auto-sync 2026-07-04T04:28:21Z workspace
Browse files- workspace/README.md +9 -4
workspace/README.md
CHANGED
|
@@ -263,12 +263,14 @@ CTT training/proxy/rollout:
|
|
| 263 |
Use `--no-markdown-report` for README-only runs.
|
| 264 |
- `scripts/eval_ctt_generated_rollout.py`: measured rollout harness that
|
| 265 |
restores states, decodes generated tangents, executes candidates, and writes
|
| 266 |
-
measured candidate rows.
|
| 267 |
-
- `scripts/eval_ctt_rollout.py`: measured-output wrapper.
|
|
|
|
| 268 |
- `scripts/build_ctt_proxy_comparison.py`: proxy comparison/gate table with
|
| 269 |
by-task/by-seed JSON outputs. Use `--no-markdown-report` for README-only
|
| 270 |
runs.
|
| 271 |
- `scripts/build_ctt_rollout_comparison.py`: measured rollout aggregation.
|
|
|
|
| 272 |
- `scripts/build_ctt_outcome_gate.py`: measured outcome acceptance gate that
|
| 273 |
combines rollout support and train-clean selector artifacts.
|
| 274 |
- `scripts/summarize_ctt_runs.py`: global CSV/Markdown summary. Markdown output
|
|
@@ -280,7 +282,8 @@ Dominance and utility:
|
|
| 280 |
- `scripts/calibrate_dominance.py`: conformal-style dominance calibration rule.
|
| 281 |
- `scripts/eval_dominance_selector.py`: LCB dominance fallback evaluation.
|
| 282 |
Reports selected success, coverage, fallback, unsafe execution, PCCE,
|
| 283 |
-
selector regret, and support/selector gaps.
|
|
|
|
| 284 |
- `scripts/eval_learned_dominance_selector.py`: ridge learned dominance
|
| 285 |
selector with basic/context/tangent/source/chart-compat and score-shape
|
| 286 |
features. Use `--no-markdown-report` for README-only runs.
|
|
@@ -289,7 +292,9 @@ Dominance and utility:
|
|
| 289 |
|
| 290 |
Metric and paper artifacts:
|
| 291 |
|
| 292 |
-
- `scripts/eval_metrics.py`: canonical measured/proxy metric evaluator.
|
|
|
|
|
|
|
| 293 |
- `scripts/audit_ctt_paper_artifacts.py`: claim-to-artifact audit for the CTT
|
| 294 |
paper. It scans forbidden wording, paper table inputs, implementation paths,
|
| 295 |
and run artifact contracts, then writes JSON/TeX audit outputs without
|
|
|
|
| 263 |
Use `--no-markdown-report` for README-only runs.
|
| 264 |
- `scripts/eval_ctt_generated_rollout.py`: measured rollout harness that
|
| 265 |
restores states, decodes generated tangents, executes candidates, and writes
|
| 266 |
+
measured candidate rows. Use `--no-markdown-report` for README-only runs.
|
| 267 |
+
- `scripts/eval_ctt_rollout.py`: measured-output wrapper. Use
|
| 268 |
+
`--no-markdown-report` for README-only runs.
|
| 269 |
- `scripts/build_ctt_proxy_comparison.py`: proxy comparison/gate table with
|
| 270 |
by-task/by-seed JSON outputs. Use `--no-markdown-report` for README-only
|
| 271 |
runs.
|
| 272 |
- `scripts/build_ctt_rollout_comparison.py`: measured rollout aggregation.
|
| 273 |
+
Use `--no-markdown-report` for README-only runs.
|
| 274 |
- `scripts/build_ctt_outcome_gate.py`: measured outcome acceptance gate that
|
| 275 |
combines rollout support and train-clean selector artifacts.
|
| 276 |
- `scripts/summarize_ctt_runs.py`: global CSV/Markdown summary. Markdown output
|
|
|
|
| 282 |
- `scripts/calibrate_dominance.py`: conformal-style dominance calibration rule.
|
| 283 |
- `scripts/eval_dominance_selector.py`: LCB dominance fallback evaluation.
|
| 284 |
Reports selected success, coverage, fallback, unsafe execution, PCCE,
|
| 285 |
+
selector regret, and support/selector gaps. Use `--no-markdown-report` for
|
| 286 |
+
README-only runs.
|
| 287 |
- `scripts/eval_learned_dominance_selector.py`: ridge learned dominance
|
| 288 |
selector with basic/context/tangent/source/chart-compat and score-shape
|
| 289 |
features. Use `--no-markdown-report` for README-only runs.
|
|
|
|
| 292 |
|
| 293 |
Metric and paper artifacts:
|
| 294 |
|
| 295 |
+
- `scripts/eval_metrics.py`: canonical measured/proxy metric evaluator. It
|
| 296 |
+
writes JSON, TeX, config, command, hash, and log artifacts; use
|
| 297 |
+
`--no-markdown-report` for README-only runs.
|
| 298 |
- `scripts/audit_ctt_paper_artifacts.py`: claim-to-artifact audit for the CTT
|
| 299 |
paper. It scans forbidden wording, paper table inputs, implementation paths,
|
| 300 |
and run artifact contracts, then writes JSON/TeX audit outputs without
|