agent-harness / docs /ANALYSIS_PLAN.md
cuber12's picture
Publish agent harness research code and paper artifacts
d61821a verified
|
Raw
History Blame Contribute Delete
2.1 kB
# Confirmatory analysis implementation
`scripts/analyze_confirmatory.py` and `scripts/analyze_live_agent.py` are the
deterministic analysis entry points for the paper. The former writes a new,
revision-addressed directory under `results/derived`.
The principal sampling unit is the task. Confidence intervals use 20,000
task-bootstrap samples with a fixed seed. Continuous paired contrasts use exact
two-sided sign-flip randomization tests; binary repair contrasts use exact
two-sided McNemar/binomial tests. P-values are Holm-corrected within each
declared experiment family. E05's three seeds are timing repetitions and are
collapsed within each task–harness–backend unit before systems summaries. E04
stale-index copies are de-duplicated by their recorded shared key.
The analyzed population is:
| Experiment | Cells | Task units | Notes |
| --- | ---: | ---: | --- |
| E01 | 150 | 15 | 10 retrieval harnesses |
| E02 | 105 | 15 | 7 model/localization harnesses |
| E03 | 100 | 10 | 10 repair harnesses |
| E04 | 135 | 15 | 3 harnesses × 3 nested doses |
| E05 | 405 | 15 | 3 harnesses × 3 backends × 3 timing seeds |
| E07 | 70 | 10 | 7 live-agent treatments |
E07 additionally audits exactly 70 task-harness cells, one code revision, raw
response count/model identity, exclusive residency, and immutable artifacts.
Its six exact McNemar contrasts receive one Holm correction. Exploratory
failure accounting distinguishes empty patches, test failures, resolution,
budget termination, assistant stops, and reasoning-text tool intent that the
runtime did not parse as a structured call.
E00 and E06 are development-only. Three superseded E02 artifacts from before
PA-001 are retained but excluded by research-code revision. E03 includes both
declared revisions on either side of PA-003 because that correction was scoped
only to the missing H019 added-symbol cell, as documented in
`PROTOCOL_AMENDMENTS.md`.
The script emits audited run-level CSVs, table JSON, statistical JSON, and all
paper figures in both PDF and PNG. Exact package versions are pinned in
`requirements-analysis.txt`.