feat(sources): source-authority scorer — P1 foundation (#111)#112
Merged
Conversation
Pure, deterministic per-URL trust scorer — the foundation for ranking which fetched sources to keep and for a source-trust signal distinct from citation support. Boost-led: gov/edu/standards/academia/official-docs = primary, wikipedia/SO/github = reputable; unknown is neutral (niche-but-legit not punished); low is a conservative seed denylist of observed content farms. Not yet wired into the keep loop — next commit. Validated on the 2026-06-15 dogfood domains: content farms -> low, AWS/Fed/redis docs -> primary.
This was referenced Jun 16, 2026
askalf
added a commit
that referenced
this pull request
Jun 18, 2026
…(P1 of #111) (#113) Wires the #112 scoreAuthority module — until now a pure, unused export — into the fetch keep-stage so authoritative/primary sources win the limited fetch slots ahead of whatever search happened to rank first. - source-authority.ts: add rankByAuthority(items, urlOf, mode) + SourceAuthorityMode. "prefer" (default) stable-sorts by authority and drops nothing; "strict" also drops known content farms, with a min-keep floor so a niche/recency round that only surfaces farms isn't zeroed out; "off" is identity. - agent.ts: rank this round's candidates before the slot-limited slice. - config.ts / cli.ts: --source-authority / DEEPDIVE_SOURCE_AUTHORITY (prefer|strict|off), default prefer; flag validated, env falls back. - 5 unit tests for rankByAuthority (prefer ordering + stability, strict drop + min-keep floor, off identity). Full build + config/agent/cli suites green (117 pass / 0 fail). Default is "prefer" per the issue — a reorder-only boost that drops nothing, so it cannot starve a query. P2 (trust signal in output/--json) and P3 (bench metric) follow.
askalf
added a commit
that referenced
this pull request
Jun 18, 2026
Ships the source-authority layer (#112/#113/#114): deterministic, no-LLM domain scoring (scoreAuthority), keep-stage ranking that gives authoritative sources the limited fetch slots (--source-authority prefer|strict|off), and a source-trust signal in the footer + --json — orthogonal to citation support, so a fully-cited answer built on content farms is flagged instead of silently confident. Version bump + CHANGELOG + README doctor sample only.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
First commit of #111 (source-authority ranking). A pure, deterministic per-URL trust scorer — no LLM, no network — that answers the orthogonal question to the lexical verifier: not "does the answer match its source?" but "is that source trustworthy?".
scoreAuthority(url) → { tier: 'primary'|'reputable'|'unknown'|'low', score, reason }*.gov/*.edu/*.mil+.ac.*TLDs; academia/standards (arxiv, aclanthology, ietf, w3.org, nature…); official docs (docs.*/developer.*prefixes + curated vendor set).reputable= wikipedia/SO/github.unknownis neutral, not punished — a niche-but-legit source shouldn't be downranked for being unrecognized.low= small, conservative seed denylist of content farms seen in the 2026-06-15 dogfood (precision over recall).Pure + unit-tested (9 cases: TLDs, subdomain matching, denylist-beats-boost, gov/edu look-alike false-positive guards, unparseable input, monotonic scores). Validated on real dogfood domains:
aiflashreport.com/gpt0x.com→ low,aws.amazon.com/federalreserve.gov/redis.io→ primary,slingacademy.com→ unknown.Not yet wired into the keep loop or output — this is the foundational unit. Next: keep-stage ranking (rank candidates by authority before the maxSources cut) and the source-trust signal in
--json/footer. No version bump (not user-facing until wired).