Skip to content

Incoming-invoice adaptability proving ground ​

PROVEN_LOCAL_SYNTHETIC_POC_NARROW_GO

Public delivery is tracked by epic #360 and sequential issues #361, #362, #363, #364, #365, and #366.

The incoming-invoice proving ground is designed to test whether one capability model can derive different process depths from evidence, risk, approval, Authority, and segregation requirements without treating company size as a workflow rule.

Intended effect ​

text
Source
→ Document
→ Synthetic extraction scoring harness (real Document-AI pilot: #378)
→ Validation
→ Supplier / PO / Receipt / Invoice matching
→ Exception and evidence-citing advisor dialogue
→ Adaptive UI
→ Receipt and evidence verdict

The process is parameterized through three scenario packs:

ScenarioIntended use
LEANMinimum justified controls for a bounded, explainable flow
CONTROLLEDMatching, tolerances, exceptions, and stronger evidence requirements
SEGREGATED_ENTERPRISESeparated roles, approvals, escalation, and audit evidence

These are process-complexity vectors, not company-size stereotypes.

Setup-agent dialogue and variant proof ​

The short-term proof adds a bounded setup-agent dialogue to the pipeline. It clarifies an altered requirement, emits a versioned configuration delta, and reuses existing capability and process variants; it does not invent new product functions.

text
requirement
→ clarification dialogue
→ versioned configuration delta
→ reused capabilities
→ execution / readback / reuse receipt

Compositions ​

Two valid compositions share the same core:

  • ERV Capability Core (self-contained): produces a local decision and evidence package from the requirement, delta, reused capabilities, and readback.
  • ERP-enhanced (optional): adds authoritative ERP reads, downstream posting delivery, and system-of-record readback on top of the core.

ERP integration adds value but is not a prerequisite for the core proof, and productive posting remains separately authorized and separately evidenced.

Baseline vs adapted variant ​

A baseline variant and a changed variant must preserve the exact core and module digests while the changed variant produces oracle-predicted, configuration-specific behavior and readback. The delta alone explains the difference.

Falsifiers ​

Core mutation, substituting an answer, omitting the delta, inventing an unsupported function or requirement, or hiding Authority are falsifiers. Any of them fails the variant proof.

Proof packages ​

  1. AP-01 — Blueprint and scenario vectors Freeze eight layers from source to proof, actors, outcomes, risks, falsifiers, scope, and scenario derivation.
  2. AP-02 — Supplier-invoice intake Bind synthetic source bytes, provenance, document version, digest, metadata, and supplier-invoice identity; deny duplicate, tampered, unsupported, and ambiguous documents.
  3. AP-03 — Synthetic extraction scoring harness Compare a deterministic baseline and a bounded synthetic model proposal against a frozen local-synthetic structured-text holdout covering layouts, line items, taxes, totals, and failures. AP-03 is a deterministic synthetic extraction scoring harness: it does not execute OCR or Document-AI model inference and makes no OCR or model-quality claim. A separate real Document-AI pilot (pinned model/OCR runtime, realistic sealed holdout, independent oracle) is tracked under #378 and remains open.
  4. AP-04 — Validation, matching, and advisor Keep supplier, purchase order, receipt, and invoice evidence distinct; version two-/three-way matching and tolerance variants; preserve exceptions and missing context. The additive versioned relational hardening slice (v2) keeps the historical v1 pack byte-identical and distinguishes amount-level matching from general invoice/PO/receipt matching: versioned amount-level modes verify only the match amounts and return UNKNOWN — never MATCHED — while a supplier, quantity, unit, or currency relation is not enforced by that mode, and relational modes return CONFLICT on a wrong relation at equal amounts. Duplicate reference kinds are denied deterministically and order-independently (no last-write-wins). The evidence package records four distinct dimensions — integrity, origin, semantics, and runtime observation — and repeating the local compiler is deterministic replay of frozen local-synthetic inputs, not an external or system-of-record readback. The standalone authority-free core remains useful without ERP or posting.
  5. AP-05 — Adaptive UI and Application Guide Derive fields and actions from scenario and evidence state while keeping all variants explainable and testable.
  6. AP-06 — End-to-end verdict Exercise positive, duplicate, tamper, mismatch, UNKNOWN, cancellation, and replay cases and emit GO, NARROW_GO, or FALSIFIED_WITH_EVIDENCE. Each released chain layer is reported with an explicit evidence class — SOURCE_BOUND (bytes bound by digest only), SCHEMA_VALIDATED (released contract shape validated) or EXECUTED (the released implementation actually ran over the bound input). A layer that was not executed is never reported as executed. The successor probe v2 additionally carries the bounded synthetic extraction-to-ERV-decision dataflow, so the EXTRACTION layer is execution evidence over the ingested document rather than only a frozen benchmark/holdout binding.

Current status ​

ItemCurrent state
Work-package specifications6/6 frozen and executed
Acceptance identifiers34/34 preserved (28 + 6 exercised)
Scenario packs3/3 defined
Public AP implementation issues#361–#366: 6/6 closed/completed at the 2026-10-01 readback; broader product and real-pilot criteria remain separate
Product implementationBounded local-synthetic proof released; general product not started
Product releasePublic local-synthetic PoC release (AP-06); general product none

Preparation artifacts may contain Blueprints, pseudocode, tentative paths, tests, gaps, and UNKNOWNs. They are not implementation or compatibility evidence.

The 34/34 count adds six acceptance identifiers—four on AP-05 (AP-05-AC05–AP-05-AC08) and two on AP-06 (AP-06-AC06–AP-06-AC07)— exercised by the released AP-05 receipt manifest and the AP-06 proof probe.

The current AP-06 proof is the evidence-class successor verification/incoming-invoice-ap06-proof-probe-v2.json (npm run incoming-invoice-ap06-proof-probe-v2:test): it re-binds the released predecessor bytes unchanged, classifies every chain layer as SOURCE_BOUND, SCHEMA_VALIDATED or EXECUTED, and carries the bounded synthetic extraction-to-ERV-decision dataflow. The frozen verification/incoming-invoice-ap06-proof-probe-v1.json is retained byte-for-byte and explicitly scoped to its historical release line; its generic exercised flag is superseded and is not reused by any current layer claim.

Dependencies and promotion ​

The complete proof depends on separately evidenced Document/Version plus Supplier, Purchase Order, Receipt, and Matching capabilities. Preparation may proceed with named gaps; product promotion must revalidate current Main, dependencies, sources, holdout boundaries, contracts, paths, and tests.

The variant proof replays on Current Main after the #372 / AP-03 terminal state, updating presentation tests, the verification DAG, and checksums in dependency order.

Standalone ERV package extensions are required and tracked separately: #374 and #375. An optional BI composition, KaleidoSphere #157, does not gate core use.

Any productive allocation or posting remains separately governed. The PoC uses synthetic data and does not claim production accounts-payable fitness, universal extraction quality, tax/accounting completeness, or autonomous booking Authority. This marker is removed now that a public GO or NARROW_GO verdict satisfying AP-05-AC05–AP-05-AC08 and AP-06-AC06–AP-06-AC07 is carried by the public AP-06 local-synthetic PoC release: the public verdict is NARROW_GO, the released capability chain is proven, and the requested 200-bps three-way variant has no released executable variant and is reported as typed UNKNOWN, not invented. The general end-to-end product is not delivered.

Return to the PanSphaira overview or inspect the Capability Matrix.

Local synthetic proof of concept — not a production release or security certification.