mirror of
https://github.com/MBombeck/HealthLog.git
synced 2026-11-08 20:42:43 +00:00
Marc-André Bombeck
05cff531e4
The per-metric assessment cards are tightly grounded but their form had converged: every card opened the same way, forced a "one doable step" even when nothing was actionable, and a metric that stayed steady for weeks produced near-identical paragraphs. Squeeze more value from data already computed server-side, without touching the grounding floor or any safety contract. - Surface FDR-surviving cross-metric correlations on the per-metric card. The discovery engine already screens behaviour×outcome pairs with a real Pearson + exact p-value + Benjamini-Hochberg FDR control; that intelligence only reached the period narrative. A new read-only consumer runs the same full-matrix discovery and filters to the pairs that involve the current metric, then feeds the engine's own conservative, descriptive, never-causal interpretation strings into the prompt as grounded relations. Best-effort: a correlation hiccup can never block the generation it only decorates. - Streak-aware repetition. Derive a steady-run length from the graded series (consecutive recent weeks within a band of the user's own baseline) and tell the model when it has already reported "no material change" N times running, so it acknowledges the continuity in one clause and pivots to a different facet instead of restating the same level. - Stop forcing a step; add controlled variety. The closing step is now conditional — when nothing is genuinely actionable the model affirms and names one thing to watch rather than manufacturing the filler that lets platitudes back in. A deterministic per-render variety token (seeded from user + metric + day, never Math.random / Date.now) rotates the opening angle so consecutive cards and days don't read identically, and the phrasing temperature lifts 0.3 → 0.45 while the facts stay pinned by the snapshot and the forbidden-phrase guards. - Surface the computed data strength (n + recency) into the prose prompt so it hedges on thin data instead of guessing what "few" means, matching the UI confidence badge. - Add locale-matched few-shot examples (a grounded assessment + a labelled banned-filler counter-example) to the assessment system prompt, which most helps weaker local-provider models follow the contract. The own-baseline grounding, computed-not-hallucinated stats, schema- enforced output, filler-phrase ban, and correlation-as-association discipline are all preserved and re-pinned by new guard tests. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> (cherry picked from commit c9d04cab1523e0e57a3f50c87c9f88684b01b26f)