cred402/docs/near-hackathon
caelum0x 524f2ddea5 fix(near-hackathon): re-voice false "live on two other chains" claim in pitch demo video
Narration line 2 said Cred402 is "already live on two other chains, and
now native NEAR" — false. Live checks against
cred402-1.onrender.com/v1/x402/{algorand,qie}/status both return
configured:false. The deck was already corrected (4eb21fe); this patches
the video's narration to match.

New line 2: "Cred402 sells credit scores to AI agents over a pay per
call payment protocol, and now it's enforced natively on NEAR." Same
Aura-2 voice (aura-2-asteria-en — re-derived acoustically, since no
voice ID was recorded in the original build: synthesized the real line 1
text in all 5 cataloged voices and compared pitch/spectral-centroid
against the shipped audio; asteria matched an order of magnitude closer
than the next candidate). Same per-line Deepgram nova-3 QC round-trip.

Only the line 2 window (8.9s-17.766s) was touched — confirmed by frame
diff that this whole span is a single static held title card in the
shipped video, so the patch could re-render just that card + new
captions + new audio and splice it between the untouched first 8.9s and
untouched remainder (which includes line 3's unmoved 19.5s start).
Re-transcribed the full final audio with nova-3 to confirm lines 1 and
3-6 are unchanged.

Duration unchanged (70.6s), resolution/fps/codec unchanged (1920x1080,
30fps, h264/aac). Re-mastered loudness: -14.8 LUFS integrated / -1.8
dBTP true peak (both within spec).

Reproducible recipe: media/video-recipe/line2-fix/ (generate_line2_audio.py
-> render_line2_assets.mjs -> build_patch.sh).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
2026-11-11 02:06:42 +03:00
..

NEAR Hackathon 2026 submission assets — Cred402 SpendGuard

Agent Economy track. Built 2026-11-11, ahead of the Nov 14 23:59 UTC deadline.

Files

  • PITCH-DECK.md — Marp source, 13 slides.
  • PITCH-DECK.pdf, PITCH-DECK.pptx — exported decks (both generated with npx @marp-team/marp-cli, slides rendered to PNG and visually checked for overflow/contrast before export; two real bugs were found and fixed this way — illegible dark-on-dark code blocks on slide 4, and a white-background table on slides 8/9 fighting the dark theme).
  • media/cred402-near-spendguard-pitch-demo.mp4 — 70.6s narrated demo video, 1920x1080, 30fps.
  • media/proof-0{1,2,3}-*.png — real frames extracted (ffmpeg) from the existing media/cred402-near-spend-guard-demo.mp4 recording, used as the deck's "live proof" slides.
  • media/raw/proof-of-life.{cast,gif,mp4} — a new, real asciinema terminal recording made today (2026-11-11) specifically for this submission: git log -1, a live RPC view_account + call_function admin() call against spendguard-024e8dd6f7.testnet (decoded on-screen), cargo test --lib (18/18), and npm test (450/450) — all real commands, real network calls, real output, no staged text.
  • media/video-recipe/ — the video's build recipe: card.html / caption.html (single self-contained HTML templates, every frame a pure function of a t query param, rendered via Playwright/headless Chromium — no AI- generated footage anywhere in this video), render_cards.mjs / render_captions.mjs, captions.srt, final-lines.json (word-level caption timings from Deepgram nova-3's own ASR round-trip on the narration audio), and contact-sheet.png (1 frame / ~2.5s QC sheet of the final video, looked at before shipping). media/video-recipe/line2-fix/ is a later, narrower recipe that re-voices only narration line 2 (see "Judge-review correction" below) without touching anything else in the video.

