Skip to content

zy839971925-zyy/Agent--deep-research-Workflow-Skills / modular/skills/deep-research

deep-research

Verified

Use when a scenario calls for real-world evidence, pre-build engineering/architecture research, or a linked hybrid; infer the route without requiring the user to label it. Keep the host-selected default model; infer proportionate workflow depth, honoring explicit depth requests such as Ultra.

references/evidence-and-provenance.md

View on GitHub
> Upgrade 2.1: read the updated host contracts (in the complete canonical edition) and migration rules (in the complete canonical edition) for executable admission, review and storage semantics. The shorter root SKILL is the active entry. Older examples are descriptive and must satisfy the current schemas.

# Evidence and provenance

**Trigger:** External, disputed, multi-source, weak-signal, quantitative, private, or consequential evidence matters.

**Reads:** material claims/questions, observations, source material, evidence gaps.

**Updates:** evidence records, source-position assessment, provenance clusters, support relations, confidence/limitations.

**May invalidate:** claims, inferences, decisions, recommendations, artifacts based on misapplied or copied evidence.

**Exit:** material claims are supported, contested, inaccessible, or bounded at the needed level.

**Related:** [retrieval and observation](retrieval-and-observation.md), [inquiry and research](inquiry-and-research.md), traceability and integrity (route to the relevant sibling Skill if needed).

**Return:** [root workflow](../SKILL.md) → Evidence judgment.

Use when external claims, multiple sources, weak signals, disputes, or high-consequence conclusions need traceable support.

## Source gate and evidence gate are different

A source may be admissible, authentic, firsthand, or authoritative and still fail to support the proposition being claimed. First ask whether the source/observation is appropriate to inspect for this question; then ask whether the inspected material actually entails or materially bears on the claim. Do not let source prestige substitute for claim-level support.

## Separate observation, assertion, and inference

Do not let a source citation blur distinct propositions. Split compound claims when support differs. Useful records are:

- **Claim:** precise proposition, scope/time, importance, evidence links, current state.
- **Evidence:** source/version, locator, inspected material or faithful observation, access extent, origin, limitations.
- **Support link:** supports / contradicts / mixed / context only, with applicability.
- **Inference:** conclusion, premises, bridge, assumptions, counterevidence, and reversal conditions.

“Verified” should mean that a source or operation was actually inspected/performed, not that objective truth has been proven.

## Evaluate evidence by fit to the claim

Authority is one signal, not a universal ranking. Consider:

- **position to know / directness:** could this source actually observe the fact?
- **provenance independence:** does it originate independently or copy another source?
- **method fitness:** was the observation or study capable of answering the proposition?
- **specificity/applicability:** same entity, version, population, jurisdiction, metric, and context?
- **temporal fit:** right period and governing version?
- **traceability:** can the supporting material be inspected and located?
- **incentives and conflicts:** what reasons exist to frame, omit, exaggerate, or under-report?
- **cross-evidence consistency:** do sources with different error mechanisms converge?

An official statement is excellent evidence of what an institution officially said; it is not automatically independent proof of every underlying assertion.

## Source position and evidence scope

Always ask whether a source can actually know the proposition it is being used to support. Position-to-know is claim-specific:

- an official body can authoritatively state its own policy or record, but may be a conflicted evaluator of its own performance;
- an operator or user can directly report local experience, but may not establish population prevalence;
- a journalist can report observed/documented events, but an internal mechanism may still be second-hand;
- a benchmark author can describe the test setup, while independent reproduction is stronger evidence of generalizable performance.

Do not upgrade a source beyond the scope of what it could observe.

## Evidence visibility and private/public routing

When material, tag source visibility as `public`, `user-provided`, `workspace/local-private`, or `confidential/restricted`. Having access for analysis does not authorize publication or sending identifiers/content to public tools. Minimize disclosure and prefer local joins/generalized public queries when private evidence must be combined with public research.

## Provenance clusters

Repeated copies do not create corroboration. Ten articles derived from one press release, wire story, dataset, anonymous post, or interview are one origin cluster for the reused proposition.

Whenever possible, trace derivative reports to the earliest accessible source. Distinguish independent observation from independent publication.

## Weak signals, rumors, leaks, community evidence

Do not exclude evidence solely because it is informal, anonymous, social, niche, or non-authoritative. Weak signals can reveal emerging events, field experience, implementation failures, practitioner behavior, or the existence of a claim before formal confirmation.

But separate:

- **“someone claims X”** — supported by the existence and provenance of the statement;
- **“X is true”** — requires evidence appropriate to X.

For a material weak signal, trace the earliest identifiable source, separate reposts from independent observations, seek corroboration with a different failure/error mechanism, and state what would upgrade or falsify the proposition. Do not collect rumor for entertainment when it cannot affect the user's outcome.

## Confidence language

Use calibrated natural language. Do not invent numerical probabilities unless a defensible model or data supports them. Distinguish supported, contested, inferred, rumor/weak signal, unresolved, inaccessible, and discarded where useful.

## Citation discipline

Citations belong near the exact claim they support. Preserve locators for tables, figures, code, filings, posts, transcripts, or passages when material. Do not make a paragraph-end citation appear to support unrelated clauses. Unread or inaccessible material cannot silently support a strong claim.



## Failure-mode independence

Independent evidence means more than different URLs or publishers. Ask whether two sources could fail in the same way: copied origin, shared dataset, shared incentive, same measurement blind spot, same sampling frame, same language/source ecosystem, or the same institutional narrative. Prefer evidence routes whose errors would be exposed by different mechanisms.

For a pivotal finding, capture a concise passage or faithful observation and its exact locator while reading. Keep scope and caveat beside it; a citation without an inspected supporting location is a lead, not passage-level verification. For multi-study comparisons, use the optional [comparative evidence synthesis](comparative-synthesis.md) card when incompatibility could distort the conclusion.

## Source conflict resolution

When credible evidence conflicts, first check whether the sources assert the same proposition. Align entity, version, time, population, definition, metric, method, provenance, and incentive. Many conflicts are scope mismatches. If conflict remains real, do not average or vote; narrow the claim, preserve bounded disagreement, and derive the next discriminating evidence need if the disagreement is material.

[← Return to root workflow](../SKILL.md)
add-product-e2e-evalpaperclipai/paperclipAdd or extend a Paperclip full-stack runner E2E workflow, fixture, matcher, or report evidence path for local or Daytona execution.Stars: 115kUpdated Claude CodeCodexVerifiedInstall
add-runner-evalpaperclipai/paperclipAdd or extend a Paperclip Runner protocol evaluation definition, roster, assertion, or report fixture with provenance and narrow validation.Stars: 115kUpdated Claude CodeCodexVerifiedInstall
diagnose-why-work-stoppedpaperclipai/paperclipCompatibility entry point for failed, stalled, looping, or over-recovered Paperclip work. Use fix-failed-paperclip-run for the full diagnosis, regression, repair, real verification, and green-PR workflow; honor an explicitly requested diagnosis-only scope.Stars: 115kUpdated Claude CodeCodexVerifiedInstall
fix-failed-paperclip-runpaperclipai/paperclipDiagnose and fix failed, stuck, slow, or incorrectly blocked Paperclip tasks and runs. Use with a task/run URL or failure report to trace the root cause, prove a regression red then green, check second-order effects, verify the real workflow, and deliver a PR with passing checks.Stars: 115kUpdated Claude CodeCodexVerifiedInstall