Skip to content

Project Plan index

Risks

The N-x novelty claims are stated in full in novelty.md; baselines (PC, B5, B-open, …) in experiments.md.

Ordered: validity risks (could make us conclude the wrong thing) first, then mechanism risks, then honest nulls (disappointing but still results), then execution.

Risk Signal Response
False negative on degradation (underpowered demo) — collapse or forgetting No degradation seen under stress Pressure-sweep incl. from-scratch + no-replay + long horizon; collapse on joint-embedding, forgetting on MAE; PC must fire first
Label-free signal can't detect in time (leading-indicator null) Tier-1 signal doesn't precede Tier-2 forgetting, or no usable lead time Oracle-driven control result still stands (decoupled); report which signals are/aren't leading indicators — a negative result of independent value; deployability becomes the open question
Coupling is regime-specific (init) Dynamics present from-scratch, absent pretrained (or vice versa) Report the init-dependence as the finding (the regime where adaptation matters); pretrained bounds realism — don't overclaim deployment relevance from a from-scratch-only effect
Synthetic-correlation artifact (STL-10) Degradation on 1a pilot but not 1b BDD STL-10 validates the apparatus only; the claim rests on BDD's real correlation; non-transfer means class-blocking drove it — say so, lean on BDD
Filter wins on collapse metric but not downstream RankMe up, probe flat Optimized the thermometer, not the patient — validate on downstream + forgetting
Actuator null — F-c can't be steered Varying α in open-loop doesn't move health The controller has no lever → F-c reduces to a static hybrid; gated (see below)
Adaptation doesn't beat frozen (RanDumb) B5 ≥ adapted Existential — reframe to selection-for-efficiency; surface in Phase 1
Closed loop doesn't beat open, with oracle signal (N-E/N-G null) Closed ≈ or < B-open Still a result (closing this loop is hard/unhelpful); keep the monitor as a passive alarm
Closed loop oscillates / destabilizes Health metrics ring Slow the controller; hysteresis; bound the α excursion — the stability finding is itself reportable
Coupling result just confirms DiSF/CCS offline No surprise Hunt loop instability / flip inversion / closed-loop benefit explicitly
Reservoir+dedup / SOFed hard to beat Bake-off flat Fine — criteria were never the contribution; weight stays on the coupling + control loop
FL infra slips / aggregation confounded FedN not ready; can't separate skew from non-IID FL science is parallel and deferrable (doesn't block the spine); the centralized reference isolates selection-induced skew from ordinary non-IID
Compute / Gaudi bottleneck; Phase-1 grid too large Habana port slow; init × pressure × method blows up CPU + small models for the demonstration; isolate Gaudi as a contained task; trim rule: full init×pressure sweep on the STL-10 pilot, carry 1–2 init levels to BDD
Drift back to the criterion bake-off Work optimizing filters, not studying the coupling The Novelty-at-a-glance (N-A…N-G) is the honesty check — if work doesn't map there, it's drift