TODO.md

TODO — MO§ES SigRank submission

Deadline: June 15 2026, 23:59 UTC. Track: Thousand Token Wood 🍄 Completed items are at the BOTTOM. Everything still TO DO is up top.

═══════════════════════════════════════════════════════════════════

⚡ TO DO — DO NOT FORGET

═══════════════════════════════════════════════════════════════════

A. Deploy to the HF Space (parked until you confirm the Space path)

• Confirm how code reaches the Space (HF git remote vs GitHub auto-sync vs manual).
• Set Space secrets (Settings → Variables and secrets) from SECRETS.local.md: SUPABASE_URL + SUPABASE_ANON_KEY (read). Add SUPABASE_SERVICE_KEY ONLY if you want HF-authed visitor rows to persist (curated demo can leave it off).
• DB is live and ready: board reads 11 rows (MO§ES + 10 tokscale) via anon; anon write blocked; sigrank_sessions table + Greatest Hits read path verified.

B. Hand off to Codex → upload to github.com/SunrisesIllNeverSee

• Codex uploads the repo to your SunrisesIllNeverSee GitHub.
• Codex-attributed commits for the $10k prize (must be genuinely Codex's work):
  • test_metrics.py — lock Υ 18,437 / lev 2042 / X 3.31 + telescoping for every SEED row.
  • Real Codex $/1M via OpenAI per-1M prices in parse_codex_submission meta.
  • NOTE: the 2:1→real-ratio anchor refinement and the estimated-row ~ marker were already done this session (Claude), so they're no longer available as Codex commits.

C. Submission (non-negotiable)

• Move Space into build-small-hackathon org.
• Record 60s demo video (script below) + paste link in README.
• Social post (X/LinkedIn, draft below) + paste link in README.
• Add a link back to your GitHub (SunrisesIllNeverSee) in the README.

D. Optional / blocked

• FIX 3 — add "MCH (tokscale read)" as a LABELED instrument-comparison row (your re-pull).
• FIX 5 — (pre-record only) install MiniCPM deps for REAL narration: .venv/bin/python -m pip install torch transformers accelerate sentencepiece

═══════════════════════════════════════════════════════════════════

📋 REFERENCE

═══════════════════════════════════════════════════════════════════

DEMO VIDEO — 60-sec script (open cold on the board)

1. (0-10s) Open ON the leaderboard. Don't explain. Let the eye hit the ~4-orders Υ gap. "This is every operator's token usage, ranked by one number."
2. (10-25s) Point at $/1M column. "The operator at the top is also the cheapest. Efficiency isn't a tradeoff against cost — it IS the cost."
3. (25-45s) Go to Clock Your Signal. Run ./sigrank (or paste ccusage claude --json). New operator drops onto the board in real time with a narrated profile.
4. (45-60s) "Υ = cache × output over input squared. You can't buy rank with volume — padding input is penalized quadratically. That's the whole game." End on board.

SOCIAL POST — draft

"Built MO§ES SigRank for the @Gradio Build Small hackathon: a diagnostic x-ray of the token economy. Paste your ccusage output, get ranked by Net Volumetric Yield — the metric where volume can't buy rank. A 0.5B MiniCPM narrates your architecture. 🍄 Thousand Token Wood. [link]"

BADGES WE'RE APPLYING FOR (confirm this list)

README tags: thousand-token-wood off-brand tiny-titan best-demo minicpm

• Off Brand ($1,500) — custom non-default UI ✓ (theme.py gold/dark)
• Tiny Titan ($1,500) — ≤4B model ✓ (MiniCPM4-0.5B)
• Best MiniCPM sponsor ($2,500) — built on MiniCPM ✓
• Best Demo ($1,000) — needs the video to land
• Codex sponsor ($10k) — needs genuine Codex-attributed commits (see B + CODEX.md)
• Model ≤32B confirmed (MiniCPM4-0.5B)

═══════════════════════════════════════════════════════════════════

✅ DONE

═══════════════════════════════════════════════════════════════════

THIS SESSION (reconcile + importer overhaul + fixes)

• Reconciled local with Devin's PRs #5/#6 (merged); dropped duplicate claudetodolist.md.
• Codex parser fixed — unified Alpha/Beta into one _codex_input_estimate helper (removed both hardcoded /9.0 copies). Beta = operator's REAL Claude input/output ratio; Alpha = AA 2:1 baseline. Verified both pathways + guard fallback.
• ./sigrank --all — runs each provider in turn; Codex reuses Claude's ratio; one failing provider doesn't stop the others.
• Instructions sharpened (app.py Clock Your Signal tab + README How-to-measure): measure each provider separately, never bare ccusage --json; what/where/where-to-input.
• Wild corpus → 10 tokscale.ai operators (replaces old 6), cited in metrics.py + README. All columns populate. MO§ES canonical Υ 18436.98 untouched.
• Supabase migration (via MCP): submitted_at + hf_user on operators, sigrank_sessions table + indexes + RLS (anon read / service insert). Verified.
• Supabase board synced to 11 rows (MO§ES + 10 tokscale, real cost stored).
• Pushed all of the above to github.com/Burnmydays/hf- (main = 39fba4e).

EARLIER

• Engine: metrics.py (4 ints → full ledger incl Avg $/1M); telescoping identity.
• narrate.py — MiniCPM4-0.5B, non-blocking, template fallback, @GPU for ZeroGPU.
• theme.py — gold/dark CSS, log-scaled Υ bars, rarity tiers, board grid.
• app.py — board hero on both tabs, profile, share card, Greatest Hits, HF login, XSS-safe name escaping, ~ estimated marker.
• db.py — Supabase REST read + service-key upsert + session history, SEED fallback.
• FIX 1 persistence-boundary honesty · FIX 2 cost provenance · FIX 4 species/quadrant reframe · FIX 6 grid check.
• local-first importer sigrank.py (auto-reads ccusage, paste = backup).