Update experiment index

README.md

@@ -0,0 +1,78 @@
+---
+license: other
+tags:
+  - osworld
+  - agent-evaluation
+  - interface-evolution
+---
+
+# evo-exp — Shadow Interface Evolution experiment archives
+
+Complete, self-contained records of OSWorld-V2 interface-evolution runs: the
+evolved interface code, every agent trajectory, the builder logs, the evolution
+database, the run configuration, and derived result tables.
+
+Each archive under `archives/` unpacks to a single directory:
+
+```
+<namespace>-<timestamp>/
+  MANIFEST.json        provenance, variant lineage, counts, caveats
+  config/              run_config.json, eval splits, seed compaction policy
+  db/                  the evolution SQLite database
+  interfaces/          git bundle of every variant tag + readable snapshots
+  builder_logs/        one JSONL per builder invocation
+  trajectories/        <variant>/<task>/ agent trajectories and per-task results
+  results/             val_matrix.csv, variants.csv, truncated_runs.csv
+  logs/                orchestrator console log
+```
+
+## Runs
+
+### `qwen_syn30v15_128k` — 5 evolution steps at a 128K context window
+
+Qwen 3.8-27B as both builder and evaluator, Prime Agent harness, 45 tasks per
+variant (30 synthetic **ungraded** train + 15 graded val), 45-VM concurrency,
+2 h per-task wall clock. 16.5 h wall clock, 270 episodes, clean exit.
+
+| variant | op | parent(s) | val partial | val binary |
+|---|---|---|---|---|
+| seed | — | — | 5.3750 / 15 = 35.83% | 1/15 |
+| **if-001** | mutation | seed | **5.8508 / 15 = 39.01%** | 1/15 |
+| if-002 | crossover | if-001 × seed | 5.5400 / 15 = 36.93% | 1/15 |
+| if-003 | mutation | if-002 | 5.2550 / 15 = 35.03% | 1/15 |
+| if-004 | crossover | if-002 × if-001 | 5.2001 / 15 = 34.67% | **3/15** |
+| if-005 | mutation | seed | 4.5750 / 15 = 30.50% | **3/15** |
+
+`val partial` is the mean raw score over the 15 val tasks; `val binary` counts
+only full credit. The two metrics disagree: **if-001** leads on partial, while
+**if-004** and **if-005** solve three tasks outright against everyone else's one.
+
+Capabilities the evolution opened that the seed did not have:
+
+| task | best before | opened by | new score |
+|---|---|---|---|
+| 010 | 0.000 (all five others) | if-005 | **1.000** |
+| 105 | 0.350 (four variants tied) | if-004 | 0.970 |
+| 069 | 0.871 | if-004 | 1.000 |
+| 046 | 0.990 | if-004 | 1.000 |
+| 079 | 0.000 | if-002 | 0.220 |
+| 058 | 0.000 | if-004 | 0.100 |
+
+### Read `results/truncated_runs.csv` before interpreting any zero
+
+**22 of 90 val runs (24%) were cut off by the 2 h per-task cap**, concentrated
+on tasks 064 and 108 (all six variants), 079 (five), and 083/058/014/048. Those
+runs write a real score — almost always 0.000 — and carry **no `failure.json`**,
+so they are invisible to infra-failure monitoring. A 0.000 on those tasks means
+"was not allowed to finish", not "cannot do it".
+
+Two further caveats, both recorded in `MANIFEST.json`:
+
+- The **seed** was evaluated under `--no-time-limit` (its results were imported
+  from a standalone eval), while if-001…if-005 ran under `--eval-timeout 7200`.
+  The seed row is not directly comparable on the truncation axis.
+- Synthetic `s_*` tasks are **UNGRADED** (constant zero by construction). Only
+  the 15 val ids carry signal; `train` scores of 0.000 are expected.
+
+Compaction was healthy throughout: 290 compactions, **0 skipped**, at a 131,072
+token window with `keep_recent_tokens = 20000`.