Product pipeline

MapToc has an interactive prototype that demonstrates source inspection, proposed evidence, human review, reviewer rationale, approval history and assembly of a cited program record using prepared sample content. It is not a production deployment. The pipeline below describes the funded build and validation sequence, not a release-date commitment.

Initial build: a reliable evidence item

The prototype establishes the interaction model. The next priority is to implement authorized document handling, shared persistence, access controls and reliable review history around that workflow. A limited program scope makes it possible to examine failure cases before expanding the product.

Pilot development: a usable program record

The next proposed layer includes an evidence inbox, program history, search and reporting support. Pilot teams should test real reporting questions, handovers and evidence retrieval tasks. Feedback should identify where terminology, review responsibilities or missing context make the record difficult to use.

Technical architecture and safeguards

This is a proposed production design, not a description of the public demo. The demo uses prepared sample content and browser-local interactions. Provider, region, retention and security choices will be confirmed before any live pilot.

Secure ingestion

• Short-lived, pre-signed upload URLs send files directly to encrypted storage.
• Future Google Drive and SharePoint connections can use OAuth and designated folders.
• File validation, malware scanning and tenant-specific processing queues limit cross-tenant exposure.
• Original files and source metadata remain attached to every processing job.

Parsing and preprocessing

PDF, DOCX, XLSX and Google Docs are converted into structured text while retaining headings, tables, author and date metadata, page numbers, source locations and document identifiers. Docling, Unstructured or Apache Tika are implementation options, not final commitments.

AI extraction

1. The parser creates a clean, bounded context from the source.
2. The model identifies relevant passages and proposes RESULT, DECISION, RISK, CHANGE, LESSON and QUESTION nodes.
3. A fixed JSON schema requires a source quotation, document reference, date and confidence score.
4. Insufficient evidence produces an uncertainty flag or empty field.

{
  "category": "RISK",
  "summary": "Stock shortages at two facilities",
  "source_quote": "Still having problems with stock at two facilities",
  "document_id": "update_2026_08_28",
  "page": 2,
  "confidence": 0.89
}

Human review and evidence linking

AI proposals remain distinguishable from reviewed records. Review can cover a document summary or a batch of findings, with uncertain or material claims flagged for attention. Reviewers see the claim beside the highlighted source passage and can approve, edit, reject or flag it. Approved items connect to intended outcomes, indicators, activities, locations, dates, decisions, risks, related documents and supporting or contradicting evidence.

Safeguards under evaluation

• Enterprise AI endpoint with contractual exclusion from public model training.
• Zero Data Retention or modified abuse monitoring where approved.
• Encryption in transit and at rest, tenant-level isolation and role-based access.
• Audit logs, configurable retention, regional deployment and customer-managed keys where required.

Integration work

Connections to existing document systems are planned directions rather than current guarantees. The useful test is whether an integration contributes relevant, authorized material without duplicating records or creating new work to maintain access and source links.

API integration design

The planned application API coordinates document processing, review and export. These are proposed responsibilities, not currently published API endpoints.

• Upload and queue: an authenticated request creates a document identifier and a processing job within the organization's workspace. File versions and content hashes help prevent duplicate processing. Job status makes failed or incomplete extraction visible.
• Workspace connectors: approved Google Drive or SharePoint folders would use scoped OAuth access. Change notifications trigger authorized retrieval; periodic reconciliation catches missed updates. Permissions and deletions must be checked as sources change.
• AI service: the backend sends the necessary source passages to the selected enterprise endpoint. It checks the response against the evidence schema and verifies that quoted text exists in the source. Schema validity alone does not establish that the interpretation is correct.
• Review and linking: access-controlled requests record approval, edits and rejection, together with the reviewer, source version and decision time. Suggested relationships between evidence items remain reviewable.
• Use and export: reporting requests assemble approved evidence with source references and review history. Draft proposals remain distinguishable from the reviewed program record.

Credentials stay on the server, and access controls apply to stored files, extracted text, search results and exports. Retention and deletion policies need to cover derived data as well as the original files. The confidence value in the example above is illustrative, not a measured accuracy rate.

Expansion: learning across programs

Portfolio views and cross-program discovery come later. They require sufficiently consistent records, comparable definitions and permissions. Finding similar language is not enough to conclude that an intervention will transfer successfully to another setting.

Stage	Question that determines progress
Initial capture	Can a reviewer approve a useful, accurately attributed item?
Program pilot	Does the record help with real questions at an acceptable effort?
Integrations	Can relevant source material remain accessible and well scoped?
Portfolio learning	Can teams compare lessons without losing context or access controls?

The roadmap can change as these questions are tested. Building more interfaces is not, by itself, evidence that the product is producing value.

Funding context and management reviews

The proposed program layer adds objectives, milestones, grant allocations and periodic expenditure imports. Material variances link to source documents, decisions and approvals. Missing data, mismatched periods and unresolved explanations should be visible.

Progress views and report preparation should reuse the same record. Institutional deployment additionally requires tested permissions, source history, exports and agreed data governance. These are development priorities, not capabilities already operating in this website demo.