Living Evidence Workspace

An empty document of the same format — where an agent and a human build the evidence base together, and export it as a living document anyone else can cross-examine.

Initializing agent interface…
Hypothesis under examination

(not set — use set_hypothesis)

You (the human)

You are the approval gate and the editor. Nothing an agent proposes enters the evidence base until you press Approve. Curate the claims, watch the ledger, and when the document is worth keeping, export it — one self-contained file, no server, no dependencies.

This workspace saves itself on this device. Export creates a portable document with a content-addressed scientific-state id and self-signed release receipt.

Your agent

  1. get_document_overview first — always.
  2. set_hypothesis to state the question.
  3. propose_study for each record, with a source and a verbatim quote; the human approves.
  4. add_claim with a declarative machine check, then evaluate_claim.
  5. export_document when the document stands on its own.

Import an evidence package

CSV, JSON, Quarto or Jupyter → review cards

Files are parsed locally and never uploaded or executed. CSV requires the complete living-evidence-smd-package/1 traceability contract: record/experiment ids, yi/vi, source, quote, locator, derivation, design, outcome, timepoint, SMD definition/direction, collection frame and explicit RoB status. See the linked schema/producer guide for the exact header. Import validates the whole file first, then stages every row for separate human approval; approval accepts a supplied extraction, not its scientific truth.

· Download sample CSV

Learning example only: invented numbers, not publications or scientific evidence. Stages three review cards in an empty workspace; it does not approve them or change your hypothesis. Review and approve records individually to see a forest plot after the second approval. Never mix this sample with real research.

No package selected. Maximum 1 MiB / 100 records. See producer recipes.

Evidence base

Currently 0 approved effect-size records; pooled random-effects (REML) estimate —, 95% CI —, p = — (I² = —, Q = —, p = —).

The forest plot appears as soon as two effect-size records have been approved, and re-renders on every change to the evidence base.

Proposed evidence records awaiting your approval

Proposed or imported records. They are not in any analysis yet; approval is not source verification.

Claims

Claims added with add_claim. Each carries a document-registered rule expressed as data (no code, no eval). Badges report only whether that rule passed, failed, or was inconclusive — not truth or evidence quality.

Reader’s Workbench

Nothing here yet. Analyses run by your agent (or by you, from the Tool console) render their figures here, newest first.

Audit ledger

Reload-persistent SHA-256 hash chain of analyses, registered-rule outcomes, mutations, receipts and import decisions. Pure reads are not ledgered. The chain detects edits and reordering; it does not establish identity or a trusted timestamp.

    Reproducibility receipt

    Seal the current scientific-state hash and audit head with a browser-session signing key. Exported documents embed the receipt; authorship assurance requires pinning the key fingerprint externally.

    Tool console

    Drive the workspace’s tools by hand (no agent required)

    The same tool surface an AI agent sees through WebMCP — same schemas, same effects, same ledger.