Skip to content

TODO.md

Latest commit

 

History

History
114 lines (66 loc) · 8.08 KB

TODO.md

File metadata and controls

114 lines (66 loc) · 8.08 KB

TODO — Open Work Items

Canonical source: msc/analysis-2026-04-02-comprehensive.md for full context and rationale. This file is the action list.

Codex Review Findings (2026-04-02, round 2)

  1. HIGH — Strategy-persistence stated for wrong mismatch object. FIXED. Rewrote Formal Expression in #strategy-persistence-schema to be explicit: schema applies to any mismatch state; what's proved is persistence of δ_s (plan-confidence error), not δ_strategic (calibration residual). Persistence of δ_strategic remains open (requires credit-assignment machinery).

  2. HIGH — Φ called "true plan success probability" but isn't under correlated failure. FIXED. Changed language in #strategic-dynamics-derivation B.5: Φ = P_Σ(θ) is now "independence-model reference value," not "true success probability." Added explicit note that δ_s tracks calibration within the independence model, not calibration to reality.

  3. HIGH — Composition bridge still open. Already honest in #composition-closure (line 161: "additional assumption beyond (A4)") and #tempo-composition (sketch status). Codex recommends presenting as "AAD plus contraction premise," not as scale invariance discharged. Action: Audit prose in OUTLINE.md and README for overstatement. No segment fix needed — the segments are correct.

  4. MEDIUM — value-object "depends on M_t alone" hides O_t dependence. FIXED. Added qualifier: "as a state variable" — O_t enters as a fixed parameter, same as π_cont and N_h.

  5. MEDIUM — Graph-structure-uniqueness overclaims. P3→Markov is conditional, not forced. Already labeled correctly in segment; language audit needed for consistency elsewhere. Covered by the "forced vs strongly motivated" audit item below.

Codex Open Questions (answers needed)

  • Canonical strategic mismatch: δ_s or δ_strategic? Answer: δ_s is proved; δ_strategic is the orient-cascade's operational target but persistence for it is open. Both are legitimate; they measure different things. Document this distinction prominently. DONE in #strategy-persistence-schema.

  • Is P̂_Σ calibrated probability or tractable heuristic? Answer: tractable heuristic (independence-model surrogate). B.5 proves calibration of the surrogate, not calibration to reality. DONE in #strategic-dynamics-derivation.

  • Bridge lemma contraction: derive from A4 or make explicit assumption? Answer: currently an explicit additional assumption. Deriving it from A4 is the open research problem. No action needed — already correctly scoped.

  • Satisfaction-gap/control-regret: intrinsic or convention-relative? Answer: convention-relative. They are diagnostics relative to a chosen continuation convention and scalarization. The convention is part of the measurement. Already documented in #value-object and #satisfaction-gap but could be more prominent. Action: Add brief note to both segments' Epistemic Status.

Exploit/Explore/Deliberate (in progress)

  • Three-way tradeoff gap. Segment written. Deep adversarial spike in progress (simulation + first-principles attack). Segment may be substantially rewritten based on spike findings. Current assessment: two-stage decomposition and δ_regret ceiling are genuine; additive objective form and ΔV_Σ approximation are hand-waving.

MEDIUM — Extensions and Refinements

  • Composition scaling with N. Whether closure defect scales polynomially or exponentially with team size. Critical for applying the theory to large teams. No spike done.

  • Multi-timescale stability. #multi-timescale-stability is a sketch; #temporal-nesting leans on it. Needs formal N-timescale singular perturbation treatment.

  • Channel independence caveat propagation. DONE. Caveats added to persistence-condition, adversarial-tempo-advantage, tempo-composition. team-persistence already had it.

  • Edge-independence scope note. DONE. Cross-cutting subsection added to strategy-dag Discussion covering buys/costs/mitigation with cross-references.

  • Communication-gain adversarial scope. Additive model fails for deception (trust is game-theoretic). Either extend or add explicit scope limitation to #communication-gain.

Missing Segments (narrative completeness)

AAD Core (01-aad-core/)

Slug Section Type Description
exploit-explore-deliberate II Written Three-way allocation — under adversarial spike
linear-ode-approximation A Written Pedagogical linear mismatch ODE
(new: adversarial-edge-targeting) III Derived? Which strategy edges most valuable to attack — the Section III gap
(new: intent-dag-development) A Aside Convergence history of AND/OR + single-p (archaeological record)
(new: prior-art-positioning) A Detail Active inference, POMDP, BDI positioning. Source: msc/02-prior-art-assessment.md

Cross-component (needed for AAD scope claims)

Slug Component Relevance
#ai-agent-as-act-agent 03-logogenic Validates Section II for Class 2 agents
#section-ii-survival 03-logogenic Which Section II results survive without directed separation
#coupled-update-dynamics 03-logogenic The coupled formulation directed-separation defers to
#developer-as-act-agent 02-tst-core Validates Section II for human agents
#causal-discovery-from-git 02-tst-core Validates CIY and loop-interventional-access for software domain

Presentation

  • Three-way presentation split. All reviewers recommend: (a) core results, (b) conditional architecture, (c) empirical programs. Single highest-leverage presentation change.

  • Prior art positioning. Active inference/FEP, POMDP, BDI relationships. Source: msc/02-prior-art-assessment.md.

  • "Forced" vs "strongly motivated" language. DONE. Three-tier classification (proved/conditional/choice) now consistent across graph-structure-uniqueness, strategy-dag, and-or-scope.

  • Composition presentation. DONE. composition-consistency and WORKBENCH fixed to flag contraction assumption explicitly.

Promotion Pipeline

30 segments at deps-verified (Gate 1 complete for batches 1+2).

Next: Gate 2 (claims-verified) in topological order. Start with the strongest candidates: sector-condition-derivation, recursive-update-derivation, mismatch-decomposition, chain-confidence-decay, persistence-condition, gain-sector-bridge, worked-example-kalman, discrete-sector-condition, satisfaction-gap, control-regret, graph-structure-uniqueness.

Lower Priority

  • Observability-dominance product formula. conf_obs = conf * obs posited but not derived. Label as formulation choice or derive.
  • Strategy-complexity-cost IB operationalization. I(Sigma_t; pi* | M_t) undefined in practice.
  • Strategic calibration aggregation. L2 norm unjustified. Label as design choice.
  • Scope architecture. "Within AAD's scope" ambiguous between adaptive and agency scope.
  • loop-interventional-access status. "exact" defensible; opening claim could be softened.
  • Between-event dynamics. g_M(M_tau) defined but unreferenced. Important for logogenic agents.
  • Fully coupled adversarial dynamics. Both agents' mismatch co-evolving. Open.
  • objective-functional labeling. "axiomatic" for scalar-comparability is a formulation choice.
  • information-bottleneck orphaned. No downstream segment uses IB objective formally.
  • Heavy-tailed disturbances. Model S assumes finite second moment.
  • satisfaction-gap/control-regret convention-dependence. "exact" but convention-relative diagnostics. Add note to Epistemic Status.
  • External validation design. Testable predictions not yet tested. Candidates: git data, RL bandits, adaptive controllers.

Project Structure (Deferred)

  • Root-level assembly index (when content beyond AAD warrants it)
  • framework/ directory for non-mathematical content
  • Multiple index support (paper, preprint, monograph)
  • Section IV standalone paper outline (draft at msc/2026-03-14-section-iv-paper-outline.md)

Tooling (Deferred)

  • Lint-md directory arguments