Research sequence
- Frame one bounded question.
- Record independent object-identity anchors before joining records.
- Prepare visible, capped searches across registered production provider lanes.
- Express one material claim at a time and cite its exact HTTPS source.
- Record what the source supports rather than what a narrative needs.
- Preserve contradictions for human resolution.
- State and test at least one alternative hypothesis.
- List missing evidence and prior-art searches.
- Assemble a canonical SHA-256-bound reviewer packet.
- Require independent researchers to reproduce, correct, and decide novelty and limitations.
AI boundary
Agents may propose queries, retrieve after a human trigger, extract candidate evidence, check citations, warn about identity collisions, detect contradictions, list gaps, and assemble review packets. Agents cannot authenticate objects, collapse identities, determine title or rights, establish novelty, contact reviewers, publish, or authorize commercial action.
Each assembled browser packet also produces a SHA-256-bound AgentTask work order.
Its eight tasks form a dependency graph with maximum candidate and model-call
budgets. Every task requires approval, retrieval begins blocked, and every
execution, conclusion, publication, contact, and payment authority flag is false.
Downloading the work order performs no task. Its public contract is
`/schemas/research-agent-work-order/v1`.
Before download, verification recomputes the packet digest, removes and recomputes
the signed work-order digest, reconstructs the complete expected work order from the
packet, and requires canonical equality. This binds task IDs, types, order,
dependencies, source plans, candidate/model-call budgets, timestamps, approval flags,
and every false authority. Rehashing an edited DAG or budget therefore cannot make it
valid. The version-1 JSON Schema separately closes task, input, budget, and authority
objects for interoperable structural validation.
Evidence classes
- Observation: directly represented in a retained source.
- Inference: a stated interpretation connecting observations.
- Contradiction: incompatible assertions retained without automatic resolution.
- Alternative hypothesis: a competing explanation that must be tested.
- Gap: evidence required before a claim can be elevated.
- Human decision: an attributable, revision-bound review outcome.
No-result searches are query-bounded observations, never proof of historical absence. Similar labels, titles, makers, dimensions, or images are candidate signals, never sufficient identity proof alone.
Reproducibility and privacy
Packets use the public version-1 schema. Canonical JSON is recursively key-sorted before SHA-256 hashing. Drafts remain in versioned browser-local storage. The application requests no name, email, telephone number, account, ownership history, or uploaded document. Researchers decide whether and how to share a downloaded packet.
Corrections and credit
Corrections cite the packet digest and affected claim, state the evidence and requested change, and preserve the prior version. See Correction and attribution policy(correction-and-attribution-policy.md).
The public contribution hub can assemble the correction envelope entirely in
the browser using the same validation core as the operator importer. Nothing is
uploaded; the researcher retains the downloaded file and its digest.
Documentation contract
This method is canonical under the repository agent-facing documentation contract and is available through `/docs`, `/api/docs/manifest`, and `/api/docs/content`.
The public JSON-LD catalog at `/research/commons/catalog` enumerates only the
already-public methodology, packet schema, contribution guide, correction
policy, expert-review guide, and no-account actions. The public packet schema is
also resolves at its declared `$id`,
`/schemas/research-commons-packet/v1`, with a convenience alias at
`/research/commons/schema`. Draft questions, unapproved
findings, participant data, and private packets are deliberately absent.
A complete fictional packet is available at
`/research/commons/sample-packet`. It demonstrates schema shape and canonical
digest verification while carrying permanent synthetic and evidence-ineligible
flags. See sample-packet.md(sample-packet.md).