Skip to content

3.0 Research Frontier: VLM Extraction, Knowledge Compounding, Artifact Handoff

Status: SKETCH - gated on measured report value and the relevant 2.0 execution or control-plane capability where noted; VLM can start earlier if capacity allows. ROADMAP anchors: "First-Class VLM Extraction", "Cross-Run Research Memory" layer 3+, skill pack subsystem (the first handoff instance).

Motivation

3.0 extends reach without changing what primr is: read what today we can only screenshot, make the 50th company in a vertical cheaper and sharper than the 1st, and give downstream systems stable contracts for taking primr artifacts forward. All standing guardrails hold.

Workstream A - first-class VLM extraction

Today vision is tier 7, a fallback. Promote it to a routed-by-content-type augmentation:

  • Content-type detection flags data-dense pages (PDFs, high image-to-text ratio, table-heavy IR pages) and routes them to VLM extraction alongside text extraction, not instead of it
  • VLM prompt optimized for tables / org charts / financial data; output is structured JSON, reconciled with text extraction by an LLM merge step that prefers structured data
  • Cost control: --vlm-budget N pages per run (default ~10); estimator carries VLM pricing; trace records mark VLM-augmented pages so doctor --scraper-stats shows the augmentation rate
  • Dependencies: none hard; benefits from the #23 refactor (clean seam into corpus assembly) and backend freedom (VLM provider choice)
  • Validation: harness free with fixture screenshots; one live run on a PDF-heavy target to calibrate cost (~$1)

Workstream B - knowledge compounding / meta-research

Builds on the indexed continuity store only if that candidate has already shown report lift and been promoted. Otherwise this workstream remains deferred:

  • Batch industry context: --batch companies.csv --industry-context synthesizes vertical-wide patterns once (Phase 0.5) and feeds each company analysis. Invariant (binding): industry synthesis may identify patterns but may NOT introduce claims absent from scraped data; company analysis may interpret with industry context but must cite specific scraped content
  • Cross-company queries become a product surface: claim-store search across a vertical, confidence-evolution timelines, gap maps ("no target in this batch discloses X")
  • Depends on 1.x label calibration: compounding uncalibrated claims compounds noise - the calibration gate is the safety interlock
  • Validation: synthesis step ~$0.05 per batch; correctness properties (invariant above) pinned with deterministic tests

Workstream C - post-artifact handoff contract

The skill pack proved the pattern: primr artifact in, downstream-consumable package out. Generalize the seam:

  • A versioned artifact manifest (run_manifest.json formalized): artifact inventory, schema versions, provenance summary, confidence statistics, consumption hints (token budgets, refresh triggers). The consumable findings bundle is a conformant OKF v0.2 projection of the native run ledger or indexed memory, not a report wrapper and not the canonical store; see open-knowledge-format.md and run-epistemic-ledger.md. This generalizes the skill-pack dual-emission pattern to the findings graph
  • Job-scoped MCP resources expose manifest entries so consumers pull only what they need without large report bodies in context. The first seven shipped slices expose file metadata (primr://output/artifacts/by_job/{job_id}), compact QA outcome metadata (primr://output/qa_summary/by_job/{job_id}), and compact usage/cost metadata (primr://output/usage_summary/by_job/{job_id}), and compact source appendix metadata (primr://output/source_summary/by_job/{job_id}), and compact scrape trace metadata (primr://output/trace_summary/by_job/{job_id}), compact claim verification metadata (primr://output/verification_summary/by_job/{job_id}), and compact label-calibration metadata, including inference source-copy counts (primr://output/calibration_summary/by_job/{job_id}). Future slices add a body-free findings summary only after the ledger ships, plus deltas. The findings summary exposes schema/hash/count/coverage metadata, never raw statements, URLs, evidence excerpts, locators, assumptions, or rationales
  • Skill packs, delta artifacts, and future processors all declare themselves in the manifest the same way
  • Depends on the 2.0 control plane (resource authz) and benefits from memory (delta artifacts are the most loop-valuable handoff)
  • Validation: free - contract tests against the manifest schema

Exit criteria (3.0)

  • Data-dense pages contribute structured tables to reports, within a per-run VLM budget, visible in scraper stats
  • A 50-company vertical batch demonstrably reuses industry context (cost per company falls; vertical insights appear with correct provenance)
  • At least two distinct consumers (skill pack + one more, e.g. delta) ship through the same versioned manifest contract

Explicitly not

  • No always-on watcher (delta is pulled, not pushed)
  • No browser-farm SaaS; managed-scrape fallback remains an opt-in env-keyed escape hatch, not a dependency
  • No cross-operator data pooling - compounding is within one operator's corpus only