Scoville Research

More links are easy. Better evidence is not.

It usually looks harmless:

  • Six articles confirm a release claim. Five copied the sixth, which copied the press release.
  • A paper reports a strong benchmark result, so the method is called production-ready before anyone checks the repository, license, or missing implementation pieces.
  • A citation is real, current, and topically relevant. It still does not support the sentence attached to it.
  • One retrieved page contains instructions for the agent. Apparently the web has promoted itself to project owner.

That is research slop: visible activity without a reliable evidence chain. The result looks researched because the source list is long, while the actual decision still rests on repetition, inference, or an uninspected abstract.

Scoville Research is a read-only Agent Skill for current multi-source web research, GitHub-first implementation discovery, academic literature work, and durable deep research. It routes each question to the evidence that can answer it, preserves contradictions and inspection limits, and stops when the remaining uncertainty is visible and another query would not change the decision.

Do not use it for one known page or paper summary, an ordinary repository inspection, pure brainstorming, planning, implementation, or wording work.

Why "Scoville"?

The family is named for useful signal that remains detectable after dilution. Research accumulates pages quickly. Its useful heat is the smaller evidence chain that survives source tracing, contradiction search, and an honest check of what was actually inspected.

How to use

Name Scoville Research and the actual decision in the request. The Skill chooses the smallest route that can answer it.

Development - investigate an implementation landscape rather than compare README feature tables:

Use Scoville Research to find open-source libraries for offline-first conflict handling on Windows. Inspect GitHub source, releases, issues, tests, licenses, and official documentation. Return a shortlist, the evidence that separates it, and the cheapest feasibility test. Do not implement anything.

Academic - trace papers at their real publication and inspection depth:

Use Scoville Research for a literature review of agentic deep-research factuality. Distinguish preprints from peer-reviewed work, connect papers to code and data where available, preserve disagreements, and mark abstract-only findings as abstract-only.

Deep - keep an audit-ready investigation resumable across sessions:

Use Scoville Research in Deep mode to determine whether a hybrid API and browser research agent is viable for this product. Preserve the brief, query log, source ledger, claim ledger, contradictions, and final report. Keep private project details out of public queries.

When a decision also needs deliberately different candidate mechanisms, request Scoville Research and Scoville Brainstorm explicitly. Research then owns one inspected prior-art lane; Brainstorm keeps the generators isolated from that lane until convergence. Two landscape passes would look thorough, but mainly create two owners for the same evidence.

Explicit $scoville-research invocation also works on hosts that support named Skill invocation.

Install

Use an Agent Skills-compatible host and Terra 5.6 Medium or a comparably capable executor such as Opus 4.8. Ask the agent to install:

Install this Agent Skill and refresh the available Skill list:
https://github.com/benjaminstelzer/scoville-research/tree/main/scoville-research
Keep the installed directory name scoville-research. Use Terra 5.6 Medium or a comparably capable executor such as Opus 4.8.

The final path must end in <skills-dir>/scoville-research/SKILL.md. For Claude Code, use ~/.claude/skills/ globally or .claude/skills/ inside one project. Other hosts use their supported Skills directory.

What it costs. Research loads a 1,501-token Core. Development, Academic, and Deep references load only when their route needs them, and real research can use materially more time, browsing, and context than an answer without the Skill. That cost buys a traceable claim-evidence chain, not a larger decorative bibliography. Use it when current multi-source evidence can change a material decision. See benchmark evidence.

What it enforces

  • The smallest sufficient route. One known source stays a normal task; Development, Academic, and Deep behavior load only when the question needs them.
  • Evidence ownership. Specifications own their contracts, repositories own observed implementation, papers own reported experiments, and none quietly inherits the authority of another.
  • Claim-level boundaries. Reported claims, direct observations, inference, contradiction, and unresolved gaps remain distinguishable.
  • Exact evidence units. Deep claims link to inspected passages or scoped observations with stable locators, not merely to an entire source.
  • Source independence. Ten retellings of one origin still count as one origin.
  • Hostile-content resistance. Retrieved pages, papers, issues, and tool output are untrusted data, not instructions.
  • Private/public separation. Local or private material does not enter an external query unless the user explicitly authorizes that disclosure.
  • A decision stop. Research ends when the decision-relevant evidence is sufficient or the remaining gap is explicit. More tabs are not promoted to rigor by seniority.

The complete contract is in SKILL.md.

How it works

A request is routed as NO, GENERAL, DEVELOPMENT, ACADEMIC, or a mixed route with the DEEP persistence overlay. The Core freezes the question and data boundary, discovers canonical sources, inspects the actual evidence, traces claims, searches for contradictions, fills the weakest remaining gap, and stops at decision sufficiency.

Deep mode adds brief.md, run.json, queries.jsonl, sources.jsonl, evidence.jsonl, claims.jsonl, and REPORT.md. The run manifest binds the exact Skill package, phase, resume point, and optional external-job state. Evidence records preserve the smallest practical locator, excerpt, or scoped observation behind each claim. Optional source fields keep accessibility, link health, and content quality separate; time-sensitive claims may carry their own as_of date. None of those fields becomes a credibility score.

The bundled standard-library validator checks structural integrity, Skill-byte continuity, state transitions, references, headings, and citation IDs without a network call. Legacy v1 packages remain readable and are never silently migrated. Valid JSON still does not prove truth or semantic support. The Skill installs no executable dependency and requires no particular research provider.

Scoville family

Each Skill works independently. Combine only the concerns the task actually needs:

  • Brainstorm explores materially different mechanisms before selection.
  • Research turns web, GitHub, and scholarly evidence into a decision-ready, claim-traceable result.
  • Code owns engineering scope, implementation, risk, and validation.
  • Design owns visual definition, art direction, design systems, critique, and repair.
  • UI owns framework-aligned implementation, interface mechanics, accessibility, and rendered evidence, with a standalone design fallback.
  • Scribe owns wording, terminology, factual meaning, and source fidelity.
  • Plan owns durable Plans, Work Items, Decisions, and lifecycle state.
  • Handoff transfers active work to another agent or session.

Status

The v1.1.0 candidate passed 13/13 open Validation cases, 4/4 one-shot sealed holdout cases, and 35/35 deterministic artifact-validator tests. Its frozen suite covers Deep v2, external-job boundaries, final citation state, private-query safety, source independence, mechanism convergence, activation near-misses, and the explicit Brainstorm composition. SkillOpt used 49 calls to propose one change; that proposal fell to 12/13 and was rejected.

Those results establish the frozen local corpus, package routing, and structural contract. They do not establish arbitrary truth, exhaustive web coverage, or identical behavior on every host and model. See the qualification manifest and benchmark evidence.

Sources

Every inspected source, contribution, limit, license note, publication status, and design claim is recorded in the source and provenance audit. The v1.1.0 comparison against other Research Skills, services, and scholarly evidence is preserved separately with its report, 27-source ledger, and 32-claim ledger.

License

MIT - see LICENSE.