Interference-Gated Dynamic Activation for Task-Agnostic Continual Learning: A Formal-Empirical Audit of Stability, Forgetting, and Failure Regimes

Contribution-and-evidence summary: The manuscript delivers a hybrid formal-empirical audit of interference-gated dynamic activation for task-agnostic continual learning, with executed benchmark artifacts, symbolic checks, negative-result reporting, and explicit claim-scope controls; evidence supports an auditable mixed-result paper rather than a positive-performance paper. Substantive strengths: comparator fairness controls are explicit and enforced, formal assumptions and applicability limits are stated in main text, empirical reporting includes uncertainty/ablation/failure-regime evidence, and figure/table integration is publication-ready with vector graphics and in-text references. Substantive weaknesses: forgetting-improvement evidence for the core mechanism remains mixed (dynamic full is largely parity with static GELU in the executed matrix), one projection-bound symbolic simplification path fails and limits formal confidence, and CORe50 remains descriptor-only so cross-substrate generalization breadth is not yet closed. Real open questions: whether stronger interference windows or replay-policy coupling can produce robust positive Delta_F under matched fairness constraints, and whether a repaired projection-bound argument can remain valid under the empirically relevant regimes. High-leverage improvements: execute one matched CORe50-inclusive rerun (or equivalent declared replacement with strict scope language), pre-register mechanism-sensitive sweeps tied to alignment windows, and either repair or further narrow the failed symbolic path with explicit theorem-to-limit wording. Minor polish: keep dataset-substrate scope wording uniform between abstract/results/limitations and keep theorem-audit caveat language tightly synchronized across formal analysis and conclusion.

- Hybrid modality fit is strong for the user intent: the manuscript combines algorithm/pseudocode, formal statements with assumptions, and executed benchmark evidence rather than forcing a purely theorem-only or purely shallow empirical framing. - Claim-evidence traceability is high: main claims are connected to concrete tables/figures and appendix diagnostics, including explicit negative outcomes and caveats. - Comparator fairness and reproducibility controls are well specified (matched replay/optimizer/schedule/memory accounting) and integrated into interpretation, reducing attribution confounds. - Manuscript quality checks passed for conference style, vector figure usage, caption presence, in-text figure/table references, and main-body page budget before references/appendix.

- Affected claim: hm-claim-1 (forgetting improvement). Missing/contradictory evidence: executed reduced matrix shows near-parity between dynamic full and static GELU with mixed support, so robust positive forgetting improvement is not established. Owning phase: validation_simulation (evidence generation) with writing_presentation/revision (scope framing). Acceptance condition: either produce matched-fairness runs with positive Delta_F and uncertainty support on primary substrate(s), or keep the contribution framed explicitly as a negative/mixed audit outcome across title/abstract/introduction/conclusion. - Affected claim: dmm-claim-1 formal confidence. Missing/contradictory evidence: theorem audit includes a failed projection-bound simplification path, limiting formal closure to local conditional diagnostics. Owning phase: derive_math_methodology with revision integration. Acceptance condition: repair the failed symbolic path with corrected assumptions/derivation and successful audit rerun, or explicitly keep theorem scope non-guarantee and limitations-bounded everywhere claim-facing. - Affected claim: cross-substrate generalization breadth. Missing evidence: CORe50 remains descriptor-only in this iteration, with fallback dataset used for executed third-substrate coverage. Owning phase: validation_simulation with writing_presentation/revision. Acceptance condition: execute CORe50 results-visible runs under matched controls, or keep generalization language restricted to executed substrates with explicit non-equivalence statement.

- Under what measurable alignment-window regimes (strength, persistence, and gate contrast) can interference-gated activation move from parity to reliably positive forgetting deltas without fairness violations? - Can replay-policy and activation-state updates be co-designed so that mechanism attribution remains clean while improving retention beyond static GELU in task-agnostic streams? - What minimal additional assumptions are sufficient to repair the projection-bound formal path without introducing unrealistic smoothness or regime constraints?

- Run one targeted CORe50-inclusive benchmark pass with the same seeds/comparator fairness predicates, and report claim status transition rules before execution. - Add a mechanism-sensitive sweep focused on alignment-window diagnostics (not just global averages) and tie each sweep axis to explicit acceptance/refutation criteria. - Perform a formal patch cycle on the failed projection simplification: restate assumptions, re-run symbolic checks, and update theorem text and appendix matrix in lockstep.