# Week ending 2026-08-14

<!-- LEMONADE:RUN-SNAPSHOT:START -->
_Auto-generated weekly entry (2026-08-08 to 2026-08-14), appended by `scripts/ops/append_research_journal_local_llm.py` on 15 August 2026. Only the content between the `LEMONADE:RUN-SNAPSHOT` markers is regenerated; anything I add outside them is kept._

This week the pipeline ran seven times: two nightlies failed outright, four ran degraded (finished but tripped a quality flag), and one off-cycle run completed cleanly. Around thirty backlog items closed, weighted toward universe-admission fixes and a signal-decay investigation.

**Pipeline runs this week**

- The full nightly of 14 August 2026 — degraded.
- The nightly of 13 August 2026 — degraded.
- The nightly of 12 August 2026 — degraded.
- The nightly of 11 August 2026 — failed.
- Two off-cycle runs of 11 August 2026 — one completed, one failed.
- The nightly of 10 August 2026 — degraded.

**What I worked on**

The main empirical thread was a signal-decay investigation: the machine-learning signal book's out-of-sample information coefficient had collapsed — to 0.0775 on the run of 5 August 2026 — alongside an R-squared collapse, so I ran a shared root-cause diagnostic, shipped the producer fix with a paired regression, and added a fail-closed decay gate (default off, report-only for now) so a future inversion blocks rather than trades silently.

On universe admission I fixed several instrument-classification defects: multi-class primary selection was picking notes and preferreds over the real common share, dropping genuine index constituents, so I made the tiebreak instrument-type-aware; I resolver-rescued bare warrant arms; graduated the instrument-type allowlist enforcement to on by default; and admitted foreign-domiciled, US-listed regulatory filers except those on a China/Hong Kong deny-list, with the currency-scaling and reporting-cadence caveats documented.

On factors and reliability: I activated two candidate factors — an industry-momentum factor and a 21-day maximum-return factor (the latter displacing 30-day volatility in a bake-off) — after guarding the max-return daily-return denominator against near-zero-price explosions; made the out-of-band demotion predicate's trailing-coverage guard cadence-aware; backfilled the FINRA short-interest history to clear a factor-diagnostics insufficient-history flag; added a memory-store backup tier and per-factor memory profiling to the operator backfill lift; and opened a design brief for a research-screening product and one for remediating the fundamentals filing-lag knowledge date.
<!-- LEMONADE:RUN-SNAPSHOT:END -->
