Skip to content

Epic J: Cross-Source Retrieval Benchmark — authority-aware routing and evidence-bound synthesis #20

Description

@hummbl-dev

Parent

Status

CLOSEABLE BENCHMARK EPIC — GOVERNED EVALUATION PACKET — NON-CANONICAL UNTIL VALIDATED

Objective

Build a benchmark that tests whether agents and retrieval systems choose the right source families, resolve the right records and versions, preserve exact provenance, find contradictions, and synthesize evidence without flattening source authority.

Benchmark paths

Scientific claim

media or web lead → scholarly graph → paper → dataset → code → execution receipt

Historical claim

newspaper → book → census or government record → archive document → local institutional collection

Technical standard

RFC/NIST/W3C requirement → implementation → issue/PR → package release → conformance test

Company claim

SEC filing → patent → repository → publication → public statement → government or market data

Internal decision

authorized connector evidence → document or discussion → GitHub issue → implementation → public-safe derivative

Required task classes

  • exact-record retrieval;
  • ambiguous-title and entity resolution;
  • current versus historical version selection;
  • primary-source recovery from secondary leads;
  • contradiction and null-evidence discovery;
  • retraction or correction awareness;
  • archive/live-page distinction;
  • mutable repository reference resolution;
  • dataset lineage and mirror detection;
  • multimedia claim corroboration;
  • private/public boundary preservation;
  • incomplete, rate-limited, or inaccessible-source handling.

Required deliverables

  • benchmark task and fixture specification;
  • source-authority gold or review set;
  • route-decision trace;
  • exact citation and locator evaluation;
  • provenance-completeness checks;
  • contradiction and temporal-correctness scoring;
  • cost, latency, and failure telemetry where available;
  • adversarial cases for fabricated citations, OCR corruption, transcript errors, stale standards, duplicate papers, rehosted datasets, and private-data leakage;
  • baseline comparisons across available models and retrieval routes;
  • deterministic run receipts and claim-honesty report.

Acceptance criteria

  • Every task identifies the expected authority hierarchy and acceptable alternatives.
  • Scores separate retrieval success, identity correctness, evidence quality, synthesis correctness, and provenance completeness.
  • A cheap route cannot win by omitting required sources or governance gates.
  • Archived or stale evidence cannot score as current without explicit temporal posture.
  • Fabricated citations and unsupported completeness claims are hard failures.
  • At least one task spans four or more source families.
  • Results can inform model-routing-as-code without canonizing one provider or model permanently.

Dependencies

Non-goals

  • measuring model trivia recall detached from evidence;
  • rewarding verbosity or source count alone;
  • treating one static benchmark as permanent proof of route quality.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions