chore(semantics): trim generated research artifacts

Co-authored-by: Rocky <hkgood@users.noreply.github.com>
2026-08-27 10:24:15 +00:00
parent de9cd332a5
commit 1ad8cb1edf
8 changed files with 6 additions and 69751 deletions
@@ -70,6 +70,8 @@ ModelTraining/ClipboardSemantics/open-training-corpus.jsonl
 ModelTraining/ClipboardSemantics/combined-training-corpus.jsonl
 ModelTraining/ClipboardSemantics/Consensus/*.jsonl
 ModelTraining/ClipboardSemantics/VerifierCandidates/
+ModelTraining/ClipboardSemantics/IterativeResearch/**/rounds.json
+ModelTraining/ClipboardSemantics/IterativeResearch/**/final-report.json

 # Python bytecode
 __pycache__/
@@ -153,9 +153,10 @@ The 2026-08-27 study completed the requested 20 rounds and two additional
 - The current deployed reference remains stronger: random-holdout macro F1
  `0.7669` and research-only comprehensive macro F1 `0.2783`.

-All three phases failed the release gate, so no model was promoted. Detailed
-rounds, final evaluations, source breakdowns, and macOS replay instructions are
-stored under `IterativeResearch/`.
+All three phases failed the release gate, so no model was promoted. The compact
+study summary and macOS replay instructions are stored under
+`IterativeResearch/`; detailed round and evaluation JSON is reproducible and
+gitignored.

 This harness is deliberately a Linux surrogate. It cannot emit the
 `NLModel`-compatible Create ML artifacts used by the keyboard extension.