FaultLine

Freeze one reviewed executable question. Replay it over recorded Git or Codex turn states. Prove only what the executions support — without trusting model intent.

Track: Developer Tools (CI / DevOps / agentic debugging evidence)

Break → Find → Prove → Fix → Prevent

Proves	Never claims
Earliest stable PASS→FAIL under a frozen witness	Model intent / unique semantic root cause
Portable offline-verifiable packages (COMMIT_PROOF on Git path)	Private Codex interception
Structured PREDICATE_* outcomes (compile ≠ FAIL)	That one Docker image fits every lockfile era

Start here (about 2 minutes)

Judging pin only: v0.1.10-buildweek
This is the only release intended for judging. Do not evaluate from main, npm, or older tags (v0.1.9-buildweek lacks EXT-01; v0.1.4 is RETRACTED).

Option A — Codespaces (fastest)

Then run:

pnpm fl judge-proof

Option B — clone (six commands)

git clone https://github.com/Mizore66/faultline.git
cd faultline
git checkout v0.1.10-buildweek
corepack enable
pnpm install --frozen-lockfile
pnpm fl judge-proof

pnpm install builds the CLI (prepare → pnpm build). Use pnpm fl:dev only for local TypeScript iteration.

What you should see

• A verified COMMIT_PROOF sample package (root sha256:f85c446d…)
• Not the historical self-incident unless you open that package separately
• No Docker and no API key required for this path

Headless: FAULTLINE_NO_BROWSER=1 pnpm fl judge-proof --export-only
Static preview: docs/self-incident-proof-preview.html
Offline verify the sample:

pnpm fl verify docs/samples/self-incident-commit-proof --expect-root sha256:f85c446dfd5ab92222b10a314e79209a8a7dc10ee69af9d2deaa04aceafeb7d9

More install detail: docs/demo-zero-install.md · persona guide: docs/personas/judge.md

Which path am I looking at?

Command	Grade	Use it for
fl investigate git / judge-proof	COMMIT_PROOF	Mature portable proof — lead here
fl investigate turns	EXPERIMENTAL_TURN	Codex-native First Bad Turn — experimental until promotion criteria
fl judge-demo	Sample UI	Evidence-model sandbox only — not a real incident

EXPERIMENTAL_TURN stays experimental until the scoreboard in src/turn-proof-promotion.ts is met (multi-OS soak still open, issue #152). Do not invent soak stats.

Optional live Docker (only if pnpm fl doctor --proof-ready says READY):

pnpm fl demo full

If doctor is not READY, stay on judge-proof / the static preview — do not chase live proof.

Evidence in this repo

Self-incident (published dogfood): Git regression at 97c3290 — docs/faultline-self-incident.md

External protocol validation (permissioned, completed): consented MUMBCS run — Turn 3 earliest stable PASS_TO_FAIL under EXPERIMENTAL_TURN, package root sha256:831885ed72814e3c2e68dd3366060f88ee94d1b0e533c4571246efd6957326b1.

• Summary: docs/publishable-summary-mumbcs.md
• Case study: docs/external-case-study-mumbcs.md

This is protocol / interoperability validation, not a claim that FaultLine caught a naturally occurring production bug, agent intent, unique root cause, or live-site impact. Byte-level re-verify of that turn package needs MUMBCS collaborator access (private repo); public readers use digests + consent + redacted narrative.

Codex + GPT-5.6

Role	What it does here
Codex (build)	Implementation, adversarial tests, hardening. Qualifying /feedback: 019f66bd-0ac1-78f3-8dc1-5968e4f2fa09
Codex (runtime)	Opt-in sidecar records public lifecycle hooks + dirty turn-tree snapshots
GPT-5.6	Blinded witness proposal + evidence-cited repair brief only — never PASS/FAIL

GPT-5.6 shapes without a key: docs/samples/gpt-5.6/. On camera with a key: fl witness propose --live.

Help is tiered: fl / fl quickstart for the short surface; fl advanced for the full command list.

Platform notes

• Node.js 22+, pnpm 10, Windows / macOS / Linux
• Docker required only for live proof (demo full / investigate), not for judge-proof
• Windows: use pnpm.cmd if ExecutionPolicy blocks pnpm. Do not insert -- between fl and the subcommand
• npm: published package may lag the pin — do not treat npm as the EXT-01 judging artifact (docs/continuous-release.md)
• CI evidence: See Actions on v0.1.10-buildweek / commit 89b8e23… (pin retargeted 2026-07-21 to current main, package 0.1.4). Prefer linking exact successful required jobs rather than an aggregate badge alone.

Docs by audience

If you are…	Start with
A judge / evaluator	docs/personas/judge.md, docs/concepts.md
A new user	docs/personas/new-user.md, docs/first-incident.md
Doing forensics / proof work	docs/personas/forensics.md, docs/proof-bundles.md
Checking security / chain	SECURITY.md, docs/security-chain.md, docs/security-model.md
Comparing to bisect / CI	docs/differentiation.md
Recording / submitting	docs/demo-partner-handoff.md, docs/build-week-submission-kit.md

Demo helpers: zero-install · rehearsal · Q&A · fallback · sidecar

Development

pnpm typecheck
pnpm test
pnpm build
pnpm coverage-matrix

On Windows, the same commands work via pnpm.cmd if ExecutionPolicy blocks pnpm.

pnpm coverage-matrix writes an adversarial scenario coverage matrix (spec/unit coverage — not an E2E benchmark).