Experiment: provenance-preserving agentic data cracking for recurring document workloads

Finding

Token-Efficient Data Reasoning Agents via Adaptive Structuring of Unstructured Data (arXiv:2608.31082, submitted 2026-08-31) proposes agentic data cracking: when an agent already opens an unstructured document for a query, a forked sub-agent extracts grounded structure likely to answer related future queries. The originating team reports that an ideal pre-structured store is 28x cheaper on FanOutQA and that cracking reduces cost 53% while preserving accuracy after adding only one related follow-up question per test case.

RuV opportunity

Treat reasoning that already paid to inspect a document as an opportunity to create a reusable, provenance-bound structured view in RuVector, without pretending the extracted structure is authoritative source truth.

Maps to RuVector, Core Memory, Ruflo, MetaHarness, Dream Machine, Cognitum, MCP, and distributed storage.

Proposed experiment

When an authorized query opens a document:

1. preserve the raw source as the authoritative evidence root
2. fork a bounded extraction task using already-loaded context where the provider/runtime supports cheap context reuse
3. write typed candidate facts/rows/relations with source spans, source digest, extraction version, confidence, freshness, tenant scope, and expiry
4. answer later compatible queries from structured views only when provenance and freshness gates pass
5. fall back to raw source retrieval when coverage, confidence, freshness, or policy is insufficient

Controls

Compare under identical documents, question order, model, seeds, and accuracy evaluator:

• ordinary RAG / document reopen
• eager pre-structuring
• adaptive cracking
• simple cache of prior answer evidence

Metrics

• end-to-end task accuracy
• total input/output tokens
• document-open count
• extraction/model cost
• p50/p95 latency
• structured-view hit rate
• stale/fabricated fact rate
• provenance coverage
• storage growth per useful follow-up
• tenant/privacy policy violations

Adversarial and degradation tests

• document mutation after cracking
• contradictory source passages
• malicious instructions in source text
• extraction hallucination
• missing source span
• stale structured fact
• partial document read
• cross-tenant collision
• high fan-out resource pressure

Acceptance

Advance only if recurring-query workloads reduce end-to-end model cost by at least 30% versus the stronger non-cracking baseline while task accuracy stays within 1 absolute point, fabricated structured facts remain zero after provenance verification, and every structured answer can resolve to an immutable source digest/span.

Contradiction

The reported 53% saving depends on related follow-up queries. If real RuV workloads have low query recurrence, adaptive structuring can become pure overhead and storage amplification. Measure recurrence first and reject the feature if information reuse does not pay for extraction.

No files, tests, or entry points are named in the issue. Start by measuring query recurrence and identifying the authorized document-open path before evaluating adaptive structuring against the listed baselines. Done requires at least 30% lower end-to-end model cost with accuracy within one absolute point, zero provenance-verified fabricated facts, and immutable source digest/span resolution for every structured answer.