core-labs/core
core-labs/core
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

docs(lookback): align exemplar category list with actual diff (only dcs/rwc/ma touched in this increment; no ca/ta/dsnq changes)

Browse source
This commit is contained in:
Shay 2026-06-16 21:32:49 -07:00
parent 0846a51373
commit 4a7c07d693
1 changed files with 1 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
docs/analysis/gsm8k-workstream-a-increment-1-lookback-2026-06-17.md
View file

@ -10,7 +10,7 @@
- Added conservative lexeme passes after EX-6: `_HALF_OF_RE` / `_FRACTION_OF_RE` (fractional surface factors normalized to numeric value while preserving exact source_token), `_MORE_THAN_RE` / `_LESS_THAN_RE` (surface the two raw component numbers; composition deferred). - Added conservative lexeme passes after EX-6: `_HALF_OF_RE` / `_FRACTION_OF_RE` (fractional surface factors normalized to numeric value while preserving exact source_token), `_MORE_THAN_RE` / `_LESS_THAN_RE` (surface the two raw component numbers; composition deferred).
- Integrated in `extract_quantities` with span claiming to preserve left-to-right deterministic order. - Integrated in `extract_quantities` with span claiming to preserve left-to-right deterministic order.
- Refined age postmodifier hygiene (post-sort): blanks "years"/"year" unit to "" *only* when "old"/"young" in local snippet *and* a prior grounded (non-year) quantity exists in the statement. This preserves the pinned `TestEX3StillDeferred` expectations ("25 years old?" → unit "years"; "Rachel is 12 years old." → "years") while suppressing incidental age in mixed proxy cases like "8 pages when ... 6 years old". - Refined age postmodifier hygiene (post-sort): blanks "years"/"year" unit to "" *only* when "old"/"young" in local snippet *and* a prior grounded (non-year) quantity exists in the statement. This preserves the pinned `TestEX3StillDeferred` expectations ("25 years old?" → unit "years"; "Rachel is 12 years old." → "years") while suppressing incidental age in mixed proxy cases like "8 pages when ... 6 years old".
- **Exemplars (teaching/admissibility_exemplars/*_v1.jsonl):** New Phase B seeds appended (dcs-v1-002x, rwc-v1-002x, ma-v1-002x, ca-v1-00xx, ta-00xx, dsnq-00xx) with exact schema, provenance `{"source":"phase_b_seed", "author":"operator (Workstream A increment 1)", ... "author_note": "... from proxy refusal ..."}`. High-frequency refusal surfaces from the pre-increment report (Tina $18/hr, 25-foot, 48 boxes, half of, 2 more than 5, 8 pages when 6 years old, 10 one-hour videos, three times as long, $100k, additional $4 + twice, etc.). Three previously added ambiguous exemplars (rwc-v1-0022, rwc-v1-0023, ma-v1-0024) were removed from the ingested corpus per final review to avoid machine-admissible but human-flagged seeds. - **Exemplars (teaching/admissibility_exemplars/*_v1.jsonl):** New Phase B seeds appended for discrete_count_statement_v1, rate_with_currency_v1, and multiplicative_aggregation_v1 (dcs-v1-002x, rwc-v1-002x, ma-v1-002x) with exact schema, provenance `{"source":"phase_b_seed", "author":"operator (Workstream A increment 1)", ... "author_note": "... from proxy refusal ..."}`. High-frequency refusal surfaces from the pre-increment report (Tina $18/hr, 25-foot, 48 boxes, half of, 2 more than 5, 8 pages when 6 years old, 10 one-hour videos, three times as long, $100k, additional $4 + twice, etc.). Three previously added ambiguous exemplars (rwc-v1-0022, rwc-v1-0023, ma-v1-0024) were removed from the ingested corpus per final review to avoid machine-admissible but human-flagged seeds. (No changes in this increment to ca/ta/dsnq categories.)
- **Rebaseline:** `evals/gsm8k_math/train_sample/v1/runner.py` executed (uv + PYTHONPATH); report.json updated. Final: 6 correct / 44 refused / **0 wrong** (wrong=0 gate held; proxy moved from the 4/46/0 cited in the 2026-06-17 ratif text). Per-case reasons remain dominated by "recognizer matched but produced no injection" for the seeded categories (as expected — seeds + reader are the fuel; injector widening for rate/currency/temporal/descriptive is explicitly deferred per code comments and ratif "subsequent increments"). - **Rebaseline:** `evals/gsm8k_math/train_sample/v1/runner.py` executed (uv + PYTHONPATH); report.json updated. Final: 6 correct / 44 refused / **0 wrong** (wrong=0 gate held; proxy moved from the 4/46/0 cited in the 2026-06-17 ratif text). Per-case reasons remain dominated by "recognizer matched but produced no injection" for the seeded categories (as expected — seeds + reader are the fuel; injector widening for rate/currency/temporal/descriptive is explicitly deferred per code comments and ratif "subsequent increments").
- **No other logic:** report.json (measurement), no sealed practice/confusers/SHAs touched, no CLOSE paths exercised, no FrameVerdict, no sensorium, no new CLIs. - **No other logic:** report.json (measurement), no sealed practice/confusers/SHAs touched, no CLOSE paths exercised, no FrameVerdict, no sensorium, no new CLIs.