My provisional favorite is OutcomeLoop. The implementation makes the most important anti-agent-theater move in the visible field: an agent message cannot declare victory. A protected external verifier must pass, and the controller seals a signed receipt. But the most complete public submission packet belongs to Incident Commander AI, while Referral Copilot has the cleanest evidence-grounded review experience. Those are three different ways to win.
The judge contract changes the strategy
The challenge opened July 13 and closes July 21 at 5:00 PM Pacific. It offers four
tracks—Apps for Life, Work & Productivity, Developer Tools, and Education—and
scores technical implementation, design and user experience, impact, and idea quality
equally. The tie-break follows that order, so technical implementation is not merely
one quarter of the score; it is the first separation when totals match. The required
package includes a working project, a public YouTube demo under three minutes, a
repository, and a Codex /feedback session ID.
Existing products are allowed, but only meaningful work built after the event began is evaluated. That makes dated delta evidence, commit history, setup clarity, sample data, and a judgeable no-surprises run path part of the product—not administrative garnish. OpenAI's own advice is equally useful: start from the problem, use Codex and GPT‑5.6 thoughtfully, and record the demo as the build comes alive.
Technical implementation
Non-trivial, working, skillful Codex and GPT‑5.6 use. First tie-break.
Design & UX
A coherent runnable product, not disconnected proof-of-concept screens.
Impact
A real audience and problem, demonstrated rather than asserted.
Idea quality
A creative, memorable mechanism—not another generic assistant.
The ten strongest public signals
Scores are editorial judgments out of 100, using the four official criteria at 25 points each. They measure public evidence available during this scan. A strong README with no runnable proof loses points; a beautiful demo with weak model necessity loses points; a broad claim with a narrow evaluation loses points. Each project gets both its winning case and the risk most likely to cost it.
Developer Tools · provisional leader
OutcomeLoop
“Codex can stop. The outcome cannot.”
A deterministic controller resumes one GPT‑5.6 Codex session until a protected external verifier passes. Least-privilege agent and verifier sandboxes, protected-path fingerprints, preflight verification, an Ed25519 identity, and a signed receipt turn completion into evidence instead of a message.
- Winning edge
- The mechanism is novel, technically deep, and demonstrated across a four-turn live handshake.
- Risk
- Its user wedge and secure verifier authoring burden are narrower than the universal slogan.
Work & Productivity · proof-bundle leader
Incident Commander AI
The public packet is formidable: typed workflow state, two recorded human approvals, a credential-free deterministic judge demo, 185 backend tests, 22 Chromium scenarios, eight safety evaluations, mutation tests, secret scans, release images, and explicit simulated-versus-live provenance.
- Winning edge
- Judges can verify the complete story without trusting a pitch or provisioning integrations.
- Risk
- Incident response is crowded; the engineering is more differentiated than the core idea.
Work & Productivity · review-UX leader
Referral Copilot
A synthetic gastroenterology referral becomes structured facts, an urgency proposal, and exact source phrase → rationale → synthetic rule traceability. The clinician can approve, edit, or reject; identifier and prompt-injection guards sit before the model; exact quotes are validated before display.
- Winning edge
- The live demo makes evidence and human authority tangible in one polished screen.
- Risk
- The evaluation is intentionally small, and demo fallback must not blur the live GPT‑5.6 moment.
Apps for Life · public-safety grounding
SISMICA GPT‑5.6
The browser sends only a canonical seismic event ID. The backend loads database facts and associated sources, hashes the input, calls GPT‑5.6 with strict JSON and store: false, validates again, caches by event version, and exposes response ID, source count, tokens, latency, and hash.
- Winning edge
- Excellent evidence grounding and unusually responsible separation from prediction and official alerts.
- Risk
- The public packet still marked a real API response pending and granted no software license at scan time.
Education · reproducible science
AstroData AI
Local scientific libraries compute the evidence before GPT‑5.6 explains a controlled metric summary. Raw FITS files and full time series stay out of the prompt. The blind synthetic fixture hides a 2.75-day transit-like signal, and the product visibly resolves the half-period harmonic without pretending it discovered a planet.
