What Drafting Does, Step by Step

Draft report runs a fixed pipeline. Knowing the steps tells you where a result came from and where to correct it.

1. Photographs and documents are already processed

Classification and extraction happen at upload, so drafting starts with a classified photograph set and a fact list from every document. Re-drafting does not re-process unchanged evidence; results are cached under your firm's workspace by content.

2. Notes become structured data

The engine turns the field notes, the questionnaire answers, the photograph captions and the document facts into structured records: one observation per building system (description, age basis with its source, condition, notes, what the POC reported, what was not observed, the photographs), a list of deficiencies (system, title, description, significance, remedy, category, life-safety flag, quantity, unit, unit-cost key, source, photographs), a list of reserve items (component, useful-life key, installed year with its source, the observer's RUL opinion if stated, quantity, unit, unit-cost key), plus deviations, limiting conditions, documents not provided and the interview record. This step is where judgment gets encoded; it is instructed to state not observed or not provided rather than invent, to keep transaction items out of the deficiency list, and to attribute every fact.

3. The cost tables are computed, not written

Table 1 is built arithmetically: quantity times the unit cost from your firm's library, rounded to an order of magnitude, the $3,000 threshold applied with the four-items rule and the life-safety exception, immediate and short-term segregated. Table 2 is built the same way with useful lives from the HUD table (Fannie 4099.F alongside), RUL from your opinion when you gave one and otherwise from useful life minus documented age, effective age as useful life minus RUL, year due as report year plus RUL, inflation compounding from the report year. No model writes a number in either table.

4. Structural checks

Before any prose: every referenced photograph must exist, every deficiency needs a remedy and a significance, ages need a source, reserve items need a sourced installed year or a flagged assumption, and lines with a unit cost but no quantity are flagged as engineer to provide. Errors block approval; warnings are shown.

5. Narrative

One section per system is drafted from that system's observation, deficiencies, cost lines and photographs, in the guide's order: description, observations and analysis, reported by the point of contact, not observed, recommendations. Then the executive summary from the sections and the tables. The prose is instructed to cite only what is in the data, never to include photograph ids in text, never to introduce a number the tables do not carry, and never to use language that accuses a party.

6. Prose checks

Numbers in the prose are compared against the data; photograph ids, banned phrases and citations of standards the guide does not reference are flagged. Findings appear in the workspace's Checks panel, with the section they came from.

7. Render

The draft is rendered to HTML and, on demand, to PDF through a headless browser, with a two-pass table of contents so page numbers are real. Until approval the PDF carries a draft-for-review mark.

What it costs

Each draft records its model usage; the job page shows the AI cost per report (typically a few dollars). Re-drafting after an evidence change costs only the steps whose inputs changed.