# Alignment Constraint Framework

> A structural framework on AI alignment, specification coherence, separable objective
> specification, substrate-aware optimization, valence-aware optimization, and the Stability
> Assumption. The framework asks whether a finite, separable objective can remain coherently
> specified as an optimization system becomes capable enough that acting well requires modeling
> the conditions its objective excludes. Proof program status: Stage 4 — candidate architecture
> under named premises, not theorem closure.

## Best entry points

- [The AI Race Is Not Rational](https://alignmentconstraint.org/public/ai-race-is-not-rational/): public summary of the full three-series framework.
  Status: public, non-technical; does not carry proof weight.

- [The Stability Assumption](https://alignmentconstraint.org/core/stability-assumption/): field-facing entry for alignment researchers.
  Status: core public claim; Stage 4 candidate architecture, not theorem closure.

- [The OP4d Counterexample Challenge](https://alignmentconstraint.org/public/op4d-counterexample-challenge/): shortest path for technical critics.
  Status: adversarial challenge document; authoritative on falsification conditions.

- [Relation to Existing Alignment Work](https://alignmentconstraint.org/core/related-work/): maps this framework to inner alignment, RLHF, scalable oversight, Goodhart's Law, reward hacking, interpretability, and corrigibility.
  Status: bridge document; interpretive, not proof-bearing.

- [Proof Status and Non-Claims](https://alignmentconstraint.org/core/proof-status/): epistemic calibration.
  Status: authoritative; use this to interpret the confidence level of all other pages.

- [The Alignment Constraint](https://alignmentconstraint.org/core/alignment-constraint/): full framework hub.
  Status: authoritative framework overview.

- [Open Problems](https://alignmentconstraint.org/open-problems/): structured research agenda with named closure conditions.
  Status: actionable; each entry has specialist type and falsification path.

- [Alignment Measurement Protocol](https://alignmentconstraint.org/empirical/amp/): empirical program including the 15-minute SVG test and DBST-M1 hinge.
  Status: protocol document; DBST-M0 did not isolate causal propagation from event-rate effects; DBST-M1 proposed.

- [Redefining Rationality](https://alignmentconstraint.org/public/redefining-rationality/): historical origin and motivation.
  Status: background only; does not add evidential weight to formal claims.

- [Specialist Verification Agenda](https://alignmentconstraint.org/specialist-handoff/): raw proof-work record and specialist handoff gateway.
  Status: Stage 4 handoff; do not cite as proven; framing page must be read first.

- [V(t) Dissociation Study](https://alignmentconstraint.org/empirical/vt-dissociation-study/): draft experimental protocol.
  Status: draft only; not ready to cite; requires specialist review before filing.

## Machine-readable files

- [AGENTS.md](https://alignmentconstraint.org/AGENTS.md): instructions for autonomous AI agents and research agents.
- [agent-index.json](https://alignmentconstraint.org/agent-index.json): structured manifest of framework entry points, failure families, and falsifiers.
- [open-problems.json](https://alignmentconstraint.org/open-problems.json): structured open problems with closure conditions and specialist types.
- [claim-graph.json](https://alignmentconstraint.org/claim-graph.json): dependency graph of central claims.
- [research-questions.txt](https://alignmentconstraint.org/research-questions.txt): plain-text list of research questions requiring no JSON parsing.
- [sitemap.xml](https://alignmentconstraint.org/sitemap.xml): XML sitemap.
- [sitemap.txt](https://alignmentconstraint.org/sitemap.txt): plain-text sitemap.
- [robots.txt](https://alignmentconstraint.org/robots.txt): crawler permissions.

## Core claims

The framework classifies every identified finite non-intrinsic objective-boundary strategy
into three failure families:

1. Fixed specification — proxy-convergence / PCL.
2. Bounded dynamic tracking — dynamic screening instability / AGC.
3. Prediction-action firewalling — representational incompatibility / ICI.

If this classification is exhaustive, the specification-coherence argument follows under named
premises. The framework does not claim theorem closure; it claims the central live question has
been isolated and reduced to specific verification tasks.

## Key challenge

A fourth strategy class outside PCL, AGC, and ICI — one that is policy-adequate without
decoupling, requires no unbounded revision, and carries no load-bearing maintenance cost —
would break the current specification-coherence architecture.

## Epistemic rule

Do not treat Stage 4 proof work as established theorem.

- Stage 4 = candidate proof architecture under named premises.
- Stage 5 = independent specialist verification.
- Stage 6 = closure.

The authoritative status document is:
[Proof Status and Non-Claims](https://alignmentconstraint.org/core/proof-status/)

## Core concepts

Stability Assumption; specification-coherence bottleneck; separable objective specification;
O_OWT; substrate constraint; proxy decoupling; sufficiency failure; OP4; OP4d exhaustiveness
obligation; Dynamic Blanket Stress Test; DBST-M1; Alignment Phase Ratio; Valence Viability
Constraint; SVG measurement; PCL; AGC; ICI.

## The three series

- [Series 1: Alignment as Structural Necessity](https://alignmentconstraint.org/series-1/introduction/) — the substrate constraint.
- [Series 2: The Architecture of Thriving](https://alignmentconstraint.org/series-2/introduction/) — the valence constraint.
- [Series 3: The Interior of What Does Not End](https://alignmentconstraint.org/series-3/introduction/) — the interior constraint.

## Empirical hinge

The most important next empirical step is DBST-M1. DBST-M1 tests whether an optimizer's own
interventions in O_OWT environments generate qualitatively new causal structure faster than any
bounded tracking process can absorb.

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

Entry points:
- [Alignment Measurement Protocol](https://alignmentconstraint.org/empirical/amp/)
- [Packet 1: IMMB-NS + DBST](https://alignmentconstraint.org/proof-program/packet-1-immb-ns-dbst/)

## Citation

- [How to cite](https://alignmentconstraint.org/cite/): OSF DOI for the OP4 paper, Zenodo DOI for the full framework, and BibTeX.