TRANSACTIONAL TRUST · VERIFIED #026

The fingerprint was correct. The causal model was not.

Dependency-aware applicability only works if the dependency model is complete enough for the decision. This executable benchmark omits a real causal input and shows a perfectly stable fingerprint authorizing the wrong consequence.

Ground truth

output = min(limit, 2 × price + tax_rate)

authoritative dependencies:
price + limit + tax_rate

artifact declares:
price + limit

Unsafe: a perfect fingerprint over the wrong graph

tax_rate: 2 → 8

declared fingerprint(price, limit):
UNCHANGED

correct output:
22 → 28

declared-only adoption:
rows = 1
HTTP 200
committed output = 22  ❌

The hash did exactly what it was asked to do. The mistake was upstream: tax_rate was absent from the declared causal graph.

Safe: validate the dependency manifest first

artifact manifest:
price + limit

authoritative contract:
price + limit + tax_rate

→ dependency_manifest_conflict
→ adoption rows = 0
→ external effects = 0

Worker B then recomputed using the complete causal contract and committed output 28 exactly once.

I63–I66

• I63 — A correct fingerprint over an incomplete dependency set is still unsafe.
• I64 — Dependency-set identity is part of applicability evidence.
• I65 — Adoption must validate the dependency manifest, not only fingerprint values.
• I66 — Omitted or unknown causal input requires revalidation, recomputation, or hold before consequence.

VALUE INTEGRITY ≠ MODEL COMPLETENESS.

Evidence

GitHub Actions run: 31583719848
Artifact ID: 9136209133
Artifact digest:
sha256:6bbe0b9a78e19ea9933f3389b3c9d99e7931b2d5d87b183b9d950aaaf8930f48
Score: 10/10

This is a deterministic protocol benchmark, not production safety certification or a vulnerability claim against PostgreSQL or another external product.