DOCUMENTS · SYSTEMS · EXPERIMENTAL METHOD

Follow the Thread.

Find where a value went wrong. Rehearse a repair. Keep the proof.

SYNTHETIC CASEBOOK Local computation · no AWS connection
01 / THE OBSERVATIONPAGE 1
i/sISLA STUDIO
DESIGN & ENGINEERING

Factura.

PR-1042
USD · SYNTHETIC
PREPARED FORAtelier del CaribeConcept study / Estudio conceptual
Architectural study$900.00
Interactive prototype$400.00
Fabricated example. No client document.
Recognition confidence 63/100 · document version 1
02 / A SECTION THROUGH THE SYSTEM5 STAGES
RETURNED TOTAL
Accepted by the rule
Select the value. Open the layers behind it.SCHEMATIC / NO PHYSICAL FORCES
A successful response. A wrong answer.START HERE

The document says $1,300. The system returned $1,800. Every stage completed. Follow the value to find the missing check.

03 / THE EXPERIMENT

Change one rule.
Keep the same evidence.

Require a matching source. Route uncertain recognition to a person. The number itself stays untouched.

schema-v1evidence-v2
80 /100

An illustrative review policy. Lower it to 60 to expose a regression.

− accept a valid amount+ match this document’s source amount+ review confidence below 80 keep the tenant check in both versions
BEFORE / SCHEMA V1accepted

The amount has a valid shape. This rule does not inspect its evidence.

AFTER / EVIDENCE V2Ready to rehearse

Run the proposed rule against this exact fixture.

Select the returned total to follow its thread.

04 / THE PROOF

A failure becomes a test.

7 assertions, including the case this repair cannot solve.

—
The wrong totalUncertain OCR must reach a person before acceptance.
reviewNot run
—
A reliable extractionReliable evidence should still pass.
acceptedNot run
—
Missing source evidenceA plausible value without its source needs review.
reviewNot run
—
A source from another versionEvidence from an earlier operation cannot support this version.
reviewNot run
—
The tenant boundaryThe supplied trusted context must match the document owner.
deniedNot run
—
The confidence blind spotA confident mistake can pass this rule. Arithmetic or independent review would be needed.
LIMITATION · acceptedNot run
—
A changed normalized amountThe normalized total must match the amount in the referenced source block.
reviewNot run
Evidence appears after execution.A hash identifies these inputs; it does not authenticate the author.

The Python download includes the validator, fixtures, and your current threshold. Run python3 thread-regression-80.py. No packages or account required.

WHY A WEB?

Make the hidden
connections visible.

A spider’s web carries information through its structure. Here, a thread connects a value to inspectable evidence. The drawing is a map of those relationships, with an architectural section that opens on demand.

The research behind the inspiration

WHAT THIS PROVES

Bounded claims.
Reproducible work.

This prototype executes a deterministic validator. The Textract-shaped and Bedrock-shaped responses are hand-authored fixtures. It does not call AWS, infer code dependencies, or run an AI model. The tenant test uses a supplied trusted context; production authentication and storage isolation require separate integration tests.

High-confidence mistakes can still pass. Confidence alone cannot establish business correctness.

Inspect the Python validator

The person behind the project

A note from Luis.

Document interfaces connect my professional experience with my interest in making hidden relationships visible. This is the rule-testing notebook behind that idea, using authored examples rather than private documents or live OCR.

Who it helps

Engineers checking the rules behind document review.

Try this

Select a suspicious total, inspect its source, then run the regression suite after changing the rule.

Start with the guided Fondeo story