Skip to content

Latest commit

 

History

History
47 lines (32 loc) · 3.55 KB

File metadata and controls

47 lines (32 loc) · 3.55 KB

FAQ

Definitions used in this repository, and direct answers to questions the measurements invite. If an answer here conflicts with docs/metrics.md, the metrics win.

Definitions

Term Meaning as used here
Architectural parameter An ISA-constrained implementer choice: a name, a value domain, and the extension that defines it
Adjusted recall Part I metric allowing documented name alignments rather than exact string equality only
WARL parameter The parameter is the set of legal values of a WARL CSR field, not the field itself
Schema-valid Passes param_schema structurally; not a claim of architectural correctness
Cross-model agreement Shared parameter names between two models; low Jaccard is a review signal, not truth
Exploratory null Result that finished under documented limitations and is reported rather than discarded

Gold names in the prompt

Were the gold parameter names given to the model?
Yes, for the published Part I-style figures. Prompts inject all 185 gold names with an instruction to reuse exact names when they match. Those figures measure grounding (locate and evidence catalogue entries), not discovery without a catalogue. Discovery without the list is unmeasured; see artifact_c/PREREGISTRATION.md.

Does that change how WARL rates should be read?
Yes. WARL at 12/24 with all correct names already supplied is an identification failure, not a missing-vocabulary failure. Prompt-only WARL guidance that increases labels without improving matches is consistent with that.

Stability and scoring

Why report exact and adjusted recall separately?
Exact requires identical names; adjusted allows Part I alignments. Headline remeasure uses adjusted; both appear in metrics.

Why does recall fall from 72.9% to 64.2%?
The gold set grew from the 185-parameter Part I freeze to the 223 UDB carried at corpus pin c184e313. Same LLM output, larger denominator. 223 is the count at that pin and not a live figure: main carried 227 when this was first published, which is corrected on #2053.

Why does a 3.8% name Jaccard matter?
Same prompt and chunk set across models, so disagreement is largely model-side. Dual-model agreement alone is not a sufficient review gate (derived non-parameters have appeared in high-confidence overlap lists).

Prompt v3 produced more WARL labels but fewer correct ones. Why keep it?
It failed cleanly: more labels, worse matched WARL rate. Prompt-only guidance is the wrong lever if the failure mode is identification.

Temporal holdout returned 0/10 in both arms. Is the harness broken?
The harness completed under a pinned model and immutable run directory. Limitations of prompt v1.2 (including label-revealing negatives) are documented in the results. That is an exploratory null, not proof for or against CSR context.

Upstream

Where is the list of UDB PRs and issues?
EVIDENCE.md, re-checked against GitHub as of the census date on that page.

Are issue/PR comments part of the prework measurements?
Yes. Several recall and taxonomy figures were first or also posted on UDB threads (#2053, #2163, #2200, #2251). They are indexed in metrics.md §8 next to the local tables. Cite the issue link when you mean the public discussion; cite this repo when you mean the offline artifact re-derivation.