audit-validator-independence

SkillDev tools

Lets your agent check whether a test or benchmark can pass only because it shares the assumptions it claims to verify.

Available today. Use it from your connected AI after setup.

Add ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.

Then ask your AI: use the audit-validator-independence skill

About this skill

Audit whether a validator, benchmark, sandbox, or test can pass by construction because it embeds the same assumptions as the theory it claims to validate. Build a claim×validator-assumption matrix and demand at least one non-circular failure channel.

What this skill tells your AI

The instructions your AI receives, as published by yogsoth-ai/de-anthropocentric-research-engine in skills/audit-validator-independence/SKILL.md and read by ahel’s review.

Purpose

Detect validators that pass by construction because they copy the target theory's assumptions.

Input contract

mode_contracts:
  validator: &validator_audit_input
    required: [target_claim, validator, validator_artifacts]
    optional: [data_generation, metric, oracle, acceptance_rule]
    constraints: [validator_assumptions_and_target_assumptions_must_be_separately_listed]
  benchmark: *validator_audit_input
  sandbox: *validator_audit_input
  simulation: *validator_audit_input

Execution protocol

Do not perform called SOP operations inline; each loaded SOP owns its contract and thresholds.

  1. You MUST load skill enumerate-validator-assumptions to enumerate embedded validator assumptions.
  2. You MUST load skill build-noncircularity-matrix to cross-tabulate target and validator assumptions.
  3. You MUST load skill detect-pass-by-construction to detect copied assumptions. You MUST load skill design-falsification-test to design an independent failure channel. Deviation: if no non-circular channel is feasible, return blocked validation rather than a pass.

Output contract

mode_contracts:
  validator: &validator_audit_output
    produces: [noncircularity_matrix, red_cell_actions, validator_verdict, adversarial_ground_truth_set]
    delta_fields: [findings, evidence_updates, uncertainties, decisions, open_questions]
  benchmark: *validator_audit_output
  sandbox: *validator_audit_output
  simulation: *validator_audit_output

Thresholds and quality gates

  • At least one validator path must be independent of the target's defining assumptions.
  • Shared or derived-from-target dependencies must be explicit in the matrix.

Failure and counterexamples

Pass-by-construction is a finding, not evidence of corroboration. Unknown dependencies remain unresolved.

Provenance map

  • resolved: circular-validation-audit

Preserved source criteria ledger

sourcesource linekindsource criterion
v4 architecturenode desctextualBuild claim-validator assumption matrix and require a non-circular failure channel.

Context checkpoint / Delta notes

Append assumptions, matrix cells, circularity findings, and proposed independent test.

Mode branches

  • validator: inspect validator implementation assumptions.
  • benchmark: inspect benchmark construction and metric assumptions.
  • sandbox: inspect environment/oracle assumptions.
  • simulation: inspect simulator and initialization assumptions.

Signals

GitHub stars
501
Forks
41
Last commit
Sep 2026
Advanced
Catalog kind
skill
Key
audit-validator-independence
Source
github.com/yogsoth-ai/de-anthropocentric-research-engine