P5: Anomaly detection

The issue says the implementation is merged and archived; the remaining full-season tuning is tracked in issue #13. Start by reading that follow-up and the linked OpenSpec change and ADR-0010 amendment. Check the backtest on full 2026 logs against known failures and the false-positive rate; completion means the follow-up's tuning is resolved.

Size: L · Depends on: IgniteRobotics/flashpoint#17, IgniteRobotics/flashpoint-private#69, IgniteRobotics/flashpoint-private#71, IgniteRobotics/flashpoint-private#74 · OpenSpec: 2026-10-12-p5-anomaly-detection

Status

Update 2026-10-19: merged (IgniteRobotics/flashpoint-private#41) and archived (IgniteRobotics/flashpoint-private#42) on 2026-10-12. What's left is [HUMAN] 9.3, tuning on the full 2026 season, tracked in IgniteRobotics/flashpoint#13. The status below is the text from when it was implemented.

Status: implemented 2026-10-11 (change p5-anomaly-detection, PR into rewrite, stacked on the can-health archive PR IgniteRobotics/flashpoint-private#40). Field check and backtest on the full 2026 logs are open ([HUMAN] 9.3, 10.4).

• What shipped:
  • flashpoint detect (also the last step of every derive; skipped when nothing changed) writes gold/anomalies/ with SPEC limits, DRIFT and STEP against each unit's own median/MAD baseline, and SIBLING comparison within the robot config's [[siblings]] groups. MODEL (isolation forest) is optional and off by default (ml extra).
  • Every anomaly carries actual vs expected, confidence, a per-match evidence series with its expected band, a "why", causes from config/anomaly-causes.toml, and "seen this before".
  • The Anomalies view (queue, sensitivity, evidence, triage), a navigation badge, Replay badges (a snapshot in static exports), and anomaly-aware History health.
  • Triage is stored append-only in the ledger: local only, backed up, and kept across rebuilds (ADR-0015).
  • flashpoint backtest scores config/failure-labels.toml (seeded with Johnson Q105 and GACMP Q7; battery-proxy removed GACMP Q7, a wiring drop): caught or missed, lead time, and false positives per method per 10 robot-matches.
• Measured: detect on a 60 × 23 season takes 1.6 s and 222 MB; the backtest on two seasons takes 3.2 s and 256 MB; an uncached Q7 brownout count takes 0.10 s.
• On real data: the Johnson lake flags the Q105 roller reboot (HIGH) and its 820 ms gap (MED), intake stalls, and brownouts in all four matches. The Q105 extension-follower degradation is missed until more 2026 logs give it a baseline (3 of 5 matches).
• Deferred:
  • the Slack/Discord webhook, until the backtest shows a tolerable false-positive rate;
  • NOISE (idle-current jitter), which needs a commanded-off feature in gold;
  • STUMPY discords and River online learners;
  • spin-test and current-limit-count metrics (no data yet);
  • a Driver Station log reader (brownouts come from the battery proxy). Still deferred after battery-proxy.
• Follow-up, change battery-proxy (2026-10-12): the battery proxy is the robot's declared low-current device (the gyro), with a per-20 ms median fallback, so a motor's wiring drop no longer reads as a brownout; the brownout hold is 0.02 s and a logging gap holds a low sample at most 0.1 s. The deferred battery-resistance metric shipped as supply path resistance per motor (mΩ, a baseline metric). See the ADR-0010 amendment.
• Follow-up, change ui-cross-tab-filters (2026-10-12): the Anomalies queue takes the shared season, robot, event, and unit filters, applied by the server (/api/anomalies; an anomaly passes when one of its flagged matches is in the chosen season and event), and can be grouped by device (a unit by serial, or a robot's battery). The navigation badge stays unfiltered. The anomaly detail names the competitions its evidence comes from, and the evidence chart labels each one with a boundary line between them. Merged as PR IgniteRobotics/flashpoint-private#44 and archived 2026-10-13 (specs synced: view-filters new; anomaly-review, lifetime-trends, match-reports).

Scope

• Tier 1, rules:
  • temperature thresholds (port 55 °C avg / 65 °C max);
  • stall seconds (high current at near-zero velocity);
  • CAN dropout counts: done from the hoots by CAN health (above), with no DS log reader. Brownout counts still need the DS log reader (LigerBots, 2026);
  • current residual P95 above a threshold.
• Tier 2, baselines:
  • robust z-score (median/MAD) of each gold feature against that unit's own history. A swapped-in unit starts its own baseline automatically;
  • also compared against the sibling slots (the 4 drive motors should look alike).
• Tier 3, models:
  • IsolationForest (scikit-learn developers, n.d.) or PyOD detectors (Zhao et al., n.d.) on gold feature vectors;
  • River HalfSpaceTrees for online scoring (River developers, n.d.);
  • STUMPY discords on silver current traces to catch odd shapes (STUMPY developers, n.d.).
• Output: an anomalies table, badges in the views, and an optional Slack or Discord webhook.
• Exit:
  • Backtest on the 2025–2026 logs flags at least one known real failure (team pit notes are the ground truth).
  • False-positive rate is low enough that students don't learn to ignore it.

Moved from the rewrite roadmap on 2026-10-19. That doc keeps the dependency graph, migration, quality gates and risks. Status lives here. Author-year citations are in references.md. pitfall N refers to 03-pitfalls-and-performance.md.