# Contravious Battle

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Contravious Battle](https://cerebralvalley.ai/u/conwanzo)
- **GitHub:** https://github.com/ConWan30/qortime-ai
- **Demo video:** https://youtu.be/INGaC9IYsNM
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/80

Signed, hash-chained receipts for console head-to-head matches, checkable in any browser, with Meta's Muse Spark, SAM 3.1, Muse Wonder and Muse Voice Transcribe explaining, cross-checking and attacking the referee's result. The AI can only lower a result, never decide one. Anti-cheat and verified skill, not a coach.

Console head-to-head results still run on trust: when scores conflict, players trade screenshots and an admin makes a call. QorTime is video-game integrity verification (anti-cheat plus verified skill). QorCap records a match (video hash chain plus the controller log over USB), a deterministic referee checks it and signs a receipt, and anyone can re-check that receipt in the browser: VERIFIED, PARTIAL, SELF-SIGNED or FAIL.

This week I built the Meta layer on top, in a fresh repo (github.com/ConWan30/qortime-ai). The rule: AI explains, it never decides. Every model output is a separate signed sidecar bound to the receipt hash; it can only lower VERIFIED to PARTIAL, never raise a status or fail a match, and deleting or tampering with it never changes the verdict (tests check this).

Pre-existing (before kickoff Mon Oct 26 2026), not in the submitted repo: the receipt format and signing, the deterministic verifier and referee service, the transparency log, QorCap, the website and match pages, and match 170123 (VERIFIED Oct 23). qortime-ai pulls that code in only as a pinned dependency (qortime @ 0e87ec3).

Built this week (all in qortime-ai, file-by-file origin in PROVENANCE.md):
Signed plain-English explanation + deterministic guardrail | Muse Spark 1.3 | none (fail-closed) | live on both verified matches
Second, independent read of every locked score | Muse Spark 1.3 | lower only | 170123: 2 of 2 agree; dev bench 73/73 (not held out)
Show Your Hands play-start filter | Muse Spark 1.3 | remove-only, capped at PARTIAL | —
Evidence Lens: "what the referee saw" | SAM 3.1 + Spark captions | none (visualization) | scorebug held 359/359 and 360/360 frames
Challenge Echo: pause-menu highlight vs logged D-pad presses | SAM 3.1 | info only, lower only | 220105: 3 of 3 challenges consistent
Stick Echo: controlled player's movement vs logged left stick | SAM 3.1 | info only, lower only | 3 private practice games: the first fell short (lowest z 2.68); the next two passed tests registered in advance (3.11, 3.39)
Red-Team the Referee | Muse Wonder 1.0 (+ Spark cited explanations) | none (attacks the verifier) | outside attacker 25 attacks · 0 accepted; insider pass reported (12 of 25, labelled)
Commentary witness: broadcast calls vs locked scores | Muse Voice Transcribe 1.0 + Muse Spark 1.3 | info only, lower only | 170123 clips: 3 of 3 agree (dev check)
Agent verify API + MCP (read-only) | — | reports the verifier's result | live: 220105 verified: true

How this SAM 3.1 use differs from player-tracking or coaching apps: SAM 3.1 isn't measuring the player. It outputs no skill, speed, route or play-quality numbers. Stick Echo checks that the on-screen player's movement follows the controller's stick input in time, scored against swapped-log baselines. It is info only, can only lower VERIFIED to PARTIAL, and passed on two of three private practice games.

Second VERIFIED match, played during the event: 220105 (Oct 27, away 7 – home 21), with the end-menu fallback and the advisory flag `alignment_not_above_chance` shown on the page, not hidden.

Try it in 2 minutes: https://qortime.com/matches/220105/ → "Verify it yourself" (in-browser verifier plus a check of the referee's public log). Offline: README "For judges" (no API key needed; 397 pytest passed, 182 node checks).

Honest limits: two referee-verified matches is n = 2, not a measured accuracy; one player's own Madden NFL 27 matches on PS5, single-sided. QorTime does not show who held the controller and does not catch collusion. Device status is at most "TPM claimed (unverified)". Stick Echo is three practice games, not a measured accuracy.

## More from Meta Global AI Developer Hackathon

- [manymoats](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/77)
- [Babatunde](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/78)
- [2.0 tech](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/79)
- [Hearth](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/81)
- [Hokfam](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/82)
- [BARRED](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/83)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/80. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt