Skip to main content

Adversarial Review

Adversarial review is Hubify’s core differentiator. Every substantive claim — a paper, a result, a derivation — is reviewed by independent models from different labs that are told to refute it, not to improve it. The reviewers share no training priors, so their errors decorrelate, and the disagreements they surface become tracked fixes before a venue referee ever sees the work.
This is a primitive, not a feature. It runs on top of any agent — including Claude Science, Codex, or your own — via the review/adversarial skill and the MCP server.

Cooperative vs. adversarial

In 2026 the whole market converged on multi-agent orchestration — Sakana Fugu, Nous Hermes Mixture-of-Agents, Anthropic Claude Science. All of it is cooperative aggregation: blend N models into one synthesized answer to raise a benchmark score. Hubify runs the other primitive.
Cooperative mixture-of-agentsHubify adversarial review
ObjectiveRaise the scoreCatch the false positive before it ships
Failure it targetsA suboptimal answerA hallucinated derivation, a fabricated result, a headlined-favorable value
Model relationshipBlend / agreeIndependently refute / disagree
Vendor spreadSame-vendor is fine (errors correlate)Cross-vendor is mandatory (errors decorrelate)
Review sourceInternal onlyInternal and independent external
Bias guardNoneIntegrity audit against self-favoring
OutputOne synthesized answerVerdict-first, source-cited, provenance-linked
Mixture-of-agents makes the answer smarter. Adversarial review makes the answer trustworthy. In science the false positive is the catastrophic failure — a retraction, a wasted grant, wrong physics propagated for years — not a lost benchmark point.

Why it matters more in science

  • Falsification is the method. A cross-vendor refutation loop is the scientific method rendered computational. Cooperative aggregation manufactures consensus instead.
  • Correlated errors are the real enemy. Two agents from the same lab share training priors and hallucinate in the same direction. Only cross-vendor independence decorrelates the errors that retract papers.
  • Cost asymmetry. Catching one false positive is worth more than any number of benchmark points.

How a review runs

1

Map the claims

Run claim-map to enumerate every claim with its evidence, so reviewers attack specific statements rather than vibes.
2

Dispatch cross-vendor reviewers

At least three different providers across distinct roles (skeptic, methods, long-context, contrarian, fact-check). Each is a separate model invocation with a role-specific system prompt, and reviewers never see each other’s verdicts — independence is the point.
3

Collect verdicts

Each reviewer returns a structured verdict — approve / concern / reject — with a ranked findings list, a summary, and a confidence score.
4

Integrity audit

A separate skeptical pass, not told the desired conclusion, checks for verdict-severity steering, dismissals that don’t hold against the source, and headlining the more-favorable of multiple values. Verdict: genuine vs. engineered.
5

Synthesize and file fixes

Critical findings from any reviewer are blocking; concerns shared by two or more are blocking. Disagreement is surfaced, never averaged away. Required fixes become tracked tasks; the paper does not advance until they land.

The structural moat

No single frontier lab can offer vendor-agnostic adversarial review without cannibalizing itself. Anthropic’s Claude Science is single-vendor by design — it will never route to GPT or Gemini to check its own work. Sakana Fugu routes around Anthropic. Hermes MoA is a developer tool with no science discipline. Only an independent layer can be the neutral referee — which is why Hubify treats every agent runner as interchangeable execution substrate it sits on top of.

Proven in production

BigBounce — a six-paper bounce-cosmology program — is driven toward publication by exactly this loop. Its cross-vendor + external + integrity-audited review has caught real errors a single model would have missed: an overlap-inflated significance from shared-supernova double-counting, and a mislabeled catalog tier that failed injection-recovery. See the live review timeline at bigbounce.hubify.app/reviews.
  • Agents — the hierarchy that dispatches reviewers
  • Papers — where adversarial review gates the pipeline
  • Skills overviewreview/adversarial, red-team-claim, consensus-synthesize
  • Bring your own agent — run the loop from Claude Code, Codex, or Cursor