Document-State Twins Back to Sovrient
ANATOP For Documents

The agent-native document layer for consequence-bearing PDFs.

Sovrient does not replace the PDF. It binds a machine-discoverable, provenance-bound, verification-ready surface to the authoritative human artifact so autonomous agents can consume document truth with low ambiguity.

Traditional digital twins simulate physical systems for prediction. ANATOP document-state twins do something different: they make consequence-bearing documents operable by autonomous agents where raw PDF consumption is structurally weak.

Open Live Reference Package Machine State Agent Manifest

What This Changes

Machine-Readable Is Not Enough

Raw PDFs can be ingested, but they do not reliably preserve table topology, stable claims, or low-ambiguity authority rules for agents.

Machine-Legible Is Still Not Enough

Tagged structure helps, but agents still need discovery, provenance, fallback behavior, and explicit conflict resolution to operate safely.

Agent-Native Is the Next Layer

The document-state twin adds declared inputs, typed surfaces, release gates, and PDF-first authority so the document can participate in verification chains.

Not A Digital Twin

Traditional Digital Twin

Replicates a physical object or system so teams can simulate scenarios and predict outcomes.

  • Replicates turbines, factories, supply chains, or cities
  • Primary question: what if this physical state changes?
  • Main buyer: engineering, IoT, operations
  • Main risk: modeling the physical system incorrectly

ANATOP Document-State Twin

Binds a deterministic machine surface to an authoritative human document so agents can verify and act with low ambiguity.

  • Binds contracts, audits, diligence reports, and consequence-bearing PDFs
  • Primary question: what does this document actually authorize or prove?
  • Main buyer: legal, risk, compliance, procurement, structured finance
  • Main risk: document opacity, extraction drift, and authority ambiguity

Live Reference Implementation

The first public reference package binds a six-page strategy PDF to a released machine surface with manifest, claims, citations, tables, pages, state receipt, attestation, normalization rules, and a table-authority report.

Reference Package Open the live document-state twin Human route for the public reference implementation. Manifest Machine discovery and declared inputs Document identity, input set, authority rule, and output map. Authority Gate Table authority report Release-blocking check that the public structured tables did not drift from the PDF authority surface. Benchmark Raw PDF vs plain text vs ANATOP twin Six fixed tasks showing why the machine twin exists.

Adjacent Artifact Class

The same ANATOP logic now extends beyond PDFs into consequence-bearing images. The first image-state twin binds a Sentinel-2 SWIR ratio heatmap to declared witness, region, and visual-summary sidecars so agents can use the image within a bounded advisory scope.

Image-State Twin Open the gold alteration heatmap twin Source image + claims + regions + field-level authority resolution. Manifest Declared image input set Visual pixels, region geometry, metrics, and advisory boundaries resolved by input role.

Three-Operator Consensus

Operator Result Shared Remaining Fail Key Confirmation
Claude Opus 10 PASS / 0 PARTIAL / 1 FAIL Raw PDF baseline Document-state twin is real; authority model is coherent.
ChatGPT 5.4 10 PASS / 0 PARTIAL / 1 FAIL Raw PDF baseline Previous table-drift objection cleared after PDF-faithful table release and authority gate.
Gemini / Antigravity 10 PASS / 0 PARTIAL / 1 FAIL Raw PDF baseline JSON and HTML surfaces resolve the structural failure of PDF-only extraction.

Current Boundaries

PDF remains authoritative Release-blocking table authority gate Public discovery via manifest and machine state Public replay still future work Hostile larger PDF testing still future work Anchors are page-level, not bbox-level

The current claim is intentionally bounded: this is a live, public, agent-native document-layer reference implementation. It is not yet a cross-issuer standard, and it still needs replay and hostile-document expansion work.