Skip to content

Research notes

Dated working memos, kept as written. They record what was measured and what was being considered at the time, and they are deliberately not revised as the system changes.

These are snapshots, not maintained pages

A claim in a memo was true when its author wrote it. Where a memo and the rest of the documentation disagree, the documentation is current and the memo is history. For the state of the system today, see Methods and Limitations.

Several of these memos propose work that has since shipped, and several propose work that was subsequently declined — see Bibliography §2.

Memo Date What it is
Evaluation review 2026-07-26 Measured-state review: what the first live week actually showed, code status against the July roadmap, a methods-literature survey mapped to measured gaps, and a sequenced recommendation. Its central finding — the binding constraint is operational, not methodological — still holds.
Improvement methods 2026-07 Literature-driven improvement plan: adaptive diurnal bias correction (EWMA hour-of-day, Kalman, SAMOS) and fixing data starvation (Ensemble API, NBM, sub-24h backfill, neighbour truth, radar minutely precipitation). Much of the first half has shipped — see Grounding.
Week 2 provider diagnostic 2026-08-04 Data-driven diagnostic: horizon census, temp_c and humidity_pct error by lead bin, findings and the actions taken.
Selection memory benchmark 2026-10-07 Fresh-process selection measurements against PR #46's original implementation, with fixture sizes, RSS spread, and runtime.
Score-reader benchmark 2026-10-10 Final source against reviewed head 6a4933f: both selection fixtures, populated dashboard and full report, three decoder policies, immutable inputs, raw samples, and payload checks.
Score fallback and report reuse 2026-10-11 Correctness and report-summary reuse against 4019cc6: 24 → 12 full load calls under mcs, equivalent report tables/charts, local calibration warnings, and retained final-wealth readback under seq_mcs.
Fitting diagnostics benchmark 2026-10-12 Isolated diagnostic collection and real fitting against merged PR #54: three fresh processes, shared thread settings, equivalent point/method/release/quantile outputs, and no diagnostic scans on successful fits.
Verified fitting diagnostics benchmark 2026-10-13 Repeats the original comparison with verified module origins before timing, three fresh processes per workload/size/variant, equivalent forecast payloads, and current promoted-failure retention.
Cross-cutting summary — Synthesis across the three: trade-offs between method families, pushback specific to this project, and what to decide up front.