GoalProof

Elevator pitch

GoalProof turns an ambitious goal into a requirement-by-requirement evidence audit. Progress can be optimistic. Completion must be proven.

Inspiration

AI agents are remarkably good at making progress, but they often stop at a plausible-looking result. A passing narrow test, a configured deployment, or a confident summary can be mistaken for proof that the original goal is complete. The person who asked for the work is left to discover the missing requirement later.

GoalProof adds a missing layer between execution and completion: a visible evidence gate.

What it does

GoalProof takes a plain-language goal and decomposes it into checkable requirements. For each requirement, the user records the strongest current evidence and assigns one of three verdicts:

• Missing — no evidence exists.
• Weak — a claim, intention, proxy, or indirect signal exists.
• Proven — direct, reproducible evidence exists.

The completion gate stays blocked until every requirement is both marked proven and backed by a non-empty evidence record. The interface continuously shows evidence coverage, unresolved blockers, and a final completion verdict. The workspace saves automatically in the current browser. Users can export a Markdown report or a signed JSON snapshot and safely restore it later.

Every JSON snapshot includes the complete audit and a SHA-256 fingerprint calculated from canonical JSON. On import, GoalProof verifies the schema, requirement uniqueness, evidence states, recomputed audit, and fingerprint before changing the workspace. A modified or damaged audit is rejected.

Everything runs locally in the browser. There is no account, upload, hidden model call, or API key.

How we built it

GoalProof was built with GPT-5.6 and Codex during OpenAI Build Week.

The core logic was developed test-first:

1. A requirement extractor turns numbered, bulleted, and sentence-form goals into a deduplicated checklist.
2. An audit engine counts only evidence explicitly marked proven with a non-empty record.
3. A completion gate returns complete only at 100% coverage.
4. A report generator exports the full evidence chain as Markdown.
5. A canonical serializer and Web Crypto produce tamper-evident audit snapshots.
6. A strict verifier rejects unsupported, inconsistent, or modified imports.

The product interface is a responsive React experience built on a Next.js-compatible vinext stack and deployed with OpenAI Sites. The visual system uses strong editorial typography and explicit status colors to make the difference between confidence and proof immediately legible.

How we used Codex and GPT-5.6

Codex researched current, eligible earning opportunities and verified the official Build Week rules before any implementation. GPT-5.6 then helped select a general-purpose problem that did not depend on private professional context.

During implementation, Codex:

• translated the idea into a minimal product architecture;
• wrote failing tests before the audit engine;
• implemented the smallest passing logic and reran the tests;
• added canonical snapshots, SHA-256 integrity verification, and safe local restoration;
• created the full responsive product experience and accessibility states;
• generated and verified an original social preview asset;
• built, packaged, deployed, and prepared the submission materials;
• audited the final output against the competition requirements.

The project thread where this work was completed is the primary evidence of the Codex collaboration.

Challenges

The hardest product decision was defining what counts as evidence without turning GoalProof into another opaque scoring model. The solution is deliberately explicit: users choose the verdict, but the software enforces a hard invariant. A proven label without an evidence record does not count.

Another challenge was making the project useful without runtime credentials. We chose a local-first design so every judge can test the complete workflow immediately and the user's goal data never leaves the browser.

Accomplishments

• A working, coherent product rather than a static concept.
• An all-or-nothing completion invariant backed by automated tests.
• A keyless, account-free demo that judges can use immediately.
• Human-readable Markdown and machine-readable signed evidence formats.
• Browser-only persistence without sending goal data to a server.
• Twelve automated checks covering core logic, tamper detection, and the rendered product surface.

What we learned

The most important lesson is that agent reliability is not only about better generation. It is also about better stopping conditions. A system that can explain why it is not done is often more trustworthy than one that produces a polished final answer.

What's next

The next version will accept evidence adapters for CI, deployments, analytics, and task systems; preserve a local history of signed snapshots; and expose the completion gate as a reusable agent tool. The local-first core will remain available without an account.

Track

Apps for Your Life

Built with

GPT-5.6, Codex, React, TypeScript, Web Crypto, vinext, OpenAI Sites, Node.js test runner

Built With

• openai
• web