VeCondo
Saved
Methodology

How evidence becomes a fact.

VeCondo is a controlled pipeline, not a black-box summarizer. Candidate observations are extracted, independently verified, deterministically validated, and — only where they pass — published as canonical facts.

01Principles

Four product principles.

01

Facts before opinions

No star ratings, no 'good/bad building' labels, no unsupported scores, no AI-generated purchase recommendations. Show more. Document more. Judge less.

02

No evidence is not proof of no problem

Never convert 'no evidence located' into 'no problem exists.' Missing information is explicit, not hidden.

03

Separate facts, reports, and inference

Every claim carries an evidence class. AI inference never visually resembles verified evidence.

04

Preserve the original source

Original PDF, webpage snapshot, file hash, page images, extracted text, OCR output, exact page citation, processing timestamp, model and version, review status.

02Evidence classes

Every claim carries its class.

Official recordGovernment registry, permit, tribunal decision, court record.
Document verifiedDocument obtained and reviewed; claim read at a specific page.
Multi-source reportedTwo or more independent sources report the same fact.
Single-source reportedOne source reports the fact; no corroboration located.
Unverified public mentionPublic mention located but not yet reviewed or corroborated.
AI inferenceModel-generated inference. Never visually styled as verified evidence.
Official recordDocument verifiedMulti-source reportedSingle-source reportedUnverified public mentionAI inference
03Event-state model

An issue is not a yes/no field.

No evidence locatedHistorical concernInvestigation proposedInvestigation approvedInvestigation in progressRepair proposedRepair approvedFunding approvedPermit issuedWork in progressCompletion reportedCompletion independently documentedSubstantially remediatedPartially remediatedDeficiency reportedPost-remediation issue reportedCurrent active issueInsufficient evidenceStatus unknown

A permit does not prove completion. A discussion does not prove approval. An approval does not prove the work started. A completion claim does not prove the problem never returned.

042+1 verification

Dual-pass extraction, selective third-pass adjudication.

Pass A (primary extraction) and Pass B (independent verification) run on different OCR/model stacks. Pass C runs only on disagreement, failure, poor source quality, validation failure, or high-impact claims — and inspects the original evidence before seeing A and B's competing answers, to reduce anchoring.

A

Primary extraction

Native PDF text, conventional OCR fallback, layout parsing, structured extraction, schema-constrained output.

B

Independent verification

Rendered page image, different vision/OCR method, different model family, independent prompt, independent output.

C

Selective adjudication

Only on disagreement or high risk. Inspects the original evidence before seeing A and B's answers.

H

Human review

A/B/C inconsistent, sensitive claim, allegation against a person, material value impact, privacy risk, structural/safety conclusion, legal risk, no clear citation.

05Deterministic rules

Validation does not rely on AI agreement alone.

  • 01Completion cannot occur before approval.
  • 02Work cannot begin before the relevant approval date unless clearly documented.
  • 03Levy amounts must be positive numeric values.
  • 04Vote percentages must be between 0 and 100.
  • 05A proposal cannot be labelled approved.
  • 06A permit cannot be treated as evidence of completed work.
  • 07“No evidence found” cannot become “no issue.”
  • 08Meeting date and document publication date must remain separate.
  • 09A document covering Tower A cannot automatically update Tower B.