ProductEvidence
Every claim has an address.Every source is kept.
Research runs plan their own queries, triage results for relevance, and store what they fetch. Each source becomes atomic claims with a locator, freshness, and credibility, and generated text cites claim IDs instead of asserting.
- one proposition each, with a locator
- Atomic claims
- snapshots kept with the project
- Stored sources
- off-topic results are never fetched
- Relevance triage
- corroboration and contradiction
- Claim linking
How it works
From a search resultto a claim you can cite.
- 01
Plan and search
A research run turns the project focus into queries and gathers candidates from search providers.
- 02
Triage before fetching
A compact judge scores each candidate for relevance. Low scorers are dropped, so off-topic pages never become claims.
- 03
Store and extract
Pages pass through a guarded fetcher, are deduplicated by content hash, and are split into atomic claims with locators.
- 04
Link and cite
Claims are compared for corroboration and contradiction. Concepts, evaluations, and documents cite them by ID.
Evidence-gap loop
Thin evidence is a task,not a verdict.
When evaluators rate the evidence weak or coverage falls short, LiteSurface turns their stated gaps and the concept’s riskiest assumptions into targeted research, then evaluates again. The new evaluation is appended; the earlier one stays exactly as it was.
- Coverage is a soft gate: it caps the recommendation instead of rejecting outright
- Research questions come from the evaluators’ own rationales
- Balanced and deep runs close gaps for the strongest concepts automatically
- Every pass is recorded, with its cost tracked against your budget
Capabilities
Provenance by default,not as an afterthought.
Locators on every claim
Claims point to the passage they came from, so a reviewer can check the exact paragraph in seconds.
Freshness and credibility
Each source and claim carries a retrieval date, a freshness estimate, and a credibility rating.
Contradiction detection
Claims that disagree are linked across sources, and both sides stay visible wherever either is cited.
Untrusted by default
Fetched text is treated as data, never as instructions, and kept apart from the prompts that steer a run.
Deduplication and caching
Sources are deduplicated by content hash and extraction is cached, so repeat research does not pay twice.
Decision relevance
Claims are scored for how much they bear on the decision, and retrieval ranks by that score.
What you getA library you can audit.
Evidence belongs to the project. Share or transfer the project and every source, claim, and link moves with it.
- Stored source snapshots with metadata
- Atomic claims with passage-level locators
- Freshness, credibility, and relevance on each claim
- Corroboration and contradiction links
- Citations from generated documents back to claims
- Research runs that resume after a failure
Stop arguing from memory.Start from the source.
Explore the evidence library in the sample project, then connect your own provider keys for live research.