disambiguation-score
One repeatable pass that keeps fak's growing namespace CRYSTAL-CLEAR - each similar-sounding concept (cache / vCache / KV cache / cachemeta / the provider prompt-cache; guard vs gate; the two witnesses) given one canonical name, a written definition, and an explicit line drawn against the siblings it is confused with. Runs the concept-disambiguation scorecard (tools/concept_disambiguation_scorecard.py) over a data dir of confusable-concept rows cross-checked against the tree (the grounding token must appear in the production corpus; the glossary anchor must exist. Use when this named workflow matches the task.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-15
- Stars
- ★ 48
- Forks
- 16
01 前言 Front Matter
- name
- disambiguation-score
- description
- One repeatable pass that keeps fak's growing namespace CRYSTAL-CLEAR - each similar-sounding concept (cache / vCache / KV cache / cachemeta / the provider prompt-cache; guard vs gate; the two witnesses) given one canonical name, a written definition, and an explicit line drawn against the siblings it is confused with. Runs the concept-disambiguation scorecard (tools/concept_disambiguation_scorecard.py) over a data dir of confusable-concept rows cross-checked against the tree (the grounding token must appear in the production corpus; the glossary anchor must exist. Use when this named workflow matches the task.
03 SKILL.md 正文
disambiguation-score - the crystal-clarity / naming pass
An instance of the generic scorecard doctrine. Read that for the five laws and the RSI loop; this file is the per-surface specialization.
Why this scorecard exists
fak has grown a large vocabulary, and several roots are badly overloaded. The word "cache" alone names at least a dozen distinct things (KV cache, vCache, cachemeta, the provider prompt-cache, ViewCache, enginecache, RadixKV, ...); "guard" and "gate" blur; "witness" means two unrelated ideas. As the system grows, a reader cannot tell the concepts apart - and a name a reader cannot pin is a name they will misuse.
The crystal-clarity law, in one line: every distinct concept gets one canonical name, a written definition, and an explicit line drawn against the siblings it is confused with - and that distinction lives in a doc that exists, not tribal memory.
This is orthogonal to its two easily-confused siblings (an irony the glossary notes):
conflation-score keeps every reported NUMBER labeled by provenance (WITNESSED vs
OBSERVED); product-score asks whether a person can USE a concept. This one asks
whether the NAME is unambiguous.
The KPIs (what it grades)
- well_formed (HARD) - required fields present, enums in the closed vocabulary, unique id, family declared.
- canonical_unique (HARD) - no two concepts share THE canonical name. A real collision is fixed by a RENAME, never a note.
- defined (HARD) - every concept carries one distinguishing definition sentence.
- disambiguated (HARD) - the heart: a concept in a family with >= 1 positioned
sibling must carry a
distinctionline AND adistinct_fromthat resolves to a real catalog id. A confusable concept that never says what it is NOT is the core debt. - grounded (HARD) - the grounding token must appear in the production corpus by
STRICT identity (a fabricated
policyloadedthat merely containspolicyis refused). - anchored (HARD) - a concept claiming
crystalclarity must have its distinction written in a glossary_anchor that exists; any non-empty anchor must resolve on disk. - clarity_consistent (HARD) - the declared verdict must match the evidence
(crystal / defined / drifting / colliding / undocumented). Calling an unanchored
concept
crystalis the overclaim this catches. - kind_grounding / hierarchy (SOFT) - advisory: kind agrees with grounding; an
optional
parentresolves and does not cycle.
Plus coverage: for each watched family the scorecard DISCOVERS the confusable
tokens in the tree (a token counts only when it spans >= min_files production files,
or is a package/dir name, or a doc heading), and coverage_debt is the count not yet
positioned. This is the ungameable engine - you cannot shrink the universe by editing
data, only by disambiguating the real concept.
The pass (the shared five-step loop)
- Run it -
python tools/concept_disambiguation_scorecard.py(work-list),--json(payload),--critical(worst-first),--gaps(the coverage backlog by family),--compare baseline.json(prove the drop). - Retire disambiguation-debt worst-first - fix the heaviest KPI by changing
REALITY: rename a true canonical collision, write a missing definition, draw the
distinction+ pointdistinct_fromat a real sibling id, replace a fabricated grounding with a real token. Close a real coverage gap atomically withfak concept position --id ID --canonical NAME --family FAMILY --definition TEXT --distinction TEXT --kind KIND --grounding TOKEN --grounding-kind KIND --glossary docs/fak/concept-glossary.md --distinct-from SIBLING_ID; classify a genuine incidental/false-positive discovery withfak concept classify --family FAMILY --token TOKEN --category CATEGORY --reason TEXT. Preview either mutation with--dry-run --json. Never hand-edit a row, family classification, glossary anchor, and generated README as separate operations. - Weigh the SOFT signals - fix the cheap real ones (kind/grounding, a dangling parent); do not chase them to zero.
- Re-measure + prove -
--compareprints the debt delta; the positioned rows must stay clean (0 clarity-debt) while coverage rises. - Commit exactly the files reported by the authoring verb - verify with
--criticaland regenerate/check the scorecard baseline when the measured metric changes. Runfak sync reconcile --apply, usefak commit --previewfollowed byfak commit --pathfor each reported path (nevergit add -A), and push viafak sync push:
End the subject with thefak sync reconcile --apply fak commit --preview -m "docs(concept): disambiguate <concept> (fak concept)" --path <paths>... fak commit --path <paths>... -m "docs(concept): disambiguate <concept> (fak concept)" fak sync push(fak <leaf>)trailer.
The anti-gaming rule (specific to this surface)
Disambiguate a concept by making the name + the written distinction real, never by
weakening the cross-check. A coverage gap is closed by ADDING an honest, grounded,
anchored row - not by dropping the family root so the token stops being discovered, and
not by listing a distinct_from id that does not exist. A crystal verdict is earned
by writing the distinction into the glossary - not by pointing the anchor at a doc that
happens to exist. The grounding cross-check is STRICT identity on purpose: if a real
concept's token genuinely is not in the production corpus, the honest fix is to rename
the concept toward the name the code actually uses, not to relax the match.
Adding a surface
When a new overloaded root appears (a fresh *-thing family proliferating across the
tree), add a family entry to _meta.json (roots, optional ignore inflections,
optional exclude substrings, min_files); the scorecard then discovers its
confusable tokens and the coverage backlog grows until you position them. Keep the
glossary (docs/fak/concept-glossary.md) as the single anchor the crystal rows cite.
ASCII gotcha
Keep the scorecard source + data + glossary ASCII-only (use - for a dash, straight
quotes). The repo's provenance pre-commit hook crashes on non-cp1252 UTF-8 before its
own escape fires.
04 同仓库更多 Skills
serve-receipt
On-demand research protocol + deterministic receipt for serving a Hugging Face model on a concrete GPU target — fan out research lanes (model facts/files, engine support matrix, serving recipes, KV/state memory math), compute the tiered feasibility verdict with `fak serve-receipt` (never narrated math), and emit a witnessed launch receipt. Use when asked for an "initial config", "serving receipt", or "how do I serve <model> on <GPUs>".
caveman
Ultra-compressed communication mode. Cuts output tokens 65% (measured) by speaking like caveman while keeping full technical accuracy. Supports intensity levels: lite, full (default), ultra, wenyan-lite, wenyan-full, wenyan-ultra. Use when user says "caveman mode", "talk like caveman", "use caveman", "less tokens", "be brief", or invokes /caveman. Also auto-triggers when token efficiency is requested.
quality-report
Score the repo quality and emit a report card.
ctx-overlay
Compact the context window and prune the memory store when a session grows too long.
quality-report
Score the repo quality and emit a report card.
ctx-overlay
Compact the context window and prune the memory store when a session grows too long.