Skip to main content

algorithm.reviews

An algorithm that reviews, and whose reviewing is itself reviewable.

Open source (MIT) · TypeScript · 19 Vitest tests, all green · Built for DeveloperWeek New York 2026.

The problem

AI answers and online reviews are unauditable. You cannot tell a grounded verdict from a hallucinated one, sources are stale or fabricated, and the AI fact-checkers are themselves black boxes. Buyers, journalists, compliance, and due-diligence teams need a verdict they can defend: a citation trail and a timestamp, not a vibe.

Paste a claim, a "best X of 2026" listicle, a vendor pitch, or a URL. A governed agent fans out across the live web, decides which sources are admissible (fail-closed) while you watch every admit and reject decision stream in real time, forms a verdict grounded only in admitted evidence, and ships a signed review receipt: verdict, per-claim confidence, timestamped citations, dissent, and an ECDSA P-256 signature anyone can check. No admissible evidence resolves to "unverifiable", never "false".

Quickstart

Everything runs keyless by default: no API key required. The offline demo fixtures make no external calls, so they cannot rate-limit or 500. Node with npm.

Clone and install:

git clone https://github.com/OrionArchitekton/algorithm-reviews
cd algorithm-reviews
npm install

Command surface

  • npm run dev: The app: paste a claim, a listicle, a pitch, or a URL, watch source admission stream in real time, and get a signed review receipt with a verdict and per-claim confidence.
  • /demo fixtures: Deterministic offline runs that make zero external calls and cannot rate-limit or 500: /demo/claude-context, /demo/nimble-best, /demo/link-rot.
  • /api/pubkey: Returns the ECDSA P-256 public key as a JWK so anyone can independently verify a review receipt signature without trusting the server.
  • npx vitest run: 19 tests across the governance policy (stale-but-high-authority override, fail-closed on no admitted content), receipt canonicalization and signing, tamper detection, and the UI reducer.

Why it is different

  • A signed receipt, not a vibe: Every review ships as a receipt: verdict, per-claim confidence, timestamped citations, and dissent, sealed with an ECDSA P-256 signature. Anyone can verify it against the published public key, and one changed byte makes verification fail. The review is itself reviewable.
  • Fail-closed source admission: Sources are admitted, not assumed. A glass-box policy decides which evidence is admissible while you watch every admit and reject stream in real time. When nothing clears the bar, the verdict fails closed to "unverifiable" instead of manufacturing a false confidence.
  • Grounded only in what was admitted: The verdict is formed strictly from admitted evidence, with confidence assigned per claim and dissent kept on the record. Policy nuance is honored too: a stale but high-authority source can still override, so the check reflects real trust, not a naive freshness heuristic.
  • Bounded and governed by construction: The agent runs under hard caps in code (at most three claims, two queries per claim, eight extracts, concurrency of two), so a review is bounded, reproducible, and cheap to audit. A two-plane architecture keeps the reasoning separate from the enforcement.

Links