- Winning edge
- A beautiful three-minute scientific reveal with an honest claim boundary.
- Risk
- The wide FITS/time-series feature set can compete with the one memorable story.
Developer Tools · underserved language wedge
AntiShadow AI
A Thai-first local egress gateway keeps sensitive prompts away from GPT‑5.6, passes short clean prompts locally, and sends only long clean prompts for structured semantic risk review. The proof console shows exactly what the model saw—or NOT SENT.
- Winning edge
- Local privacy architecture plus a culturally specific enterprise gap.
- Risk
- The detector benchmark is small, zero entities is not a guarantee, and its license restricted commercial use.
Education · deterministic evidence state
Misconception Studio
GPT‑5.6 extracts literal spans in two bounded rounds; deterministic code alone advances evidence state after quote offsets, marker ownership, and completeness checks. The 24 fictional responses, exact denominators, TTL cleanup, reset, and public-claim scan show rare epistemic discipline.
- Winning edge
- The model/deterministic boundary is exemplary and the “insufficient evidence → probe → changed evidence” arc is judgeable.
- Risk
- Public deployment and independent content review remained incomplete.
Apps for Life · clearest human need
Okusuri Toban
An Android app answers one family question: has someone already given this pet's dose? It uses append-only records, explicit cancellation events, double-record warnings, a one-screen handoff, and an account-free judge APK—with no dosage advice.
- Winning edge
- Outstanding focus, human stakes, restraint, and judge accessibility.
- Risk
- GPT‑5.6 is more visible as build collaborator than essential runtime intelligence.
Apps for Life · coherent consumer workflow
Airline Said No
Airline refusal documents become facts, interpretation, recommendation, and an editable reply. Structured outputs, request-scoped document handling, store: false, keyboard support, reduced motion, fictional samples, 78 tests, and a polished short demo make it easy to understand.
- Winning edge
- A complete problem-to-action arc with excellent consumer clarity.
- Risk
- No hosted judge path and a comparatively familiar document-explanation pattern.
Developer Tools · research originality
Sieve
A causal faithfulness auditor changes one stated rationale claim, constraint, or hypothesis and measures whether the resulting patch and tests move. Five recorded baselines and 15 interventions make the offline report reproducible without pretending to reveal private reasoning.
- Winning edge
- It asks a genuinely new evaluation question and frames the limitations correctly.
- Risk
- The current UX is a research instrument, and the illustrative task set limits broad impact claims.
The winning shape is remarkably consistent
Make the model's authority smaller than its intelligence.
The best projects let GPT‑5.6 classify, explain, extract, or propose. Deterministic code owns state transitions, permissions, source validity, and external effects.
Turn the dangerous failure into the demo.
OutcomeLoop fails the verifier. Incident Commander blocks remediation. AstroData exposes a harmonic. Misconception Studio begins with insufficient evidence. Failure is where trust becomes visible.
Ship a judge path that needs no faith.
Frozen synthetic fixtures, one-command setup, account-free builds, deterministic rehearsals, and explicit live-mode receipts reduce evaluation friction without faking the model.
Leave a receipt at the exact claim boundary.
Source IDs, input hashes, response IDs, latency, storage flags, approval events, signed evidence, and exportable reports make “it worked” inspectable.
The gap: continuity after the model interaction breaks
Healthcare entries are already strong. Competing head-on with referral triage would be foolish, and “AI visit summary” is both crowded and too easy to imitate. The unclaimed territory is the moment after capture: a voice intake drops, a note remains unsigned, an attachment is referenced but absent, or the originating clinician leaves shift. The product problem is not generating more prose. It is keeping patient context, clinician intent, ownership, and one concrete next safe workflow step together.
That is the continuity invariant behind our sibling doctor-first care operating system: captured consults, closed handoffs, and a living patient timeline. For Build Week we isolated the invariant into a new, public, synthetic-only surface with no production data dependency and no care-action authority.
What we cooked
Continuity Relay
The visit broke. The thread didn’t.
A judge chooses a frozen synthetic case, injects a handoff failure, and runs a live
GPT‑5.6 continuity pass. The browser sends only two enums. The server owns the facts,
hashes the grounded input, calls the Responses API with strict structured output and
store: false, validates every cited source ID, and requires clinician review.
