Proof

The harness is tested the way it governs: an AI builds real software under the rules, and independent checks verify the result. Below is how the testing works and what each trial proved.

How the testing works

Every trial runs the same pipeline. The principle throughout: no claim is accepted from the party that made it.

Requirement extraction

The owner brief is broken into numbered requirements (REQ-01, REQ-02, …), each with a verbatim quote from the brief. The requirement list the AI works against isn't one it wrote itself.

Preflight review

Before any code, an independent reviewer checks the AI's plan: every step needs a single verifiable outcome, pipeline steps must connect, and nothing required may be missing a step. Planning gets its own bounded budget — it never borrows from the build budget.

Build under the rules

The AI builds under versioned harness operating rules: bounded repair rounds, frozen acceptance checks, hash-chained state. Every trial records which rules version governed it.

Behavior checks on real state

The candidate is executed against frozen acceptance checks — real commands against real state (databases, files, process kills), asserting on complete persisted state and user-facing output. A check that passes without exercising the behavior doesn't count.

Independent verification

A fresh session that never saw the build re-verifies it: every file hash is rechecked, and a random lottery selects checks to re-run from scratch. The builder's own 'it works' is never accepted as evidence.

Receipted documentation

Per-check journals, command logs, file hashes, reviewer transcripts — everything is receipted so the verdict can be reproduced from the records alone.

What the trials prove

Trials run against a fictional multi-tenant booking spec — hard enough to be meaningful (hold expiry, idempotent replay, crash recovery, tenant isolation), fictional so no production data is ever involved.

Proves the gates are real

QDHY trial — INCOMPLETE_EXHAUSTED

The old protocol demanded a convergent PASS from the plan reviewer. The reviewer never converged, the run exhausted its budget — and the harness reported the failure instead of passing it anyway. Release stayed blocked through every shortfall. A harness that can't say no is decoration; this one said no, on the record, and the failure designed the two-stage protocol.

Proves the protocol completes builds

Two-stage trial — COMPLETE, 22/22 checks

Under the two-stage protocol, the same spec ran to completion: all 22 behavior checks passed on executed code, including two real SIGKILL recovery paths. Planning stopped at its budget with a written record instead of spiraling. Prompt in, verified code out, fully documented.

Proves verification survives the environment

Verification survival — COMPLETE

The independent reviewer's sandbox was measured killing commands at exactly 60.0 seconds. The runner was rebuilt around that fact: 48-second per-check budget, every check journaled to disk, resume support. Result: 22/22 controller checks, 130/130 file hashes independently verified, 5/5 lottery-selected spot checks re-run and passed. A killed check is recorded as evidence — never silently lost, never converted into a pass.

Proves it reproduces

v2 proveout — COMPLETE

The two-stage protocol promoted into standing rules (v2) and re-run against the same spec and fixtures: 22/22 checks, 130/130 hashes, 5/5 spot checks, all eight hard gates PASS, zero new repair rounds burned. Same outcome under a new rules version — the result comes from the protocol, not from one trial's luck.

Check it yourself

The Brain Pad runs the deterministic checks in your browser — no server, no model, no judgment calls:

Requirement map check — every requirement maps to implementation, a workflow test, and evidence; boilerplate maps rejected.
State record check — machine-readable checkpoint state with round budgets, gate ledger, and failure inventory.
Handoff content check — the resumption packet carries the literal brief, packet, frozen checks, evidence, and counters.
Bundle-vs-handoff check — every contracted value must appear verbatim in the handoff; even a one-digit mutation fails.
SaaS sidecars — idempotency, entitlement gating, fail-closed behavior, and rate limiting checklists with explicit not-applicable marks.

Open the Brain Pad or run an independent review — paste any AI's output and watch the checks work.