ProgrammeWorld Model.
Understanding how complex programmes change, with a traceable path from the source report to the model’s view of the world.
Research prototype · Evidence pipeline and synthetic kernel
Pilot snapshot · 9 September 2026 · One World Bank project, P126507. These are engineering and coverage measures, not predictive accuracy scores.
From reports
to programme histories.
Programme delivery leaves a fragmented record: implementation reports, revised dates, ratings and decisions. Before a model can learn from that record, we need to establish what was reported, which entity it concerns, and when it could have been known.
This project begins there. It preserves original documents, extracts cited observations and reconstructs differences between available reports. The longer-term aim is to explore dependencies, test conditional scenarios and update programme state as evidence changes.
Three foundations.
Every observation has a source.
Original bytes and hashes, report identities, exact excerpts and PDF-page references. Missing information remains unknown.
Publication is not the event.
Report dates, disclosure metadata and retrieval times stay separate. Historical availability requires its own verification.
Agreement is not ground truth.
Versioned evidence, independent-review packets and explicit reconciliation. Observations and synthetic assumptions remain distinguishable.
A finding that
changes the question.
In the pilot, a project summary’s revised closing date differs from the closing date of its IDA credit. The credit and a trust fund have separate dates. A changed summary cell therefore cannot automatically become a label for a legal extension of the same loan.
This is why entity scope and source review come before model training.
What exists today.
A reproducible public-report evidence pipeline, a versioned SQLite index, resumable acquisition, review and reconciliation records, backup recovery, and automated checks on Python 3.11 and 3.14.
A separate synthetic simulation kernel supports testing explicitly assumed mechanisms. It is not empirically calibrated.
The pilot contains 121 reported values and 14 explicit unknowns. Sequence 7 is missing from the acquired series. Independent extraction accuracy, historical source versions and financing-specific outcome definitions remain open.
The research path.
There is no trained programme predictor yet. Future datasets and experiments will be shared with their provenance, evaluation and limitations when ready. This Space is a project overview, not an interactive forecasting tool.