Repository guide · 0 diagrams

View source on GitHub ↗
Original documentation, preserved from the repository. Historical projections and scenario ambitions are not evidence of live performance. See the current readiness record for deployment requirements.

SUCCESSOR Ω engineering security review

Review scope: local source review and adversarial tests of mission commitments, signing/proof, action authorization, persistence, portability and the existing computer-work bridge. This is an engineering review, not an external security audit, independent mission examination or production commissioning.

Base: 5b4cebb309a83a7a6749d8911d8bf96a1921e042. Generated release evidence records the checkout revision, working-tree state and executable source digest; the release PR identifies the final head. The acceptance matrix records validation scope; release notes preserve remaining gates.

Reviewed findings and corrections

Finding Correction and evidence
A signed mission-A context could refer to a mission-B candidate Proof verification now binds the candidate's mission to the signature context; the cross-mission test in test/proof-runtime.test.mjs rejects it
A trusted signature could carry internally contradictory PASS fields Verification checks comparator roster, finite statistics, costs, attempts and hard-gate consistency as well as signatures; contradictory signed reports fail the proof/runtime tests
A signer bypassing the proof issuer could promote public fixture evidence to I2 Proof verification now independently rejects synthetic-public I2–I4 claims, even when the trusted signer permits those levels. An adversarial regression signs the receipts directly, checks rejection, and preserves valid I0/I1 fixture labels
Distinct case hashes could be mistaken for independent samples Protocol explicitly commits independent-case groups, one case per group and custodian assurance; repeated correlation groups fail. Bounds are labelled separate one-sided bounds; actual independence remains an external evidence obligation
Credential aliases or mislabeled authority levels could imply extra permission Key fingerprints/custody, named principal trust, admission maximum level and effect classification are checked; A3 cannot relabel a consequential-payment effect to avoid A4 dual controls
Dependency closure was initially separate from active runtime state Compilation registration connects verified seals and mission-qualified source/job dependencies to durable impairment; linked proof and authority contract while unrelated missions remain intact. test/institution.test.mjs checks actual compiled invoice and fresh-lease denial after rights loss
CLI input could be replaced between a path check and reading; the proof harness checked its size only after allocation src/json-file.mjs opens without following a final symlink, checks and reads the same descriptor with a hard byte ceiling, refuses nonregular inputs without blocking, detects observed mutation and closes on every path. Malformed JSON errors are generic, so parser snippets cannot echo signer/trust material into logs. CLI and proof harness share it; six regressions cover replacement, symlinks, growth, mutation, UTF-8 and FIFO handling
A state path could follow a symlink; malformed knowledge expiry could remain current Store rejects nonregular/symlink paths, including dangling symlinks; knowledge expiry is validated and compared as time. test/store-pack.test.mjs covers both reproduced cases
Known credential aliases or modified restored lineage could enter portable data Additional credential-field checks and original-pack revalidation reject those cases. Public journey export also matches supported fixture bytes and recomputed results; store/pack and journey tests cover altered content
Outer metadata could be stripped or changed while reusing a task-v1 approval Protected admission binds successorManifestSha256; removing metadata cannot downgrade a sealed job into legacy execution. The orchestrator bridge tests cover unchanged task/changed outer envelope
Retries or revocation races could repeat effects Durable reservations, nonce/idempotency bindings, a pre-dispatch recheck and unknown-outcome reconciliation prevent blind replay in tested fixture drivers. Revocation after committed dispatch remains an in-flight event
Economic observations could be confused with mission proof Local-chain reconciliation binds original specification/result bytes and exact registry payout/finality evidence; a paid FAIL evaluation remains failed and unadmitted with no authority
A generic proof rehearsal could be mistaken for examination of the actual WORLD candidate src/examination.mjs freezes all 16 actual bindings and reexecutes the 60 public cases against the actual comparators. test/examination.test.mjs checks the signed I0 FAIL, tamper rejection, absence of private keys and independent-admission denial

Package test paths in the table are relative to packages/successor-core/; bridge tests are apps/orchestrator/__tests__/successorComputerWork.test.ts, and the actual local-chain case is test/v2/JobRegistry.test.js.

Boundaries that remain material

Keep source, manifests, failure findings and recovery evidence. Any material code, configuration, supplier or trust change requires the affected checks again and, where applicable, new proof/admission. No clean local test suite establishes universal ASI, customer Alpha or production authority.

← Back to the collection