What the prototype is designed to do
- Parse common validation-document formats
- Flag candidate omissions and inconsistencies
- Connect findings to source context
- Record reviewer disposition and rationale
Deterministic review engine · AI-assistance roadmap
CSVDoc Reviewer
Turn validation packages into traceable candidate findings and a structured human decision record—without treating automation as approval.
Verified public implementation
Open prototype, inspectable controls
The public repository currently implements a deterministic, rules-based review engine with explicit parsing, review, disposition, audit, and export stages. It does not send documents to a cloud service or reduce a package to one generic large-language-model prompt.
- Repository status
- Public prototype
- Current engine
- Deterministic review orchestration
- Public demo
- Reviewer-only, instance-local demonstration mode
- Production boundary
- OIDC, durable encrypted storage, qualification, and validation required
01Sanitized draft→02Deterministic analysis→03Human disposition→04Controlled export
What it must never do
- Approve or release a regulated record
- Replace accountable technical review
- Accept production, patient, personal, or confidential data in the public demo
- Turn automated output into a compliance conclusion
Public demonstration boundary
Use sanitized, nonconfidential sample material only.
Do not upload patient information, personal data, employer-confidential information, regulated records, credentials, supplier-confidential documents, proprietary methods, or security-sensitive material. The linked service is external to CSVtoCSA.com and has its own availability and data-handling controls.
Current access note: demonstration identity is unverified, records are instance-local and may be lost, and the service is not a regulated system of record.
Assurance position
AI assistance is a bounded extension—not the current compliance claim
The companion methodology explores source-grounded AI assistance for prioritization and review. Any model, prompt, retrieval, or corpus capability must be introduced behind intended-use limits, evaluation evidence, human override, change control, and monitoring.
Intended use
Decision-support for qualified reviewers working with permitted, sanitized draft material.
Critical evidence
Finding recall, false findings, source correctness, parsing coverage, access, security, and human override behavior.
Lifecycle control
Rules, parsers, interfaces—and any future model, prompt, retrieval, or corpus—trigger proportionate reassessment when changed.