Commit Graph

10 Commits

Author SHA1 Message Date
claude be71ac406b router/semantic: slice 23 corpus fast-path reconciliation — DeriveFastPath over the real router replaces the regex mirror, factory/merge validated, corpus rebuilt (dataset_hash unchanged), drift diagnostic and rerun tooling 2026-09-08 04:02:00 +04:00
claude 29f74dd3cc router/semantic: slice 22 residual non-action router — emit step, Go harness (legacy + heads modes), Python experiment 2026-09-08 01:26:56 +04:00
claude fa98e4722e router/semantic: slice 21 deterministic execution-frame guard engine, fixtures, runner and emit step
The guard answers one question — may this utterance become an executable
action — as a three-way policy gate (permissive / blocked / ambiguous) and
never decides what the utterance is. Rules are the encoding of the measured
slice-20 dev-pool discriminators: 126 capability-question rows are 42/42/42
addressed / bare-ability / bare-future; 127 bare можешь+пожалуйста rows are
100% action; can-you-please is 100% action. Reuses the shipped prohibition
parser, morph finiteness and lexicon fillers; reason vocabulary is closed.

Slice 21 (task/725, brief after the accepted slice 20).
2026-09-07 23:43:47 +04:00
claude dc73fc4e31 router/semantic: slice 20 pretrained-rubert-tiny pragmatics specialist tooling + tokenizer audit gate
cmd/semantic-router-experiment/slice20_pretrained.py (516 lines, over
300-line hook): tokenize/grouped/lfo/metrics/runtime/onnx/ceiling
subcommands, MAX_LEN=25, lr grid 1e-5/2e-5/5e-5 x seeds 42/17/7, 4-thread
fp32, early stop on val PR-AUC. slice20_audit.py: WordPiece audit gate
(UNK rate 0.0026, seq p99 19, no loss above 96). artifacts under /tmp/mvn-s20/.
2026-09-07 21:40:49 +04:00
claude 7d31de5a6b router/semantic: slice 19 from-scratch sequence pragmatics specialist tooling + eval index 2026-09-07 17:58:09 +04:00
claude f2b65cd5d2 router/semantic: slice 18 sparse lexical action-gate probe tooling + eval index 2026-09-07 15:35:41 +04:00
claude e80d45f4b3 router/semantic: slice 17 nonlinear MLP probe tooling + eval index 2026-09-07 15:23:39 +04:00
claude 774c0660b0 router/semantic: slice 16 diagnostic — action/non-action boundary analysis
Experiments:
- Residual-only OOF safety metrics
- False-action decomposition by semantic family (7 families, 24 split groups)
- Paired e5 geometry (gap=0.010 between action and capability question)
- Binary action probe (PR-AUC=0.707, ROC-AUC=0.863)
- Cost-sensitive linear classification
- Expanded action-threshold sweep (7 thresholds)
- Voice-like punctuation stress evaluation
- e5 + tiny structural features
- Four-hypothesis comparison table

Verdict: e5 representation is the primary bottleneck. Action seeds and
capability questions are nearly indistinguishable in embedding space.
2026-09-07 14:02:00 +04:00
claude 9df223907b router/semantic: sklearn grouped CV linear head experiment (slice 15)
Python experiment: 6-class logistic regression over frozen e5 embeddings,
grouped cross-validation, calibration, abstention curves, contrast-family
analysis, action-threshold sweep.
2026-09-07 13:52:12 +04:00
claude f8ec77de8a router/semantic: e5 embedding cache for linear head experiment (slice 15)
Go tool that computes e5-small embeddings for every corpus row,
assigns frozen/dev split and grouped CV folds, outputs JSON.
2026-09-07 13:51:54 +04:00