Independent Validation Hardening Contract

Purpose

Define the minimum record for proportional independent validation when a Work Slice needs more than maker self-assessment. This contract strengthens validation without making LLM review proof, creating a new lifecycle, or requiring critic review for every low-risk task.

Boundary

Independent validation is a review layer over existing AletheIA authorities:

Intent and expectations remain governed by the Intent-to-Evidence Work Slice Extension.
Separate verifier requirements remain governed by the Maker-Checker Policy.
Objective gates remain governed by the Objective Gate Policy.
Harness evidence and action records remain governed by AHGE, AHC, Execution Audit Record and Agent Action Audit Record.

This contract adds the shape of the independent validation context and review record. It does not assign reviewers, run benchmarks, approve merges, deploy artifacts, expose hidden chain-of-thought, or replace human review.

When to use

Use independent validation when one or more of these conditions are present:

independent_validation_recommended_when:
  - governance_error_risk
  - contractual_error_risk
  - human_intent_may_be_misread
  - evidence_coverage_is_partial
  - maker_checker_mandatory_when_applies
  - high_cost_or_high_rework_consequence

Independent validation is optional for low-risk, low-impact work when an objective gate is sufficient and no maker-checker trigger applies.

Critic context

A critic receives artifact and evidence context, not hidden actor reasoning. The minimum context is:

independent_critic_context:
  context_id:
  work_slice_ref:
  artifact_refs: []
  intent_ref: not_needed | string
  expectations_ref: not_needed | string
  evidence_refs: []
  source_refs: []
  known_constraints: []
  known_gaps: []
  review_focus: []
  privacy_boundary:
    prompt_content: excluded | summarized | authorized_excerpt
    restricted_sources: metadata_only | not_present
    secrets: excluded

Required rules:

source_refs must identify the source records used to form the review context.
known_gaps must stay visible; unavailable evidence is represented as unavailable, not invented.
Hidden chain-of-thought, private prompts, secrets and restricted source content must not be transferred to the critic.
The critic may inspect artifacts, expectations, evidence and permitted summaries; it may not become the source of truth.

Evidence-to-expectation gate

A clean proceed verdict is invalid while a required expectation remains unproven.

evidence_to_expectation_gate:
  gate_id:
  expectations:
    - expectation_ref:
      required: true
      evidence_status: proven | partially_proven | not_proven | not_applicable | unavailable
      evidence_refs: []
      gap_reason:
      disposition: proceed | proceed_with_gap | revise | block | human_review_required
  overall_disposition: proceed | proceed_with_gap | revise | block | human_review_required

Disposition rules:

proceed requires every required expectation to be proven or not_applicable with a source-backed reason.
proceed_with_gap requires explicit accepted limitations and reviewer or human sign-off reference.
revise means the maker should change the artifact before closure.
block means the slice must not close or advance until the blocking gap is resolved.
human_review_required means the critic cannot resolve the risk or authority boundary.

Validation review record

validation_review_record:
  review_id:
  work_slice_ref:
  reviewer:
    independence: independent_critic | separate_agent | separate_human | mixed
    reviewer_ref:
    relationship_to_maker: none | separate_role | separate_team | unknown
  critic_context_ref:
  evidence_gate_ref:
  findings:
    - finding_id:
      severity: low | medium | high | critical
      claim:
      expectation_refs: []
      evidence_refs: []
      source_refs: []
      recommendation: proceed | revise | block | human_review_required | log_only
  verdict: proceed | proceed_with_gap | revise | block | human_review_required
  confidence: low | medium | high
  limitations: []
  human_review_ref: not_required | pending | string
  audit_refs: []

A validation review record is not valid when:

it has no source_refs for a finding;
it reports proceed while a required expectation is not_proven, partially_proven without accepted limitation, or unavailable;
it treats an LLM critic as proof or final authority;
it hides the review limitations.

Before accepting a validation review, confirm:

The reviewer is not the maker, or the lack of independence is explicitly marked unknown and cannot satisfy maker-checker requirements.
The critic saw the artifact, expectations and permitted evidence, not hidden chain-of-thought.
Each important finding points to source_refs and evidence_refs when available.
Required expectations with missing evidence cannot produce proceed.
Human review remains required for sensitive, irreversible or critical decisions.
The record states limitations and unavailable evidence neutrally.

Observatory projection

Read-only observability may show the existence, disposition, confidence, missing evidence and source references of independent validation. It must not show private prompts, hidden reasoning, secrets or restricted source content.