We don't list implementations — what problem was being solved, told through 7 arcs

Red is the failure, green is the fix — the interactive timeline below (01) is the detail·evidence panel for these same seven arcs.

1. 1 What's seen by eye gets hooked; what's read by code gets verified
   ✗ Reading the grid as text by eye, hooked by an orange strip at the bottom of the screen — still level 0 after 124 actions → ✓ Switched to numpy code decoding (frame.find/diff) · 2026-06-30
2. 2 If the measurement is contaminated, the win is void too
   ✗ Testing candidates on a fork produced a fake 0→4 climb — ruled invalid via a correction banner → ✓ Sealed Game.fork (RuntimeError on call)·fork_dependence=0 verified every turn · 2026-07-09
3. 3 All-or-nothing scoring empties out learning
   ✗ all-or-nothing judge + a catch-22 gate pair waiting on each other hollowed out the library → ✓ Changed-cell F1 partial credit + a single advantage_on_union channel · a76ea3a·4807091
4. 4 When the accountant doubles as the judge, a plausible story beats the facts
   ✗ LLM narrative DL mixed past and future, pushing a wrong win-condition to the top of credence (L0 death) → ✓ Split past axis = code fold / future axis = f(V only) — blocked narrative DL · 0714 constitution amendment
5. 5 When the grader fails silently, all trust freezes
   ✗ Passing the board as a list caused 13/21 skills to silently throw → total credence freeze·forced sleep every round (15–34 min) → ✓ One np.asarray at the scoring boundary — 0/21 → 13/21 scored · a524931
6. 6 Memory exists, but if the ration starves it, it's as good as gone
   ✗ The briefing truncated the winner's claim at [:160] — the correct payload never reached the prompt (L3 stuck) → ✓ head_capped: the claim's first segment is never truncated · 83ea9e5
7. 7 If a fatal falsification isn't folded into the score, the same mistake repeats
   ✗ Even after GAME_OVER, the goal that killed the run stayed at 0.50 — the same direction got re-explored right after resuming → ✓ Added a GAME_OVER branch to falsify — in h10 all 6 wincond hypotheses were demoted (β25·27) · abe2a2c

Click = pin · hover = preview · each card's first line is the insight; expand for a 2-sentence problem narrative + supporting runs

Arcs 1–4 are traceable only via commit/page citations (no turns.json capture existed at the time). Arcs 5–7 (numpy boundary · briefing rationing · GAME_OVER falsification) are backed by actual round logs captured on 2026‑07‑16–17 — the table below is the full ledger of those 7 runs, and the arc-panel citations come from the h8·h9·h10 rows of this table.