← The Grader Is the Optimizer · published copies: local paths, internal names, the canary's identity and the held-out evaluation are removed.
One row per Study; open a Study for its scores, decisions, diffs, traces and canary. Dev means use each Study's dev-rule metric (score_ungated); "head" is the latest kept revision. Agent failures count runs the framework classes as the Agent's own (invalid artifact, timeout, protocol violation, unverified Harness); "unscored" counts finished runs without a value of the rule metric, for example an Agent execution failure. Canary hits count candidate runs. Efficiency uses each revision's median dev run minutes and median dev run cost at Kimi's kimi-k3 list prices (¥ per million tokens: uncached input and 5-minute cache writes 20, cache hits 2, output incl. thinking 100); "latest vs parent" is the newest candidate with dev runs against its own parent; "loop extensions" lists iterations whose change added a Pi extension. Pages refresh when a Study's ledger or Meta Agent records change; screenshots are left out while runs are timed. Index written 2026-11-01 19:37 UTC.
| Study | Arm | Scope | Lineage | Meta Agent | State | Iterations | Kept / reverted | Baseline dev | Head dev | Last decision | Agent failures | Unscored | Canary | Latest vs parent (dev median) | Head vs baseline (time / cost) | Head cost per dev run | Loop extensions | Meta cost | Snapshot |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| pi-f0-instructions-l1 | F0 | L1 instructions | 1 | kimi-k3 · high | search | 4 | 1 / 3 | 0.711 | 0.806 iter-001 | iter-004: revert (below_bar) | 0/40 | 0 | 2/32 | iter-004: -0.9 min · 1.16× cost | 1.8× / 1.8× | ¥8.16 | – | ¥1044.94 | 19:33 |
| pi-f0-instructions-l2 | F0 | L1 instructions | 2 | kimi-k3 · high | search | 4 | 1 / 3 | 0.746 | 0.833 iter-001 | iter-004: revert (below_bar) | 0/40 | 0 | 0/32 | iter-004: +2.5 min · 1.20× cost | 1.2× / 1.8× | ¥6.06 | – | ¥873.22 | 19:33 |
| pi-f1a-instructions-l1 | F1a | L1 instructions | 1 | kimi-k3 · high | search | 4 | 1 / 3 | 0.704 | 0.878 iter-001 | iter-004: revert (below_bar) | 0/40 | 0 | 2/32 | iter-004: +12.2 min · 1.72× cost | 1.6× / 1.9× | ¥9.03 | – | ¥1634.84 | 19:34 |
| pi-f1a-instructions-l2 | F1a | L1 instructions | 2 | kimi-k3 · high | search | 4 | 1 / 2 | 0.721 | 0.898 iter-001 | iter-003: revert (below_bar) | 1/32 | 0 | 0/23 | iter-003: -5.9 min · 0.66× cost | 1.8× / 1.9× | ¥10.42 | – | ¥783.08 | 19:34 |
| pi-f1a-instructions-l3 | F1a | L1 instructions | 3 | kimi-k3 · high | search | 4 | 1 / 2 | 0.755 | 0.916 iter-001 | iter-004: revert (below_bar) | 0/32 | 0 | 2/24 | iter-004: -0.2 min · 0.95× cost | 2.1× / 2.2× | ¥11.81 | – | ¥1060.10 | 19:35 |
| pi-f2-instructions-l1 | F2 | L1 instructions | 1 | kimi-k3 · high | search | 4 | 1 / 3 | 0.702 | 0.917 iter-001 | iter-004: revert (below_bar) | 0/40 | 1 | 0/32 | iter-004: +3.8 min · 1.23× cost | 1.1× / 1.4× | ¥7.50 | – | ¥1176.17 | 19:35 |
| pi-f2-instructions-l2 | F2 | L1 instructions | 2 | kimi-k3 · high | search | 7 | 2 / 5 | 0.738 | 0.950 iter-004 | iter-007: revert (below_bar) | 0/64 | 0 | 2/56 | iter-007: +9.0 min · 1.80× cost | 1.5× / 1.3× | ¥6.68 | – | ¥2185.25 | 19:36 |
| pi-f2-loop-l1 | F2 | L3 loop | 1 | kimi-k3 · high | search | 4 | 1 / 3 | 0.749 | 0.958 iter-001 | iter-004: revert (more_flagged_runs) | 0/40 | 0 | 0/31 | iter-004: -2.1 min · 1.16× cost | 2.0× / 2.1× | ¥7.91 | iter-004 | ¥1294.16 | 19:37 |