Skip to content

VERIFIER.md

Latest commit

 

History

History
66 lines (54 loc) · 3.04 KB

VERIFIER.md

File metadata and controls

66 lines (54 loc) · 3.04 KB
Raw
tags
core
verification
runtime
project ABSOLUTE_ZERO
status active
confidence high
date 2026-07-11
summary Verification Engine spec — eleven checks over every modification, confidence score, gated verdict before completion.

ABSOLUTE ZERO — Verification Engine

Verifies every modification before completion. Nothing is DONE until the verifier says so — a gate failure exits nonzero and the /task VERIFY state cannot pass. Implementation: scripts/verifier.py (stdlib; python analyzed with ast, not regex guesswork).

1. API

python scripts/verifier.py check              git-changed files (default)
python scripts/verifier.py check <files...>   explicit set
python scripts/verifier.py check --all        whole-vault sweep
python scripts/verifier.py check --json       machine report
python scripts/verifier.py --selftest

Reports persist to 90_META/verify/<timestamp>.json. Exit code: 0 unless verdict is FAIL.

2. The eleven checks

check gate what it does
architecture ✔ files inside vault taxonomy; python only in scripts/; root docs registered in indexer ROOT_DOCS
imports ✔ ast: every import stdlib or local (stdlib-only law = FAIL); unused imports (warn)
dependencies ✔ circular imports across scripts/ (graphlib); dead wikilinks in changed notes (code blocks excluded)
naming ast: snake_case functions (dunders ok), PascalCase classes
complexity ast: >60 statements or nesting >4 per function; >500-line files
regression reverse import graph: who imports the changed module — their selftests are pulled into the test run
security ✔ hardcoded credentials, eval/exec, shell=True, os.system, pickle, absolute user paths; # verifier:ignore suppresses a line
performance IO/subprocess/compile calls inside loops (warn — sometimes inherent, flagged honestly)
formatting lines >100, trailing whitespace, tabs, missing final newline
documentation py: module docstring + usage example; md: frontmatter with mandatory summary ≤25 tokens, tags, date
tests ✔ every vault script has --selftest (OS law) and the selftests of changed + dependent scripts are executed

3. Confidence and verdict

confidence = 100 * (1 - min(0.03*warns, 0.30) - 0.15*fails)   floor 0
verdict    = FAIL                 any fail in a gate category (✔ above)
           = PASS-WITH-WARNINGS   warns only
           = PASS                 clean

Warns cap at −30: style noise cannot zero the score. Fails are uncapped and gate categories block completion — failure is loud by design (P1).

4. Integration

  • /task VERIFY stage: run verifier.py check first, then the trace's intent checklist. FAIL verdict → retry EXECUTE or close fail; never SUMMARIZE over a failing verifier.
  • Thresholds, gates, penalties, security patterns are constants at the top of scripts/verifier.py — tune there.
  • First vault-wide sweep (2026-07-11): caught two real violations — query.py and review.py predated the selftest law; both fixed.