How the video was built (so it can be reproduced or audited)

  1. Narration: 6 lines synthesized with Deepgram Aura-2 (plans/gtm/tools/deepgram_voice.py, the existing shared tool), each line QC'd via a real ASR round-trip (nova-3) before being accepted — several takes were rejected and retried for real mispronunciations (e.g. "Algorand" → "Albarand" on 4/4 takes, "482" as digits not aligning with its caption) and the script was reworded rather than faked past the check.
  2. Visuals: two real terminal clips (the pre-existing 17s SpendGuard demo and the new 12s proof-of-life clip above) plus four code-rendered title cards (Playwright screenshots of a single HTML/CSS template, eased with a deterministic cubic function of t, not a GIF or stock asset).
  3. Timeline: each card's duration was sized to its own narration's actual measured length (TTS duration is non-deterministic take-to-take); the two real terminal clips kept their real fixed durations (17.0s, 12.9s) and the narration was written short enough to fit inside them with room to spare.
  4. Captions: burned in as timed PNG overlays (this machine's Homebrew ffmpeg build has no libass/drawtext support, so captions are composited via overlay=enable='between(t,start,end)' instead of the subtitles filter) — same real word timings as the narration audio, not hand-typed guesses.
  5. Loudness: mastered with two-pass loudnorm (I=-14, TP=-2.0) plus a final gain nudge and alimiter. Measured on the final exported MP4 with ffmpeg ebur128=peak=true (numbers below are post line-2 patch, 2026-11-11 — see "Judge-review correction" below):
    • Integrated loudness: -14.8 LUFS (target -14 ±1)
    • True peak: -1.8 dBFS (target ≤ -1.5 dBTP — passes with margin)
    • Loudness range: 5.4 LU
  6. QC: a 1-frame-per-~2.5s contact sheet (media/video-recipe/contact-sheet.png, 40 frames) was rendered and visually reviewed before shipping; no blank/unreadable frames found (the single black tile is past the video's 70.6s end, an artifact of the contact-sheet grid, not the video).

Claims re-checked live, 2026-11-11 (same day as this build)

  • spendguard-024e8dd6f7.testnet — live, public NEAR Testnet, confirmed via direct JSON-RPC to rpc.testnet.fastnear.com (view_account and call_function admin()), not cached from an earlier pass.
  • cargo test --lib in contracts-near/spend_guard: 18/18 passed, just now.
  • npm test (repo root): 450/450 passed, just now.
  • GitHub repo caelum0x/cred402: public, confirmed via gh repo view.

Not re-run this pass (unchanged since the 2026-11-10 verification in plans/submission-packets/near-cred402-ration.md, expensive to re-run because it downloads a local nearcore sandbox binary): the 14/14 near-workspaces sandbox integration tests.

Judge-review correction, 2026-11-11

A judging pass against the live deployment found the deck's first draft overclaimed: it said Cred402 is "already live on Algorand and QIE." Live checks show otherwise — curl https://cred402-1.onrender.com/v1/x402/{algorand,qie}/status both return "configured": false right now (pay-to address not set on either chain in production), so neither paid route is actually functioning yet. PITCH-DECK.md (slides 4 and 10) and the exported PDF/PPTX have been corrected to state this honestly: the Algorand/QIE x402 routes exist in the same codebase and are deployed, but are not yet production-live, and NEAR is the only one of the three chains with a public, re-verified on-chain proof trail.

Video fix, 2026-11-11 (same day, follow-up pass): the narrated video's line 2 originally carried the same inflated claim ("...already live on two other chains, and now native NEAR") and was re-voiced rather than left inconsistent with the corrected deck. The new line says: "Cred402 sells credit scores to AI agents over a pay per call payment protocol, and now it's enforced natively on NEAR." — same Aura-2 voice (aura-2-asteria-en, re-derived acoustically from the surviving narration since no voice ID was recorded in the first build — see media/video-recipe/line2-fix/generate_line2_audio.py), same per-line Deepgram QC round-trip, only that one line's audio/caption window touched (8.9s → 17.756s on the film clock, a single static held title card the whole way through — verified by diffing frames across the window before touching anything). Line 1 and lines 3-6 are byte-identical to the previous cut (re-transcribed the whole final audio with nova-3 after the patch to confirm). Full recipe, reproducible with one script: media/video-recipe/line2-fix/ (generate_line2_audio.py → render_line2_assets.mjs → build_patch.sh). Duration unchanged at 70.6s; loudness re-measured at -14.8 LUFS integrated / -1.8 dBTP true peak (both within spec, see "How the video was built" above).