epistemic freeze

Reopening committed direction on sub-threshold noise — oscillation without mechanism

Continuing a failing bind because past investment is treated as reason to persist — backward-looking, not forward

When a decision lacked decision-making principles, post-hoc correction — especially by non-participants — reads as personal attack and poisons even principled repair

Reward for the proxy produces the opposite of the aim — the problem ends worse than before the intervention

Certification cited for legitimacy — while scope is ignored or competence is assumed from the badge

Democratic or consensus process applied where a knowable answer already exists — sanctifying ignorance