# Week ending 2026-08-07 _Auto-generated weekly entry (2026-08-01 to 2026-08-07), appended by `scripts/ops/append_research_journal_local_llm.py` on 8 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 six times. One cycle completed cleanly; the other five ran degraded — that is, they finished and produced output but tripped a quality flag along the way, which is a softer condition than an outright failure. A large batch of backlog items closed, weighted toward research-governance and reliability fixes rather than new empirical results. **Pipeline runs this week** - The full nightly of 7 August 2026 — degraded. - The nightly of 6 August 2026 — degraded. - The nightly of 5 August 2026 — degraded. - The nightly of 4 August 2026 — degraded. - The nightly of 3 August 2026 — degraded. - An off-cycle run of 2 August 2026 — completed. **What I worked on** On the research-governance side I added a trust-tier floor for the signal model's predictors, so a deprecated or low-trust factor is kept out of the traded model, and surfaced that evidence on the signal-promotion sidecar. I also made the factor-evaluation floors cadence-aware, so annually- and quarterly-reported fundamentals are judged against a sample floor scaled to their reporting frequency rather than a daily one, and I resolved the orthogonality residual book to daily-cadence factors only, since a sparse-grain factor had been poisoning the all-books-finite mask. On reliability and tooling: I cut peak memory in the signal-panel loader so a full-history sweep no longer runs out of memory; added a pure Newey–West heteroskedasticity-and-autocorrelation-consistent t-statistic kernel; stood up the public research-publishing workflow and the book's part-and-chapter structure; and fixed a time-zone bug in the filing-lag change-control cooling-off window plus a Unicode console-output crash in its command-line tool. I also disabled two descriptor predictors from the machine-learning signal manifest, and recorded the decision to floor the max-return factor's daily-return denominator at one cent to stop a near-zero adjusted-close price exploding the ratio. Around thirty backlog items closed in total, spanning documentation, tests, and features.