Method
How Uozan reconciles project evidence
Project assurance is mostly reconciliation: does what the reporting says match what the plan, the registers and the minutes say? Uozan automates the mechanical part of that work and shows its evidence, so you can spend your time on the judgement.
1. Extraction, with provenance
Each document is read once per version. Text and tables are extracted along with where they came from: the page of a PDF, the sheet and row of a spreadsheet, the heading and table of a Word document, the slide of a presentation. Files are checked by content, not just extension; password-protected, corrupt and image-only (scanned) PDFs are detected and reported rather than silently skipped. A content hash means an unchanged file is never processed twice.
2. Normalisation
Registers and tables — RAID logs, milestone plans, action trackers, budget sheets, readiness checklists — are mapped column by column into risks, issues, actions, milestones, dependencies, decisions, budget lines and readiness criteria, with owners, dates, statuses, scores and amounts. This step is deterministic code; no model is involved.
Narrative text — status report commentary, meeting minutes — is read by an AI model that identifies the same kinds of items. Every item it proposes must quote the passage it came from, and Uozan keeps it only if that quote, and every date, owner and amount it reports, can be found verbatim in the source. Anything else is discarded.
3. Reconciliation
Items are matched across documents by identifier (R-12, M4, A-07) or by a strong title match of the same type. Near-misses are deliberately not merged. Deterministic rules then compare them:
- different dates for the same milestone or commitment, and how many times a date has moved;
- green or on-track statuses on items whose date has already passed or slipped against baseline;
- items complete in one document and open in another, taking into account which document is newer;
- risks closed or mitigated while their mitigating actions remain open;
- overdue actions, unowned risks, dependencies and actions, decisions past their required date;
- different budget or forecast figures for the same line, and forecast against budget;
- readiness claimed in reporting while readiness criteria are incomplete;
- stale registers, undated documents and missing types of evidence.
A newer document that supersedes an older one is not, on that basis alone, treated as a contradiction — but if it disagrees with the plan of record, the finding explains which source is newer and what to confirm.
4. Prioritisation and explanation
Findings are ranked by severity and evidence quality. Each states what was found, why it matters, the rationale for its severity, the supporting and conflicting evidence with quotes and locations, its limitations, and a suggested next step. Each is labelled by what it rests on: a stated fact, a calculation, a conflict between sources, an inference, or missing evidence.
A delivery-confidence indicator is calculated only when there is enough structured evidence, from four published inputs and weights. It is capped when critical findings are open, rounded to the nearest five, and replaced with an explicit limitation when the evidence is insufficient.
5. Review and repeat
You can acknowledge, resolve or dismiss each finding (with a reason) and add notes; every change is kept in its history. Findings have stable identities, so the next period’s assessment shows what is new, what has gone, what has worsened and which milestones moved.
What Uozan does not do
- It does not see anything you have not uploaded.
- It does not perform critical-path analysis, which needs task-level logic links that reporting documents rarely contain.
- It does not re-score your risks, invent missing figures or fill gaps with estimates.
- It is not an audit, and its reports are not approved by anyone. It is decision support for a professional who verifies before acting.