- 1BreakInject interrupted ownership.
- 2GroundLoad frozen server-side facts.
- 3RelayGPT‑5.6 proposes continuity.
- 4ValidateReject unknown evidence.
- 5ReviewA clinician remains the gate.
We designed it backward from disqualification
Why this can place—and what would still beat it
The bet scores well because the problem is legible in seconds, the break is interactive, GPT‑5.6 is necessary but bounded, the new-work delta is clean, and the receipt is the interface. Work & Productivity is a better track than Apps for Life because the buyer and user are clinic teams coordinating work, not consumers seeking medical answers.
It still loses if the live model path is flaky, the three-minute video does not show the network boundary, the output feels like an ordinary summary, or we imply care validation. It also loses to a product with representative user evidence. This version is a rigorous interaction and safety prototype, not proof that the workflow improves care. The next honest evidence is repeated blinded fixture evaluation followed by clinician usability review—not a larger prompt.
Appendix · references and methodology
How this ranking was made—and how to challenge it.
A. Collection window and source order
The scan ran July 17, 2026 Pacific time. Official OpenAI and Devpost pages defined eligibility, timing, required artifacts, categories, and the scoring rubric. The public field came from GitHub repository searches for “OpenAI Build Week” and “openai-build-week” with attention to repositories created or materially updated after July 13. Each shortlisted README was read directly; linked live demos, releases, videos, CI, judge guides, validation reports, and Devpost receipts were included when public. Search-result snippets were not treated as product evidence.
B. Inclusion and exclusion
- Included: explicit Build Week 2026 identification, a judgeable product description, credible GPT‑5.6/Codex use, and enough public implementation evidence to score at least three criteria.
- Excluded: empty shells, name-only repositories, projects whose Build Week delta could not be distinguished, and projects without enough public evidence to evaluate.
- Not inferred: private submissions, unlinked deployment state, unpublished tests, API calls claimed without receipts, or adoption and efficacy not measured in the repository.
C. Scoring formula
Each official criterion received 0–25 points. Technical implementation rewarded non-trivial model/Codex use, architecture, reproducibility, safety, tests, and live receipts. Design rewarded a coherent end-to-end flow, accessibility, clear state, low judge friction, and honest failure UX. Impact rewarded specificity of problem, audience, stakes, and demonstrated usefulness. Idea quality rewarded a memorable mechanism and distance from generic assistant patterns. Scores were not adjusted for stars, participant popularity, team size, or prose volume.
D. Limitations
- The official project gallery was unpublished, so recall is unknowable and the ranking cannot represent the full field.
- Public repositories can change after this snapshot; scores are intentionally date-stamped.
- We did not clone and execute every project or independently reproduce every performance/test claim.
- Repository quality and documentation may over-represent engineering-heavy entrants relative to video-first submissions.
- Continuity Relay is our own project; its analysis is disclosed separately and it is not inserted into the ranked public field.
E. Primary rule and model references
- OpenAI Build Week — event timeline, judges, submission package, and evaluation framing.
- OpenAI Build Week on Devpost — tracks, prizes, required artifacts, criteria, and deadline.
- Official rules — eligibility, Build Week delta, judging stages, repository/run access, and tie-breaks.
- Official resources — challenge guidance and current credit availability.
- Project gallery — visibility boundary at scan time.
- OpenAI reasoning guide — GPT‑5.6 starting guidance and Responses API recommendations.
F. Candidate evidence
- OutcomeLoop repository
- Incident Commander AI repository
- Referral Copilot repository
- SISMICA GPT‑5.6 repository
- AstroData AI repository
- AntiShadow AI repository
- Misconception Studio repository
- Okusuri Toban repository
- Airline Said No repository
- Sieve repository
G. Reproducible packet
The repository carries the machine-readable scores, collection boundary, submission
copy, failure analysis, and under-three-minute demo script under
memory-bank/artifacts/openai-build-week-winning-contender-2026-07-17/.
Change a score if the evidence changes; do not move the caveat.