Skip to content

VERIFICATION.md

Latest commit

 

History

History
125 lines (101 loc) · 5.94 KB

VERIFICATION.md

File metadata and controls

125 lines (101 loc) · 5.94 KB
Raw

V7 Verification Record

This record defines the evidence required before the V7 development branch can be merged. Passing one methodology never substitutes for another.

Method One: Static and Structural

The static gate verifies source structure, generated artifacts, documented surfaces, immutable inputs, raw-result provenance, and privacy.

python3 -m py_compile mind.py src/mind/*.py bench/*.py tools/*.py
python3 tools/build_single.py --check
python3 tools/claims.py check
python3 tools/privacy_scan.py
git diff --check

Required invariants:

  • the modular source builds byte-for-byte into mind.py;
  • every documented command, flag, environment variable, protocol tool, test count, version, checksum, and result link matches executable state;
  • every public result comes from a clean tree and carries the current mind.py digest;
  • mutation reports use the current corpus, a green product-suite baseline, complete classification, and isolated confirmation for every parallel kill; the three self-referential claims tests run after report generation, and the report records both this exclusion and the exact baseline test count;
  • English and Arabic documentation expose the same marked sections;
  • no private path, email address, secret pattern, private audit, or local task ledger enters the public tree.

Method Two: Dynamic and Adversarial

The dynamic gate exercises normal behavior, boundary crossings, corruption, crashes, concurrency, privacy erasure, protocol handling, and long-horizon autonomy.

python3 -W error::ResourceWarning -m unittest discover -s tests -q
python3 bench/bench.py
python3 bench/multilang.py
python3 bench/discrim.py
python3 bench/slots.py
python3 bench/soak.py
python3 bench/fuzz.py --quick
python3 bench/autonomy.py --days 1825 --payload-chars 5000

The test suite includes a sparse 100 MB archive crossing, CRLF and BOM foreign writers, exact concurrent scheduler updates, one-winner leases, crash-resumable privacy operations, exact restore, tampered backup rejection, persistent semantic-server failure handling, malformed protocol messages, and deterministic journal-merge convergence.

Method Three: Differential and Metamorphic

The differential gate proves equivalence across representations and degraded conditions rather than merely checking one happy path.

python3 -m unittest -q \
  tests.test_distribution \
  tests.test_round14 \
  tests.test_merge \
  tests.test_storage \
  tests.test_scheduler

Required comparisons:

  • modular source versus generated one-file distribution;
  • offline ranking versus whole-ranking fallback after any semantic failure;
  • original export versus CRLF, BOM, and mixed-line-ending transformations;
  • both branch orders and randomized interleavings for journal merge;
  • backup digest versus restored digest, including interrupted restore recovery;
  • explicit empty protocol input versus end-of-file behavior;
  • current format versus prior-format fixtures.

Public Evidence

Area Raw result
LongMemEval, offline longmemeval-offline-v7-dev.json
LongMemEval, BM25 longmemeval-bm25-v7-dev.json
LongMemEval, concept sidecar longmemeval-concept-v7-dev.json
Paraphrase traps paraphrase-v7-dev.json
10,000-fact bulk ingest bulk-v7-dev.json
Five-year autonomy autonomy-five-year-v7-dev.json
One-file mutations mutation-mind-v7-dev.json
LongMemEval mutations mutation-longmemeval-v7-dev.json

The 50-question LongMemEval subset is deliberately reported as a subset. BM25 leads both evidence metrics on it. The benchmark does not measure graph traversal, temporal validity, contradiction handling, privacy lifecycle, merge, or autonomous consolidation, and therefore is not an overall product ranking.

Mutation rates are test-sensitivity measurements, not product-quality scores. Every candidate kill produced under parallel execution is rerun in isolation; parallel-only failures are reclassified and retained as diagnostic evidence.

Traceability

Scope Primary implementation evidence Primary test evidence
Availability and integrity, C-01 through C-05 filesystem, graph, scheduler, cortex, dream domains test_round14, test_scheduler, test_mind
Retrieval and benchmark trust, C-06 through C-15 embedding, recall, benchmark, claims tooling test_longmemeval_bench, test_mutation_bench, test_claims
Semantics and privacy, C-16 through C-22 graph metadata, lifecycle manager, storage commands test_lifecycle, test_storage, test_typed_memory
Governance and distribution, C-23 through C-26 workflows, ownership, release dry run, deterministic build test_distribution, remote required checks
Portability and hygiene, C-27 through C-34 invocation, cache bounds, parser, conflict provenance test_round14, test_mind, test_provenance
Roadmap A1 through A14 protocol, bulk ingest, sidecar, tiers, merge, doctor, policy, scoreboard domain test modules and all eight raw results
Auto-first contract policy capture, scheduler lease, growth receipts, five-year lifecycle test_policy, test_scheduler, autonomy result

Remote Completion Gate

Local evidence is necessary but not sufficient. Completion also requires:

  • all nine operating-system and Python matrix cells green on the pushed commit;
  • the release dry-run workflow green;
  • required review and conversation controls verified through repository state;
  • no open code-scanning, dependency, or secret-scanning finding attributable to the change;
  • the merged commit reachable from the default branch and a clean local worktree.