LIVE RUN
No run
Pipeline progress
Launch a pipeline to see stage-level progress.
RESEARCH OPERATIONS
Evidence pipeline
Connecting
CHILDSAFE EVIDENCE PIPELINE V3
Literature to framework,
Literature to framework,
fully traceable.
Select one risk category, retrieve recent literature, label natural paragraphs, induce a taxonomy from verified evidence paths, align it to the seed, and update three age-specific rubric cards.
No active pipeline
EVENT STREAM
Latest activity
No pipeline events yet.
RUN HISTORY
Jobs
| Run | Status | Created | Attempt | Mode | Risk category | Actions |
|---|
METHOD AT A GLANCE
Agentic evidence-to-framework workflow
Each run accepts exactly one F01–F10 risk category. Evidence first induces a taxonomy branch; the complete frozen seed is consulted only during alignment, then the selected branch and its three age-specific rubric cards receive candidate updates.
LLM Agent
Deterministic system
Verification gate
Research output
AAggregate & verify evidenceSearch, paragraph-label and fact-check
INPUTOne risk categoryExactly one of F01–F10; no taxonomy file path is
entered.
AGENTQuery PlannerGenerates versioned, engine-specific youth and protection
queries.
SYSTEMRetrieve, deduplicate & screenFour core engines, Scholar coverage, 2023+, dual Agents and
Dynamic Drop.
SYSTEMNatural paragraph chunkerCreates stable paragraph IDs with paper, page, offsets and
text hashes.
AGENTParagraph LabelerLabels every paragraph; one neighbouring-context retry is
allowed.
AGENTAtomic Claim ExtractorExtracts one supported conclusion per claim from selected
paragraphs.
INDEPENDENT GATEClaim VerifierChecks exact quote, entailment, age, sample, limitations
and causal wording.
OUTPUTVerified evidence packagesParagraph + labels + spans + supported claims +
provenance.
BLocal type-path inductionInterpret each evidence package
AGENTLeaf-Type AgentInfers the most specific risk mechanism stated by each
package.
AGENTPath GeneratorCreates one or more root-to-leaf type paths for that
package.
GATEPath ValidatorFixed risk root, progressive specificity, maximum depth
five, no duplicate nodes.
OUTPUTValidated local pathsEvery path retains package, claim and source evidence
IDs.
CGlobal path mergingBuild one graph from all local paths
SYSTEMLabel normalizationNormalizes case, punctuation, plurals and
abbreviations.
SYSTEMMiniLM candidate retrievalFinds the Top-5 possible synonym pairs without merging them
yet.
AGENTNode Relation VerifierDecides same, broader, narrower, related, distinct or
uncertain.
OUTPUTWeighted path graphRetains aliases, document support, path support and claim
IDs.
DHierarchy verificationAccept only supported is-a structure
SYSTEMDeterministic pruningRemoves self-loops, inverse edges and cycles.
INDEPENDENT GATEEdge VerifierChecks each is-a relation and its supporting source
evidence.
PASSEvidence-induced treeMain parent tree plus verified cross-links.
FAILQuarantineUnsupported, conflicted or malformed structures remain
auditable.
ESeed alignmentCompare after evidence induction
FROZEN INPUTComplete seed taxonomyAll ten formal v1 risk branches remain unchanged.
AGENTSeed Alignment AgentClassifies equivalent, narrower, broader, novel, split,
merge or conflict.
PATCHStatement-level patchUpdates only the selected branch and retains all evidence
IDs.
OUTPUTComplete candidate taxonomySelected branch updated; the other nine branches are
preserved.
FRubric constructionConvert verified structure into evaluation rules
AGENTFramework Statement AgentDrafts definitions, age conditions, protective principles
and context limits.
INDEPENDENT GATEStatement VerifierChecks each statement against its node, claims and exact
paragraphs.
AGENTRubric BuilderUpdates cards for ages 7–9, 10–12 and 13–17.
INDEPENDENT GATERule & lineage verifierRubric → node → path → package → paragraph → paper.
OUTPUTCandidate releaseComplete taxonomy, 30 cards, ontology, ledger, diffs and
trace report.
END-TO-END LINEAGERule → Statement → Claim → Span → Page → Source → Query
FAILS A GATEQuarantine with reason; never silently propagated
VERSION CONTROLFormal taxonomy and rubric v1 remain frozen
DISCOVERY
Retrieval coverage
Top-100 coverage is shown before eligibility screening. Missing engines remain visible.
Google Scholar coverage & downloads
Checking live provider availability…
Select a run to inspect Scholar coverage.
Select a run to inspect engine coverage.
Source funnel
No retrieval data.
Dynamic Drop
No cutoff decisions.
MULTI-AGENT REVIEW
Screening decisions
Jev checks title and abstract relevance first. Population-first and Risk-first Agents then make independent decisions; material disagreement invokes adjudication.
Ranked papers and Agent probability
| Rank | Family | Paper | Origin | RRF | Citation | Agent p | Jev p | Decision |
|---|
PROVENANCE
Evidence explorer
Inspect natural-paragraph labeling and verified evidence packages. Public excerpts are limited to 280 characters.
Natural paragraphs and labels
| Paragraph | Page | Label | Scores | Evidence roles | Excerpt |
|---|
Trace an entity
Enter an entity ID or open a paper from the screening table.
EVIDENCE-INDUCED STRUCTURE
Taxonomy induction
Local type paths are merged globally, hierarchy edges are independently verified, and only then aligned to the frozen seed.
Validated local type paths
No paths yet.
Verified taxonomy nodes
No induced nodes yet.
Seed and rubric patches
No candidate patches yet.
OUTPUTS
Candidate release
Formal taxonomy and rubric v1 remain unchanged. Public downloads are sanitized allowlisted artifacts.
No completed release selected.
Artifacts
No artifacts available.