Keeping the scorecard current — the two cadences
An outward scorecard rots in two directions. This is how each is caught and closed.
1. As the industry changes
New techniques appear and published SOTA bars move. Two mechanisms catch it:
- New dimension → coverage drops. When the field starts competing on something new (a new quant format, a new decoding trick), add it to
tools/industry_scorecard.data/_taxonomy.json. It is immediately uncovered, socoveragefalls and the dimension shows up in--gapsuntil fak is positioned on it. - Stale SOTA bar → industry-drift backlog. Every dimension carries a
source_dateand every competitor alast_reviewed.python tools/industry_scorecard.py --stalelists the bars past theindustry_review_window_dayswindow — re-check them on the web and update the number + date. (Advisory, never parity-debt: a number does not become false the day it crosses the window, it wants a look.)
2. As fak changes
- A benchmark lands → a gap becomes a measured row. When a new fak number ships (traced in
BENCHMARK-AUTHORITY.md), turn the relevantno-claimposition into alead/parity/trailsrow, citing the commit/artifact.--gapslists the honest gaps that are candidates to measure. - A fak number ages → re-confirm.
measured_ondrives thefreshnessKPI; old measurements are flagged (advisory) to re-confirm when a bench node is free. - A number changes → never hand-edit the doc. Edit the data file, regenerate.
The commands
python tools/industry_scorecard.py # the scorecard + both work-lists
python tools/industry_scorecard.py --gaps # coverage backlog (what to position/measure)
python tools/industry_scorecard.py --stale # industry-drift backlog (SOTA bars to re-check)
python tools/industry_scorecard.py --verify-sources # fak numbers still match their artifacts
python tools/industry_scorecard.py --markdown-dir docs/industry-scorecard # regenerate this folder
The one rule that overrides everything
Never invent a number or a win. A fak figure must already exist in BENCHMARK-AUTHORITY.md / a committed artifact; a competitor figure must cite a real source. If a fix would require manufacturing evidence, the honest row is a no-claim gap — not a guess. The /industry-score skill enforces this.