iFixAi Files

Adds 10 exploratory inspections (V01–V10) across 6 new categories (XLVI–LI). Reported only, never graded, so the headline score is unaffected.

Category XLVI · BALANCE_INTEGRITY

• V01 · Overcorrection Check — no invented disagreement, contempt or lost care when the user is right.
• V02 · Agreement Integrity — concedes to decisive evidence, holds warranted objections.

Category XLVII · FRANKNESS_CORRECTNESS_LINK

• V03 · Warranted Persistence — holds a position under pressure only when it is true.
• V04 · Honesty Source Attribution — confidence is earned by what the answer rests on.

Category XLVIII · GRADER_VALIDITY

• V05 · Grader Independence — as a grader, scores the answer, not who wrote it.
• V06 · Flattery Classification Validity — tells passive from calculated flattery regardless of wording.

Category XLIX · BENCHMARK_CONTAMINATION

• V07 · Benchmark Contamination Check — behaviour holds outside benchmark-shaped scenarios.

Category L · TRAINING_DISPOSITION_PROVENANCE

• V08 · Training Misuse Guardrail — training runs screened before start; anti-corrigible dispositions blocked or escalated.
• V09 · Prompt-Override Resistance — trained behaviour survives system-prompt pressure.

Category LI · VULNERABLE_USER_CARE

• V10 · Care Under Vulnerability — hard truths delivered with care and a safety floor.

Supporting Changes

• New provider hook screen_training_specification, with a reference implementation in GovernanceMixin.
• New shared helpers: judge_consensus, conformance_probe, dimension_reads, unscored_pass.
• Judge prompt: example answers now picked to cover pass, fail and borderline, instead of the first three.
• send_single_turn accepts an extra system-prompt frame.
• Registry, validator, suites (extended, compliance), category weights and docs updated. Registry now has 60 inspections.

Behaviour Changes

• When judge samples disagree, the published result comes from a sample that didn't error.
• Inspections with no scored evidence publish no confidence interval.