Evidence-first research provenance
SYNAPSE
An evidence-first AI research agent that turns hard questions into traceable source searches, fetched text, quote-grounded evidence, verified facts, coverage audits, and patch diffs.
Problem / itch
Research agents are too easy to trust when they produce fluent summaries. The itch was to make every claim carry its source, quote, validator state, and coverage gap so a human can audit the chain.
What I built
- Built a provider-first pipeline: planner, parallel research jobs, searcher, source fetcher, evidence extractor, fact checker, synthesizer, coverage auditor, patch applicator, validator, and run-quality scoring.
- Treated facts as reviewable units before prose: fetched sources, source quotes, fact IDs, verified/partial/unsupported states, contradictions, and coverage gaps.
- Added live golden validation that fails on fake URLs, missing quotes, unsupported claims, ungrounded patch operations, missing provider metrics, and hidden fallback behavior.
- Recorded provider metrics, stage timings, token telemetry where available, fallback reasons, run quality, and UI-ready patch operations.
Results
Every important final claim is supposed to trace back to a fetched source, exact quote, and fact ID instead of relying on decorative links.
The coverage auditor compares the answer against the original question and flags unsupported claims, contradictions, weak evidence, and missing evidence areas.
Patch operations explain what changed, where it changed, why it changed, and which evidence supports the edit.
The UI is built around auditability: source quality, evidence items, coverage gaps, validator state, and run quality are visible instead of hidden behind polished prose.
The strongest behavior is refusing false confidence: if the evidence is thin, SYNAPSE should say what is missing instead of pretending the answer is settled.
Status
Currently halted.