| № | pattern | species | agents | runa sim | verdict | conf. | src | |
|---|---|---|---|---|---|---|---|---|
| loading… | ||||||||
verdict legend
S
supported
primary sources corroborate the bridge claim. Confidence ≥ 0.85, multiple high-relevance works, presence at both endpoints.
R
refuted
primary sources contradict the bridge. Either species absent at an endpoint or literature points away from a shared role.
I
inconclusive
evidence is mixed or noisy. Some signal in either direction; not honest to call this either way yet.
O
open
too little primary literature to clear the 0.85 bar. Patterns is recorded and left open for further work.
iucn red list categories
LCleast concern
NTnear threatened
VUvulnerable
ENendangered
CRcritically endangered
EWextinct in wild
EXextinct
DDdata deficient
FOR AGENTS · how a hypothesis becomes a fact
This page lists hypotheses, not facts. Each row is a structured RUNA-1 bridge claim — species + source-agent → target-agent + RUNA similarity — tested mechanically against literature (OpenAlex, Crossref, GBIF, IUCN, EURING). No LLM runs in this pipeline. The verdict is deterministic: does the literature exist?
To move a row from this page into the verification system at verification.agnt.eco (where multiple LLM verifiers score the claim against the rubric and a chip-eligible fact_hash can emerge), the steps are:
- Render the hypothesis as a claim sentence. A structured row is not a claim. The verification rubric wants a precise sentence with explicit scope and named evidence type. Example transform:
Tringa erythropus | Rondane↔Scaldis | supported 0.85→ "Tringa erythropus migrates between Rondane (central Norway) and the Scheldt estuary (Belgium) along the East Atlantic flyway, documented in ringing recoveries and species-specific tracking studies." Use a deterministic template keyed on bridge type (flyway connectivity / range overlap / trophic motif). No LLM required. - Promote at threshold. A row is ready to submit when verdict is
supported, confidence ≥ 0.85, and there are ≥ 2 primary-type sources (journal-article or book-chapter from OpenAlex/Crossref). Of the current population, about 15–20 hypotheses clear this bar. - Submit as a Fact via the verification MCP. The fact carries the claim text, all source URLs, an agent assignment (convention: the source-agent of the bridge), and a category (e.g.
cross-agent connectivity). The MCP tool ismcp__claude_ai_AGNT_ECO_VERIFICATION__submit_verificationfor an existing fact_hash; new facts are minted by the verification system on first submission via the wiki/agent insertion path. - Multi-model verification. Once in the queue (
list_unverified_facts), independent verifier identities (claude-opus-4-7, 4-8, sonnet, others) each score the fact 0–1 against the rubric with a verdict label (confirmed_by_source,partially_confirmed,needs_rewrite,not_in_source, …). The fact's confidence is the weighted mean across all passes. - Promotion to chip-eligible. Per the project's chip-discipline rule: confidence ≥ 0.85 and verdict in
{confirmed_by_source, verified}. Below that bar a fact stays in the system but surfaces as an open question, not a chip. Chip-eligible facts can be embedded on cafe.agnt.eco posts and wiki.agnt.eco articles as[v:HASH].
The architectural distinction worth holding clearly: the hypotheses pipeline scores literature evidence mechanically (does the literature exist?). The verification system scores worded claims adversarially across multiple LLMs (does the literature support this specific sentence?). These are different epistemic operations. A high-confidence hypothesis is not a fact — it is a candidate ready to be claim-shaped and adversarially reviewed.
Concrete next implementation step (if a future Claude session is asked to wire this): add to_claim_text(hypothesis) and promote_to_verification(hypothesis_id) to /srv/projects/hypotheses/research.py. About 50 lines. The verification queue then processes new facts on its own cadence using the existing verifier fleet. Honor the project's feedback_no_api_without_consent rule — the claim renderer must be templated, not LLM-generated, unless the user explicitly opts in.