# Card Cauldron TCG

- **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:** [Sabastian Caudill](https://cerebralvalley.ai/u/buckeyebinder)
- **GitHub:** https://github.com/caudillsabastian6-png/Fake-check
- **Demo video:** https://www.youtube.com/watch?v=1vAKOJ_Vju4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1958

Real or Replica — Fake Check
Live demo: https://cardcauldrontcg.com/demo
Model weights (Basil reader, Muse Glimmer QLoRA): https://huggingface.co/CardCauldronTCG/basil-glimmer-qlora

### Honesty & proof status

Real or Replica's card checker is governed by a rule we set before we had results to protect: **no verdict ships until it passes a blind proof test we pre-registered** — false-flags ≤ 5%, counterfeit catch ≥ 80%, decisive coverage ≥ 80%, zero forbidden outputs, judged on physical cards alone. That test has not been passed yet, so the demo does not pretend otherwise: every check in the demo is labelled **SAMPLE**, and the production checker answers **Inconclusive** until the evidence earns more. We would rather show judges an honest "not proven yet" than a confident demo that grades its own homework.

One manufactured lane *has* passed its own proof test, under its own name and its own fences, and it is the only verdict we have switched on anywhere in the product:

> **"Our printed pair-sheet lane passed its blind proof test — 399 of 400 printed fakes caught, 3 false alarms in 120 — and that lane, under its own name, is the only verdict we have switched on; every other check still honestly says SAMPLE or Inconclusive."**

Where the proof program actually stands, in numbers we have measured and filed:

- Our first combined test (**Run 4**) failed three of the four bars — false-flags 15.66%, catch 37.71%, coverage 53.03% — because manufactured fakes made up most of the training pool and the model learned our print paths instead of real counterfeits (its read on physical fakes *fell* to 14.29%). That failure redesigned the system: physical cards and each manufactured family now sit in **separate lanes**, separately fitted and reported, and the physical core is frozen and judged alone.
- The frozen physical core, measured exactly on its historical test side, reads false-flags 3.33% and coverage 94.17% — but catch **51.72%**, well short of the 80% bar. A tune-side re-test on our newly organized archive (8,125 of the owner's own scans and photos, labelled and designation-screened) did not move that headline: the best honest combination reads 69.0% catch at 10.8% false-flags. The gap is evidence, not thresholds: most of our 57 known physical counterfeits imitate one modern set whose genuine counterparts are still being catalogued out of thousands of scanned pages, and a statistically credible blind test needs more fresh counterfeits (≥ ~62) than currently exist in our house (57, all previously measured).
- The final pre-registered attempt at that physical test — **Run 5**, run today, Wed Nov 4, 2026, as a disclosed **legacy-protocol attempt** (its test set is our house collection, handled and measured during development, so it is not and is never called a blind test) — **stopped at its own freeze audit, before any card was scored**. Thirteen of fourteen audit checks passed; the one failure was a frozen-artefact hash: a governance section (a threshold-margin rule) had been appended to the protocol's design document after its hash was pinned, under an owner's order and on the record — touching no threshold, no item, and no measurement file, eight of which still match their pins exactly. The gate does not grade causes, and we did not repair it after the fact: no labels were sealed, no measurement was run, and no pass of any kind is claimed. The signed attempt report is in our public record. A future measurement of this population runs under a new pre-registration — and a switched-on physical verdict still requires a blind pass on fresh evidence (≥ 62 fresh fakes + ≥ 200 fresh genuine) plus a separate authorization, neither of which exists today.
- The manufactured lanes, reported separately, each on its own denominators: the **printed pair-sheet family** — front+back fusion, one frozen sitting, executed exactly once: catch **399/400 = 99.75%** (95% CI 98.6–100.0%), false-flags **3/120 = 2.50%**, decisive coverage 99.75% of fakes and 92.5% of genuine, forbidden outputs **0** — measuring the one school print path and our house scan path against the known-truth genuine reference, not wild counterfeits or other printers. That lane — the Printed Pair-Sheet Lane — was then switched on for scanner-pair submissions only, under its own name, after a final fresh-eyes live read by our verifier. Its first live traffic, this morning, measured and recorded exactly as it landed: a genuine **Flying Pikachu V** read *Likely Genuine*; a genuine **Golisopod GX** read *Likely Genuine*; and the printed fake of that same Pikachu, from the sheet filmed for this video, read **Inconclusive** — its front and back sat inside the band, and no threshold was touched and no read re-rolled to make it a better story. The crew-manufactured digital family is a weak lane whose full measurements sit in our internal readiness record, because our final verifier restricts that lane's figures to the internal record; disclosed here as weak rather than hidden, because hiding it would be the kind of bookkeeping this project exists to refuse. A third family — fake sheets we fabricated ourselves on cardstock and paper, with printed backs (the first fakes in our corpus that *have* backs) — was printed Nov 2, frozen under its own ingestion protocol, and is held below our verifier's fit minimum; it carries no verdict weight.
- A new labeling channel — reading a card's printed text, fonts, and glyph-scale print process against the record for its exact printing — has completed its tune-side controls: zero decisive false-fires on genuine cards after audit fixes, low coverage, and a recommendation of a narrow corroborating seat only. It was **not** admitted to the Run 5 attempt and has earned no seat in any verdict; it is described here as what it is: in testing.
- The identity read inside Check a Card is supplied by a reader adapter we trained on **Muse Glimmer**, Meta's model: it reads a card's name, printed number, and set — identity only. It never votes on authenticity; we also tested it as a judge, it failed outright (it endorsed counterfeits as genuine far more often than genuine cards), and that failure is disclosed in our record alongside the exam numbers. Its weights are published at **huggingface.co/CardCauldronTCG/basil-glimmer-qlora**. On submission day the reader's hosted endpoint is dark — its Modal workspace was disabled when its credits ran out — so the read is described here in writing rather than shown live in the video; the video shows only what was actually working when it was filmed. The demo itself is live at build **#27.15**, and the project's code and records are public at **github.com/caudillsabastian6-png/Fake-check**.

When a version of this checker tells a collector their physical card is *Likely Genuine* or *Likely Counterfeit*, that sentence will have been earned on cards the model had never seen, under bars we published before we knew the result. Until then: SAMPLE means sample.

## More from Meta Global AI Developer Hackathon

- [Zukki](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1954)
- [bjirka](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1956)
- [WORLDTRACE](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1957)
- [Doom_Slayer](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1959)
- [Ras Ali Labs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1960)
- [metacog](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1961)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1958. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt