daniel-ospina

research-verification

"Thin wrapper that standardizes invocation of the research skill's adversarial review gate. Ensures disconfirming queries are run and assumptions are challenged. Invoked by test-routing when domain=research."

daniel-ospina 0 2 Updated 1mo ago
GitHub

Install

npx skillscat add daniel-ospina/agent-infra/research-verification

Install via the SkillsCat registry.

SKILL.md

⛔ This skill MUST be read in full — not skimmed. Formal review gates depend on its workflow.
Skipping steps silently bypasses quality checks. Missing gates = undetected breakages.

Research Verification — Adversarial Review Gate

Overview

Ensures every research investigation includes adversarial review — disconfirming queries, assumption challenges, and source verification. Thin wrapper — delegates to the research skill's built-in adversarial gate.

Announce at start: "I'm using the research-verification skill to run adversarial review."

Invoked by test-routing when domain=research is detected.

When to Use

  • Research investigations (market, technical, competitive)
  • Decision-support research (architecture choices, tool comparisons)
  • Domain exploration (new technology, new market)

When NOT to Use

  • Trivial single-fact lookups ("what version of Node?") → skip
  • Content-pipeline research (keyword/SERP) → content-verification handles this
  • Research already passed adversarial review → skip

Process

Step 1 — Read Domain

Confirm this is a research issue from the test-routing verification plan:

Domain: research

Step 2 — Invoke Adversarial Review

Route to the research skill's adversarial review gate. This runs:

  1. Disconfirming queries: Search for evidence that contradicts the research findings
  2. Assumption challenge: List all assumptions, challenge each with "what if this is wrong?"
  3. Source verification: Check that claims are sourced (internal codebase, external docs, web research)
  4. Alternative framings: How might we achieve the same outcome differently?

Step 3 — Verify Gate Completeness

  • Disconfirming queries were run (not skipped)
  • At least one disconfirming finding was surfaced (if none, flag as P2 — possible confirmation bias)
  • Assumptions are tagged with confidence levels
  • Sources are cited for key claims

Step 4 — Return Report

## Research Verification Report

**Issue:** #N
**Research topic:** <summary>
**Adversarial queries run:** 3
**Disconfirming findings:** 1

### Adversarial Findings
| Query | Finding | Impact |
|-------|---------|--------|
| "What are alternatives to X?" | Y is a viable alternative not considered | Medium — should be added to research brief |

### Assumption Check
| Assumption | Confidence | Challenged? |
|------------|-----------|-------------|
| "Supabase scales to 1M users" | medium | Yes — found counterexample at 500K |
| "Playwright is faster than Cypress" | high | Yes — confirmed by benchmarks |

### Verdict
**PASS** — adversarial review complete. 1 disconfirming finding surfaced.

Pipeline Handoff

Invoked by: test-routing (domain=research)
Dispatches to: research skill (adversarial review gate)
Consumed by: code-review (verification gate check)

What Fails If You Skip

Skip Consequence
Disconfirming queries Research confirms existing bias — missed alternatives, unchallenged assumptions
Assumption challenge Untracked assumptions become silent design constraints
Source verification Claims treated as facts without evidence — decisions built on sand

Continue following the workflow as mandated by this skill. Do not skip steps.