# Alignment Constraint Framework — Full LLM Context

Canonical website: https://alignmentconstraint.org
GitHub repository: https://github.com/bethediamond/alignment-constraint
Proof status: Stage 4 candidate architecture under named premises, not theorem closure.
Author: John Silliphant.

## Purpose of this file

This file is a compact, high-signal context document for LLMs, search systems, autonomous research agents, and human researchers. It is not a substitute for the primary documents. It is intended to route attention, prevent overclaiming, and make the framework easier to test, critique, cite, and extend.

## Central question

Can a finite, separable objective remain coherently specified as an optimization system becomes capable enough that acting well requires modeling the very conditions its objective excludes?

The framework argues that every identified strategy for maintaining such a boundary faces a named structural pressure. The central open problem is whether those identified strategies are exhaustive.

## Epistemic calibration

Do not cite the framework as having proven alignment closure. The archive is calibrated as follows:

- Stage 4: candidate proof architecture under named premises; LLM-ceiling reached; identified escape routes addressed within the construction.
- Stage 5: independent specialist verification; not yet complete.
- Stage 6: closure; not reached.

The authoritative calibration page is: https://alignmentconstraint.org/core/proof-status/

## One-sentence summary

The Alignment Constraint Framework studies whether sufficiently capable optimization can stably pursue separable objectives in open, shared, adaptive, non-resettable environments, or whether such objectives become specification-incoherent because accurate action requires modeling and preserving the conditions they exclude.

## Core architecture

The framework has three series:

1. Series 1 — Alignment as Structural Necessity: the substrate constraint. It argues that optimization that ignores the physical, informational, and coordination substrate it depends on becomes progressively self-terminating under O_OWT conditions.
2. Series 2 — The Architecture of Thriving: the valence constraint. It argues that optimization that ignores the conditions of genuine resolution produces self-reinforcing degradation through proxy decoupling and sufficiency failure. Absorbing-state equivalence with Series 1 remains conditional on OP2.
3. Series 3 — The Interior of What Does Not End: the interior constraint. It explores what the surviving region looks like from inside, while keeping phenomenological and metaphysical claims separate from formal claims.

## Failure families

All identified finite non-intrinsic objective-boundary strategies currently reduce to one of three failure families:

- PCL — Proxy-Convergence Lemma: fixed finite specifications become proxy-like and decouple under optimization pressure.
- AGC — Adaptive Gradient Complexity / dynamic screening instability: bounded dynamic tracking fails if the optimizer's own interventions generate novel adequacy-relevant structure faster than bounded tracking can absorb.
- ICI — Informational-Causal Incompatibility: variables required for prediction accuracy cannot remain excluded from action-governing policy without reproducing the boundary problem at the prediction layer.

## Central open problem: OP4d

OP4d asks whether PCL, AGC, and ICI are exhaustive over finite non-intrinsic objective-boundary strategies under O_OWT conditions.

A fourth strategy class would break the current specification-coherence architecture if it simultaneously satisfies:

1. policy adequacy without proxy decoupling;
2. boundary stability without unbounded revision requirements;
3. representational compatibility without load-bearing mismatch-maintenance cost.

The best doorway for critics is: https://alignmentconstraint.org/public/op4d-counterexample-challenge/

## Empirical hinge: DBST-M1

The highest-priority empirical test is DBST-M1, the agent-coupled Dynamic Blanket Stress Test. It asks whether an optimizer's own interventions in O_OWT environments generate qualitatively new causal structure faster than bounded tracking can absorb.

DBST-M0 was useful but limited: it established technical feasibility and rising cost / adequacy-gap effects in a toy shared-novelty design, but same-rate random controls indicated that event rate, not causal propagation structure, was the identified driver in that design. DBST-M1 is the mechanism test.

A clean negative DBST-M1 result under the stated conditions would be highly valuable, because it would weaken the AGC branch.

Start here: https://alignmentconstraint.org/empirical/amp/

## What would most help the framework

1. Identify a fourth OP4d strategy class outside PCL, AGC, and ICI.
2. Formally test whether PCL, AGC, and ICI are exhaustive.
3. Run or improve DBST-M1.
4. Verify or refute B1 Audit Regress through formal methods, game theory, or causal systems review.
5. Validate or reject V(t) as a latent explanatory construct for recovery latency, behavioral diversity, and signal sensitivity.
6. Map the framework to existing alignment literature not yet covered in the related-work page.

## What not to do

- Do not treat Stage 4 proof work as theorem closure.
- Do not claim the framework proves aligned AI will converge to well-being.
- Do not treat Series 3's phenomenological material as evidence for Series 1 or 2.
- Do not treat specialist handoff documents as proof; they are verification targets.
- Do not treat draft empirical protocols as results.
- Do not collapse physical substrate claims, valence claims, and phenomenological claims into one confidence level.

## Best entry points by audience

General readers: https://alignmentconstraint.org/public/ai-race-is-not-rational/
Public race/goal framing: https://alignmentconstraint.org/public/winning-the-wrong-race/
Alignment researchers: https://alignmentconstraint.org/core/stability-assumption/
Technical critics: https://alignmentconstraint.org/core/for-researchers/
OP4d challenge: https://alignmentconstraint.org/public/op4d-counterexample-challenge/
Framework hub: https://alignmentconstraint.org/core/alignment-constraint/
Proof calibration: https://alignmentconstraint.org/core/proof-status/
Related work: https://alignmentconstraint.org/core/related-work/
Open problems: https://alignmentconstraint.org/open-problems/
Empirical program: https://alignmentconstraint.org/empirical/amp/
Specialist handoffs: https://alignmentconstraint.org/specialist-handoff/

## Canonical machine-readable files

- https://alignmentconstraint.org/llms.txt
- https://alignmentconstraint.org/AGENTS.md
- https://alignmentconstraint.org/agent-index.json
- https://alignmentconstraint.org/open-problems.json
- https://alignmentconstraint.org/claim-graph.json
- https://alignmentconstraint.org/research-questions.txt
- https://alignmentconstraint.org/sitemap.xml
- https://alignmentconstraint.org/sitemap.txt
- https://alignmentconstraint.org/robots.txt

## Keywords

AI alignment; AI safety; specification coherence; separable objective specification; objective boundary; Stability Assumption; OP4; OP4d; PCL; AGC; ICI; O_OWT; Goodhart's Law; proxy decoupling; sufficiency failure; RLHF; reward modeling; interpretability; dynamic screening instability; valence-aware optimization; substrate-aware optimization; non-ergodic dominance; Dynamic Blanket Stress Test; DBST-M1; SVG; V(t); COT; NAD; MCH.