Research Proof

Pressure-test research claims with falsifiable evidence plans, adversarial checks, frozen verifiers, evidence certainty checks, and proof ledgers.

Why this exists

Agents are too eager to implement.

Research Proof forces them to state the claim, freeze the verifier, name rejection gates, charge hidden costs, and update a proof ledger before moving forward.

The mental model behind Research Proof is simple: treat every promising idea as a claim under test. Borrow the habits that strong research teams use across AI, medicine, mathematics, engineering, and design: freeze the verifier, separate evidence from interpretation, search for counterexamples, charge hidden costs, and only upgrade confidence when the claim survives transfer pressure. References shape the method; they are never treated as proof of a user claim.

Research Proof image generation concept

Release 1.3.0

This release keeps the skill compact while making the proof path easier to follow.

Highlights:

• Makes the selected proof method and next method-specific moves explicit.
• Keeps compact and schema-style answers from dropping raw evidence or rejected shortcuts.
• Adds practical eval-harness guidance for reusable helpers instead of one-off scripts.
• Keeps status language tied to the evidence in the answer.
• Updates the Claude Code plugin to 1.3.0.

The references are intentionally plain: proof ladders, causal attribution, live-source review, clinical evidence review, design research, cross-domain transfer, and eval harness structure.

See CHANGELOG.md for release notes.

Install

npx skills add tonyblu331/research-proof --skill research-proof

Global install:

npx skills add tonyblu331/research-proof --skill research-proof -g

List available skills before installing:

npx skills add tonyblu331/research-proof --list

Manual install:

git clone https://github.com/tonyblu331/research-proof.git

Then copy skills/research-proof into your agent's skills directory.

Claude Code Plugin

This repo is a Claude Code marketplace. Install it with:

claude plugin marketplace add tonyblu331/research-proof
claude plugin install research-proof-plugin@research-proof

Invoke it with:

/research-proof-plugin:research-proof

The plugin wrapper lives here:

.claude-plugin/marketplace.json
plugins/research-proof-plugin/
  .claude-plugin/plugin.json
  skills/research-proof/

Local plugin test:

claude plugin marketplace add .\
claude --plugin-dir .\plugins\research-proof-plugin

Optional local checks:

node .\tools\validate-research-skill.mjs
claude plugin validate .
claude plugin validate .\plugins\research-proof-plugin

Use It For

Use Research Proof when a claim is promising but still vague:

Use research-proof to pressure-test this claim: our agent loop can improve a prompt library overnight without human review.

Good fits:

• research roadmaps
• benchmark reviews
• proof ladders
• cross-domain mathematical transfer
• evaluator-gated loops
• research TDD scenarios
• clinical or intervention evidence questions
• systematic reviews and evidence-certainty checks
• causal inference and observational-data claims
• mathematical innovation by borrowing invariants, constructions, or proof tools from distant fields
• SIGGRAPH-style artifact, rendering, simulation, and perceptual-system claims
• tool-grounded scientific workflows and live-source research claims
• clinical AI reporting, calibration, validation, and deployment-readiness claims
• design research and prototype-readiness claims
• research-program strategy and funding decisions
• adversarial follow-up tests

Tiny Before / After Examples

Vague benchmark claim

Before:

Our new retrieval strategy beats the baseline.

After:

Claim: On frozen dataset D, candidate retrieval C beats baseline B only if answer accuracy improves by at least 5 points, p95 latency stays under budget, and no added manual labeling is required.
Rejection Gates: Reject if the evaluator changes, wins only on visible examples, hides latency cost, or fails a held-out query family.

Messy research paper claim

Before:

This paper proves the method generalizes.

After:

Claim: The paper supports generalization only for domain D, metric M, and comparison B used in the study.
Enemy Terms: Dataset leakage, weak baselines, missing ablations, narrow distribution, and author interpretation separated from measured results.

Agent loop claim

Before:

The agent can improve the repo by looping on issues overnight.

After:

Verifier Boundary: Freeze issues, tests, budgets, writable files, and review authority before the loop starts.
Rejection Gates: Reject if the agent edits the verifier, creates duplicate work, exceeds budget, needs manual cleanup, or cannot reproduce the ledger from evidence.

What It Produces

Research Proof forces the agent to define:

Claim
Verifier Boundary
Baseline / Candidate Family
Current Evidence
Enemy Terms
Rejection Gates
Proof Ladder / Transfer Path
Verdict
Proof Ledger Decision
Next Pressure

Evidence is labeled as PROVEN, SUPPORTED, REJECTED, or OPEN.

Not a magic proof machine

Repo evidence is not universal proof. Research Proof improves research discipline by making the claim, verifier, rejection gates, hidden costs, and ledger explicit, but transfer still needs held-out or production-style validation.

That honesty is part of the product.

Quick Example

Messy claim:

Our autonomous loop can improve a prompt library overnight without human review.

Research Proof rewrite:

Claim
For prompt set D and baseline B, candidate loop C wins only if held-out task score improves by +5% while latency, token cost, regressions, and human review stay within budget.

Verifier Boundary
The evaluator, held-out tasks, scoring rubric, and regression set are frozen before the loop starts. The candidate can edit prompts only. It cannot inspect held-out answers, change tests, widen budgets, or mark its own outputs as accepted.

Rejection Gates
Reject if the candidate changes the evaluator, fails regression, exceeds token budget, improves only visible tasks, or requires manual cleanup.

Proof Ledger Decision
OPEN until it wins the frozen harness and survives transfer.

Next Pressure
Run a transfer test on a new prompt family with the same scoring rules.

See examples/fuzzy-claim-proof-ledger.md for the full worked example.

Distribution

This repository ships the same skill through Claude Code plugins and the open skills CLI. The source of truth is skills/research-proof; the plugin skill directory is a distribution copy kept in sync by the repo checks.

• Source: github.com/tonyblu331/research-proof
• Releases: github.com/tonyblu331/research-proof/releases

Repository Layout

assets/
examples/
plugins/research-proof-plugin/
skills/research-proof/
  SKILL.md
  evals/evals.json
  references/
tools/
  lib/
  validate-research-skill.mjs
.github/

License

MIT