How the 2025 analysis gets carried through 2024 and 2025, with the AI step built so you can check it.
Same two methods, same continuum, same definitions. What changes is the reach of the data and the evidence behind every number.
The pipeline will be built and tested before the December release lands, so the update runs within days rather than months.
The purpose codes stay the official narrow measure, the virtual marker the wide descriptive one, and the continuum keeps its five categories.
Where a definition needs sharpening, this group decides the wording.
Two models read every description independently, a stronger one settles the disagreements, and anything still unclear is queued for a person. Last year's report named this as the next step.
The workbook tables are built from your questions, not my guesses. Code and documentation are handed over at the close.
Each record carries a short description written by the donor. That description is the only place housing shows up when the code does not name it.
Every one of them gets read. Not a sample.
16030 housing policy and 16040 low-cost housing. Whatever donors report under them, nothing more. Official, uncontested and narrow.
Each one places the activity on the continuum and writes down its reasoning, from the same definitions, never seeing the other's answer. Where they agree, the label stands.
It sees the description and both answers, decides, and records how sure it is.
Where the arbiter is unsure, the activity goes to expert review. Everything else settles into its category, with in-donor refugee costs always reported separately.
Every decision is stored, so any number traces back to the activities behind it.
Several of you are joining this work for the first time, so here is the shape of it from start to close.
Last year's analysis rebuilt from today's data, so we start from a baseline we can defend.
We are here, starting todayEvery new activity description read by two models, disagreements arbitrated, unclear cases queued.
Second half of OctoberAn expert classifies a sample blind, and you settle the borderline categories.
Late October and NovemberDataset, Excel workbooks, findings report, methodology annex.
NovemberThe new OECD release run through the same pipeline within days.
DecemberFinal package, closing presentation, code and documentation handed over.
JanuaryThree controls, all of them visible to you, and one caution.
Last year's report rebuilt from today's data with last year's labels, so differences get explained rather than argued.
Every data problem written down with what was done about it. No silent fixes.
An expert classifies a sample without seeing the AI labels. The only honest accuracy test.
Figures also move because the OECD revises history and the price base changed. Both are quantified separately.
A decision or a slot in the calendar, not much of your time.
About 200 activities, half a day in total, and it can be split across several people. I send the pack.
About twenty questions for the Excel workbook, so the tables are built around real use.
Should a drainage or municipal infrastructure project that never mentions housing count as housing? Your call.
Your advocacy changed the classification. Code 16040 becomes adequate housing development, and the new wording reads almost exactly like the Global Housing Continuum. It applies from 2027, which raises a question best answered with three views side by side.
The two codes as donors report them now. The official picture.
The new wording applied to the same activities. A simulation, because nothing will be reported under the new codes for years.
The virtual marker across the whole continuum. The widest view, and the reason this research exists.
The manual review in the second half of October, and your questions by the end of the month.