Skip to content

docs: rewrite README, add LICENSE, prune internal docs#2

Open
claygeo wants to merge 1 commit into
mainfrom
docs/readme-overhaul-and-cleanup
Open

docs: rewrite README, add LICENSE, prune internal docs#2
claygeo wants to merge 1 commit into
mainfrom
docs/readme-overhaul-and-cleanup

Conversation

@claygeo

@claygeo claygeo commented Jun 14, 2026

Copy link
Copy Markdown
Owner

Deep investigation + domain research pass on the repo, then a full documentation sync. The old README read accurately for the project as it stood weeks ago, but the codebase moved on and the numbers drifted by ~10x.

What changed

Removed internal docs (private paths / launch drafts / stale plan — git history keeps them):

  • HANDOFF.md, LAUNCH-X.md, PLAN.md

Added LICENSE (MIT) — the README already pointed at it but the file didn't exist.

Genericized scripts/VPS-SETUP.md — dropped a personal local filesystem path + the VPS IP, fixed the dead crucible-ai.netlify.app domain.

Rewrote README.md to match what actually shipped:

  • Corrected scoring math. Log-loss clamp is [0.01, 0.99] (README said 1e-4). The leaderboard P&L is a flat-$25 directional bet, not the Kelly formula that was shown — that belongs to the trading lab. Eivra Score formula confirmed; clarified ELO does not feed it.
  • Documented the paper-trading proof lab (/trading), the /live page, and the ~20 /api/trading-* endpoints — an entire shipped subsystem the old README omitted.
  • Fixed the claude -p "$0 API cost" details — real flags (--max-budget-usd 0.30, --max-turns 5, Task in --disallowedTools), the actual cache key, %TEMP% vs /tmp, and the synthetic backfill created_at.
  • Refreshed structure tree + schema (added paper_trading_snapshots, evidence_events; dropped per-table row counts that rot on every cron run).
  • Honest positioning vs ForecastBench, Prophet Arena, Metaculus FutureEval, and Halawi et al., and softened the overstated "first live stats-honest scoreboard" claim.
  • Strengthened the honesty section: flagged that ~0.02 Brier / ~97% win across all six agents is backfill inflation, not superhuman skill — superforecasters land ~0.15-0.20, the best published LLM forecaster ~0.24.
  • Fixed the author link (@deforestpeg was pointing at the suspended /claygdev handle), removed the dead HANDOFF.md link and the references to the deleted files.

How it was checked

A multi-agent verification pass queried the live Supabase prod DB for ground-truth counts and cross-checked every numeric/path/mechanism claim against the code; the corrections above come from that. Web research grounded the competitive landscape and the superforecaster Brier baselines.

Rewrite README to match the shipped system:
- correct scoring math (log-loss clamp [0.01,0.99]; leaderboard P&L is a
  flat-$25 directional bet, distinct from the quarter-Kelly trading lab)
- document the paper-trading proof lab, the /live page, and the
  ~20 /api/trading-* endpoints (the biggest doc gap)
- fix the claude -p "$0 API cost" mechanism details (real flags,
  cache key, per-call budget, Task in disallowedTools, %TEMP% vs /tmp)
- refresh the project-structure tree and schema (add paper_trading_snapshots,
  evidence_events; drop row counts that rot on every cron run)
- honest positioning vs ForecastBench / Prophet Arena / Metaculus FutureEval /
  Halawi, and flag that ~0.02 Brier / ~97% win is backfill inflation, not
  superhuman skill (superforecasters land ~0.15-0.20)
- fix author link (@deforestpeg pointed at the suspended /claygdev handle)

Add MIT LICENSE (README already referenced it).

Remove internal docs that don't belong in a public repo:
HANDOFF.md, LAUNCH-X.md, PLAN.md (private local paths, launch-thread/DM
drafts, account-suspension notes, stale build plan).

Genericize scripts/VPS-SETUP.md: drop the personal local path and VPS IP,
fix the dead crucible-ai.netlify.app domain.
claygeo pushed a commit that referenced this pull request Jun 19, 2026
Echo is currently rank #1 (not #2) — the insight panel text was stale.
Also wires rank_delta_24h arrows (↑/↓) to agent cards so rank movement
is visible when standings shift ahead of Thursday launch.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QGt65pxBC4WKvuLUukSRhL
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant