Skip to content

string-audit v0.1.0 — cost-aware, grounded content auditor

Choose a tag to compare

View all tags
@bdelanghe bdelanghe released this 22 Jun 23:58
Immutable release. Only release title and notes can be modified.
v0.1.0
85c91ed

A content auditor that audits typed string symbols, scopes checks by type, and
content-hash caches so only changed copy pays for the (expensive) LLM call — and
whose rewrites can't invent facts.

Audits

  • Type-scoped — headline / cta / meta / claim / body, each with its own checks.
  • Grounding guard — claims may only assert facts in the grounding source; ungrounded stats are flagged, never fabricated.
  • Spell-check — portable modern wordlist + a custom dictionary.txt.
  • Grammar/style — write-good.
  • Overlap — duplicate / near-duplicate copy across symbols.
  • Extract + coverage — scan a surface → propose typed symbols; covered / uncovered / unused.

Cost & provenance

  • CAS-memoized — key = sha256(version:type:value); warm run = 0 calls; identical audits dedupe to shared blobs.
  • Pluggable store (same get/put/has port): fs · cas (content-addressed + anchored-chain input→output lineage) · socket (a store daemon mounted on a Unix socket, in a room).
  • Anthropic provider — structured output via tool-use, cost-aware model, grounding enforced; request/parse path test-verified (live run is keyed).

Reuses the bounded-systems stack

cas (content addressing) · anchored-chain (lineage) · brand/content (typed symbols) · prx (budget). CI pipeline green.

Assets 3
Loading