Dated milestones for the Framler engine. Per-horizon calibration roll-in, factor activation phases, security and disclosure work. Updated as each phase ships — no aspirational dates, no marketing fluff.
BOCPD posterior + conformal halfwidth + Kalman DLM weights flipped from literature priors to MLE-tuned on realised forward returns. 1d + 7d horizons measured at IC 0.055-0.102 on n=1,091 samples across 3 regimes.
Five engine-moat scalars (overall_raw, coherence, confluence_delta, confluence_tail_alignment, regime amplifier coefficients) stripped from the RSC hydration payload on every ticker page + every public API. Adversaries can no longer reverse-engineer regime amplifier coefficients via View Source / API scraping.
Two new academic-pedigree factors (Corsi 2009 HAR-RV cascade volatility forecast; Bali-Cakici-Whitelaw 2011 implied-volatility risk premium) now populate stock_universe daily with engine weight 0. Composite math unchanged — observation phase runs ~2-3 weeks before measured IC justifies activation.
7-day forward-return calibration window matures (~30 weekday samples post 2026-05-17 flip). Mondrian conformal halfwidth + per-horizon alpha scale on the 7d projection roll from analytical priors to MLE-tuned on measured data.
30-day forward-return calibration window matures (~30 weekday samples post first 30d outcome). Becomes the primary headline forecast horizon — most multi-factor research targets 30d for the cleanest signal-to-noise ratio.
After 2-3 weeks of parallel HAR-RV vs Huber-60 measurement, calibrate the optimal blend weight α via MSE on realised vol. Engine variance prior switches from Huber-only to a blended estimator — sharper intervals on volatile names, identical on stable ones.
After IVRP score column accumulates ~6 weeks of OOS measurement, calibrate-weights cron measures factor IC. If IC > 0.03 (professional-quant band), weight promotes from 0 toward inverse-cov calibrated value. Adds Bali-Cakici-Whitelaw cross-section signal to the composite.
90-day forward-return calibration window matures. All 4 horizons (1d/7d/30d/90d) running on MLE-tuned posterior — analytical priors retire to fallback role.
Sunday 12:00 UTC cron refits factor weights via inverse-covariance + copula-blend across 4 regimes × 4 horizons. Adapts to factor decay without manual retuning.
Every push runs type-check + 783 vitest invariants + knip dead-code scan + lint:quality (blocks Cyrillic / dated / founder-attribution comments) + npm audit. Merge fails on any red. No release without four green gates.
Dates marked "~" are tied to calibration windows — they shift when the relevant horizon accumulates enough realised forward-return samples (~30 weekday observations). Roll-in is automatic via the weekly calibrate-weights cron, not a manual deploy.