Content
# Proof Mode
Phase 1 foundation/ledger kernel, Phase 2 broker enforcement and the locked
five-tool STDIO MCP surface, plus the Phase 3 compiler and orchestration
integration.
The normative backend source is [Proof Mode Backend Schema Implementation Specification v1.0](docs/Proof_Mode_Backend_Schema_Implementation_Specification_v1.0.md). Earlier backend drafts are historical context only.
## Development
```powershell
uv sync --all-packages
uv run pytest
uv run pytest -q tests/test_phase2_broker_mcp.py
uv run proof-mode-db migrate --database .proof-mode/proof-mode.db --artifacts .proof-mode
uv run proof-mode-db verify --database .proof-mode/proof-mode.db --artifacts .proof-mode
```
The demo reset command is deliberately gated:
```powershell
uv run proof-mode-db demo-reset --data-dir .proof-mode-demo --confirm RESET
```
Phase 1 and Phase 2 acceptance are complete. Phase 3 now persists every
structured model interaction, enforces bounded compilation and exact-hash
confirmation, and records Codex run provenance. P0 trust-separation hardening
is complete: exact attempt records, L3 separate-process observation, exact
filesystem target binding, portable bundles, an offline verifier, and
tamper-invalidating verdict support are covered by the suite. The current
local regression suite is 80 passed and 3 skipped (the skips are the opt-in
Codex and live GPT-5.6 provider gates).
The P0 REST deployment profile is single-writer: run one API process with one
Uvicorn worker. Compile and confirmation mutations use an application-level
lock; multiple API processes require an external distributed writer lock.
The authenticated real-Codex and real GPT-5.6 compilation gates remain
intentionally opt-in because they require provider capacity and saved Codex
authentication:
The Phase 3 implementation is in `packages/mission_compiler` and
`packages/command_broker`: strict structured contract models, trusted
repository-binding checks, bounded repair, ambiguity refusal, canonical
contract hashing, durable interaction records, exact confirmation, and
Codex JSONL provenance.
```powershell
$env:PROOF_MODE_RUN_CODEX_E2E = "1"
uv run pytest -q tests/test_phase2_codex_runtime.py -s
```
The test preserves JSONL/stdout/stderr and exit-code evidence in
`.proof-mode/codex-runtime-evidence.json`. Provider capacity is reported as
`PROVIDER_UNAVAILABLE`; MCP startup or broker failures remain implementation
failures. The authenticated Codex/MCP gate is preserved as passed evidence;
the live GPT-5.6 compile remains `PENDING_PROVIDER_QUOTA`. P0 evidence is
recorded under `.proof-mode/p0-hardening/`.
Connection Info
You Might Also Like
ai-native-pm-os
The exhaustive guide to mastering Claude for Product Managers. Build your...
Train-in-Silence
The first Task-Aware MCP server and automated VRAM calculator for LLM...
stacklit
108,000 lines of code. 4,000 tokens of index. One command makes any repo...
AppClaw
AI-powered mobile automation agent — describe what you want in plain...
pdf-mcp
Production-ready MCP server for PDF processing with intelligent caching....
kotadb
Local-only code intelligence API for AI developer workflows (Bun +...