Evidence for connorjarvis.co.uk/ai
==================================

Raw data behind the Rich to Richer numbers on the AI page. Rich to Richer is a Roblox game built by
studio-kit, Connor Jarvis's autonomous studio of Claude, Codex and Gemini agents. The studio's code is
private; these files are its own logs, stripped to the fields the numbers need.

Snapshot: 24 Oct 2026, 22:55 UTC (co-op log to tick 95).

Files
-----
ledger.jsonl        One row per model request (26,018 rows): time, tick, role or developer, model, and
                    input, output, cache-read and cache-write tokens. Written by a Claude Code plugin
                    hook in the studio (one row per request, tool call and subagent spawn; tool rows are
                    left out here because they hold shell commands).
coop-events.jsonl   Every co-op event (7,058 rows): cards queued, claimed, handed back, checked by the
                    merge gate (review-prep: stage and ok), reviewed (verdict: result and reasons),
                    merged, bounced and reclaimed. Free-text notes, branches and commits are left out.
example-rejection-RR-942.txt
                    One reviewer verdict in full: Claude Opus reviewing a change by the Gemini developer
                    and sending it back, with the failing acceptance criteria and what "fixed" means.
verify.mjs          Recomputes every number below. Node 18+, no dependencies:
                      node verify.mjs                                         (files beside it)
                      node verify.mjs https://connorjarvis.co.uk/ai/evidence/ (fetches them)

What it prints (this snapshot)
------------------------------
Model requests            26,018, 19 Oct to 24 Oct 2026
Input tokens incl. cache  5.12B
Output tokens             38.1M
Cost at API rates         $1,975.92  (Opus $1,491.98, Sonnet $426.39, GPT-5.6-Luna $38.90, Gemini $13.97)
Merges                    655 (654 distinct cards)
Merges by developer       Claude 365, Codex 240, Gemini 50
Review verdicts           655 pass, 105 bounce, 103 reclaim, 28 stale
Rejection reasons         acceptance criterion missed 67, idiom 31, check red 17, scope 10,
                          text not by id 6, out of claim 4, not reproducible 3, setting voice 2,
                          tests not first 2, style budget 1
Merge-gate failures       checks 48, merge conflicts 29, file claims 4

Caveats
-------
- The cost is what the same tokens would cost at public API rates. The studio runs on subscription
  plans, so it is not a bill. The price table is in verify.mjs with its sources.
- 223 ledger rows name no model and one model has no published price; both count as $0, as they do
  in the studio's own dashboard.
- This is self-published data. It shows the numbers are computed, not typed; it cannot prove the logs
  were never edited.
- Test counts are not in these files. The latest is in the tick 94 commit: "check --no-cache
  5351/0/302" (passed / failed / skipped).