prove-it-ai-gate 0.3.2

A lightweight local CLI acceptance gate for AI-generated engineering work

prove-it-ai-gate

Don't trust agent output. Accept evidence.

A lightweight local CLI acceptance gate for AI-generated engineering work. Checks agent transcripts, evidence folders, workspace state, and validation outputs before accepting AI-generated work.

Warning: This tool checks evidence quality, not absolute correctness. It is not a substitute for human review on high-risk changes.

Project Status

prove-it-ai-gate is currently an early beta.

Stable core:

• transcript-check
• workspace-check
• evidence-check
• accept

Experimental:

• reuse-scan
• capture
• wiki upload
• evidence cleanup/archive

Use the stable core for local dogfooding. Treat experimental features as evolving APIs.

Installation

pip install prove-it-ai-gate

Or from source:

git clone https://github.com/hrishikesh-thakre/prove-it-ai-gate.git
cd prove-it-ai-gate
pip install .

Requires Python 3.9+.

Quick Start

ai-gate init
ai-gate transcript-check --transcript transcript.jsonl
ai-gate workspace-check --repo ./my-project
ai-gate evidence-check --evidence ./evidence --task-type audit

# Full acceptance gate
ai-gate accept \
  --repo ./my-project \
  --evidence ./evidence \
  --transcript transcript.jsonl \
  --task-type audit

# Scan a knowledge wiki before starting work (experimental)
ai-gate reuse-scan --brief brief.md --local-wiki ./wiki-exports

# Capture reusable learnings after acceptance (experimental)
ai-gate capture --evidence ./evidence --output ./learnings

Decision States

Decision	Meaning
ACCEPT	Required evidence exists and checks passed
ACCEPT_WITH_CONDITIONS	Mostly acceptable, but follow-ups remain
REJECT	Evidence proves a failure or unacceptable risk
BLOCKED	Required evidence is missing or invalid

Built-in Checks

Check	What it catches
transcript_truncation_check	Agent used truncated tool output without replacement evidence
workspace_hygiene_check	Agent modified target repo during read-only task, scratch files in repo
evidence_folder_schema_check	Missing, empty, or placeholder evidence files
confidence_claim_check	Unsupported high-confidence claims, heuristic overclaim
scope_completeness_check	Bare "Complete" claims without scoped corpus definition
heuristic_extraction_check	Filename/keyword matching claimed as full coverage
closeout_transcript_crosscheck	Closeout claims "no changes" but transcript shows write events
evidence_inventory_integrity	Inflated inventory counts vs transcript tool outputs
evidence_validation_integrity	Test/lint output mismatch with closeout claims
evidence_depth	Shallow/minimal evidence stubs
git_status_authenticity	Fabricated git status not matching transcript

Dogfood

Run automated dogfood cases to verify the gate catches known failure patterns:

python scripts/run_dogfood.py

See dogfood/dogfood_report.md for results and examples/dogfood/ for test case fixtures.

Policy Packs

Three built-in policies in prove_it_ai_gate/policies/:

Policy	Use Case
read_only_audit.yml	Audits, reviews, listings, evaluations
code_change.yml	Code modifications with tests and validation
corpus_audit.yml	Full-corpus scans with extraction artifacts

Evidence Folder Contract

Every agent task should produce:

evidence/
  brief.md, plan.md, commands.jsonl, changed_files.txt
  closeout.md, risks.md
  validation/test_output.txt, lint_output.txt, typecheck_output.txt, audit_guard.txt
  artifacts/inventory.json, extracted_findings.json
  workspace/git_status_before.txt, git_status_after.txt, diff_summary.patch
  transcript.jsonl

See docs/evidence-folder-contract.md for requirements by task type.

Docs

• Evidence Folder Contract
• Acceptance Policy Model
• Transcript Schema
• Workspace Hygiene
• Confidence Rules
• Prove It Without Me
• Release Readiness

License

MIT — see LICENSE.