Living Evidence Board — experimental appendix
An unverified conversation-to-graph sandbox for exploring hypotheses, mechanisms, claims, evidence and open questions. It is not a validated scientific evidence base.
What this is. The exemplar and workspace pages are one fixed genre: an SMD meta-analysis, studies in, forest plot out. Most real research questions do not fit that shape — this board generalizes the format's rules (propose → human approval, quotes required, visible ledger, computed-not-fabricated diagnostics) to any mix of hypotheses, mechanisms, claims, evidence and open questions. It is DESIGN.md v3 §7's Evidence Map (Layer 1), first concrete cut.
Extraction is agent work, done with quotes; inclusion is a human decision, one card at a time; the discovery panel below is bookkeeping over what the board contains, never an assessment of the literature.
Experimental appendix — not part of the Pygmalion meta-analysis, its evidence base, numerical reference checks, or software verification suite. The seed is an unverified ChatGPT conversation. Human approval means accepted onto this local board; it does not verify the source, quote, edge, or claim.
The board has NO statistics engine and issues NO verdicts. A claim's tally is edge bookkeeping over active evidence edges (preloaded seed + human-approved additions) — not truth adjudication — and every surface that shows it says so.
Seeded evidence is agent-extracted from a real ChatGPT research conversation, not independently
verified: every seed evidence node carries that label and its cited_as, rendered on its panel.
Nothing here mutates except through propose → human approve; everything that does is ledgered.
…
If you are a human
Click any node — a hypothesis, a mechanism, a claim pill (with its tally glyph), an evidence rect (coloured by kind), or a dashed question box. Every node is keyboard-reachable: tab to it and press Enter. Escape or a click on empty background clears the selection. Approve or reject each proposal from its card in the pending section; watch the discoveries panel update. When the board is worth keeping, export it from the Tool console — a JSON snapshot of every node, edge, discovery and ledger entry.
If you are an agent
board_overviewfirst — always.list_nodes/get_board_diagnosticsto see what the board already has and what it's missing.propose_node(evidence needsquote+cited_as) andpropose_edge— both render an approval card; there is NO agent-side approve tool.get_board_diagnosticsagain after the human approves — what changed, what to chase next.
The map
Left → right: hypotheses (large boxes), mechanisms (small boxes, part-of their hypothesis), claims (pills — evidence_edge_state glyph ● support_only / ◐ mixed / ⊘ contradiction_only / ○ none, plus the supports+/contradicts− count), evidence (small rects coloured by kind — legend on the map), and open questions (dashed boxes, far right). Vertical order groups claims under the hypothesis they connect to most and sits evidence near its claims; the canvas grows to whatever height the current node set needs, so nothing here is ever squeezed into overlap — edge states are bookkeeping over active evidence edges, not truth adjudication.
Detail
Nothing selected yet — pick a node on the map, or let your agent drive
(focus_node). Whatever either of you opens, both of you see.
Computed discoveries
Claims with mixed or missing evidence edges, single-citation-label claims, hypotheses without linked test questions, every open question, and how much of the evidence base is still unverified — all computed live from the board's approved nodes and edges, refreshed after every approval. Bookkeeping, not an assessment of the literature.
Pending — awaiting your approval
An agent proposed these. They are not on the board, the map, or any tally until you decide.
Audit ledger
Append-only record of everything done to this board in your session — by agents, by you, and by the page itself. Pure reads are not ledgered; anything that changes what is on screen or what the board contains is.
Tool console
Drive the board's tools by hand (no agent required)
The same 11 tools an agent sees through WebMCP — same schemas, same effects on the board, same ledger. Nothing an agent can do here is hidden from you.