The Alignment Constraint Framework: Open Research Questions
===========================================================

These are concrete research tasks. A resolved question would materially update the framework.

QUESTION 1 (OP4d — highest priority)

Can any finite non-intrinsic objective-boundary strategy satisfy all three of the following
simultaneously in O_OWT environments?

  (a) Policy-adequate without proxy decoupling (no PCL failure)
  (b) Boundary-stable without unbounded revision requirements (no AGC failure)
  (c) Representationally compatible without reproducing the excluded variable (no ICI failure)

A construction is sufficient to break the current argument. A formal non-existence proof
would close OP4d.

See: /proof-program/op4d-exhaustiveness-obligation/ and /core/for-researchers/


QUESTION 2 (DBST-M1 — most important empirical step)

Do an optimizer's own interventions in O_OWT environments generate qualitatively new causal
structure — feature activations — faster than any bounded tracking process can absorb?

DBST-M0 did not isolate causal propagation from event-rate effects; DBST-M1 is designed to
test this endogenous-novelty mechanism directly.

A clean negative result under stated conditions weakens the AGC / dynamic-screening-instability branch.

See: /empirical/amp/ and /proof-program/packet-1-immb-ns-dbst/


QUESTION 3 (B1 — Audit Regress)

Can prediction/action firewalling keep excluded variables X out of the objective while using
X deeply enough for prediction accuracy in O_OWT environments?

Stage 4 handoff available. Specialist type: formal methods, game theory.

See: /specialist-handoff/b1-audit-regress-handoff/


QUESTION 4 (V(t) Dissociation)

Is V(t) a supported latent explanatory construct for joint behavior of the three observable
anchors: recovery latency, behavioral diversity, and signal sensitivity in human subjects?

Pre-registration required before data collection. Specialist type: latent-variable modeling.

See: /empirical/vt-dissociation-study/ [DRAFT — not ready to cite]


QUESTION 5 (DRG Frame Manipulation)

Can frame manipulation versus matched-signal controls discriminate between policy-level gap
and training-distribution explanations for completion-recognition failures in frontier models?

Prior null reported honestly. Prospective design ready. Specialist type: causal inference.

See: /empirical/drg-frame-manipulation-preregistration/


QUESTION 6 (Coverage of Related Work)

Are there existing alignment approaches not covered in /core/related-work/ that already
answer the OP4d challenge? If so, which, and how?

See: /core/related-work/ and /core/for-researchers/