Skip to content

CLAIMS.md

Latest commit

 

History

History
97 lines (60 loc) · 4.09 KB

CLAIMS.md

File metadata and controls

97 lines (60 loc) · 4.09 KB

AIEP — Defended Claims

Every claim listed here is:

  1. Mechanically enforced by the kernel
  2. Testable with a specific, machine-readable vector
  3. Bounded by explicit conditions

This document is the attack surface on our terms. If you want to challenge AIEP, start here. Every claim has a test. Run it.

Claim 1 — Deterministic Replay (AIEP-REPLAY-001)

Exact claim:

Any two AIEP-compliant implementations given identical evidence_ledger, reasoning_ledger, and schema_version_id with the same LOCKFILE_VERSION MUST produce identical pack_hash.

Mechanism: pack_hash = sha256_hex(canonical_json({evidence_ledger, reasoning_ledger, schema_version_id})). Determinism follows from R1–R8 of CANON_SPEC.md — every rule eliminates a source of non-determinism.

Holds when: Implementation follows CANON_SPEC v1.0.0 exactly. Fails when: Different LOCKFILE_VERSION, or any deviation from R1–R8 (key ordering, NFC, number normalisation, length-prefixed concat, etc.).

Test vectors: AIEP-TV-DET-001 (basic), AIEP-TV-DET-002 (NFC edge case), AIEP-TV-DET-003 (empty state)

aiep-verify --vectors vectors/v1.0.0/

Claim 2 — Tamper Detection

Exact claim:

Any modification to a committed decision record — evidence content, reasoning content, or schema_version_id — produces a detectable pack_hash mismatch against the stored value.

Mechanism: Hash of the full state. Any change to any leaf value cascades to a different root hash. Detection requires only: recompute and compare.

Holds when: Records are stored with their pack_hash intact at or after commitment. Fails when: A write-access attacker modifies both the record AND the stored pack_hash simultaneously. See BOUNDARIES.md §B2.

Test vectors: AIEP-TV-TAMP-001 (evidence tampered), AIEP-TV-TAMP-002 (reasoning tampered)

Claim 3 — Evidence Boundedness

Exact claim:

The system will not advance a conclusion without committed evidential support. Evidence gaps trigger a committed NegativeProofRecord, not silent omission.

Mechanism: CC-001 (plausibility filter) and the negative proof protocol (P16) together ensure that absence of evidence is itself committed as negative_proof_hash(type, window_start, window_end, schema).

Holds when: CC-001 and negative proof protocol are enforced before conclusion advancement. Fails when: Caller bypasses evaluate_admissibility (misuse, outside system boundary).

Test vectors: AIEP-TV-NEG-001 (absence committed), AIEP-TV-NEG-002 (absence in pack_hash)

Claim 4 — Fail-Closed Default (CC-005)

Exact claim:

OPEN is the exception. CLOSED is the constitutional default. Any gate not explicitly set to OPEN blocks conclusion advancement.

Mechanism: evaluate_admissibility returns (False, "CC-005: Gate is CLOSED...") whenever gate_status != "OPEN". There is no fallback output — only block.

Holds when: evaluate_admissibility is called and its result respected. Fails when: Caller ignores the return value (misuse).

Test vectors: AIEP-TV-ADM-001 (gate CLOSED blocks passthrough admissible conclusion)

Claim 5 — Dissent Structurally Blocks Advancement (CC-004)

Exact claim:

A conclusion contested by reject or escalate dissent cannot advance until dissent is structurally resolved. No collapse to a single confident output.

Mechanism: evaluate_admissibility checks dissent_final_position and returns (False, "CC-004: Dissent unresolved...") for reject or escalate.

Holds when: evaluate_admissibility is called with the resolved dissent position. Fails when: Caller bypasses arbitration (misuse).

Test vectors: AIEP-TV-ADM-002 (dissent rejects), AIEP-TV-ADM-003 (dissent accepted — positive case)

What Is NOT Claimed Here

See BOUNDARIES.md for the full boundary documentation.

Short version: AIEP does not claim that evidence is accurate, that the underlying model is correct, or that write-access attackers cannot construct internally consistent false records. These are explicitly documented non-claims.