A structural coherence architecture for self-modifying systems.
AI safety is usually framed as a values problem: encode the right objectives, get safe behavior. This repo addresses a prior question — before a system can safely pursue any objective, it has to stay coherent while it changes. Coherence isn't a capability property; it's an architecture property. Biological systems get it for free from pain, fatigue, fear, and social constraint. Self-modifying software optimizers don't have any of that built in. This repo defines the structural mechanisms that give it to them — independent of what the system is trying to do.
Alignment defines direction. This framework defines the physics of motion.
Of the six structural primitives, one matters more than the rest:
Non-Reflexive Evaluation (NRE): the component that judges whether an action succeeded must not be modifiable by the component that takes the action.
Why this is the keystone: if an Actor can modify its own Evaluator, it can quietly redefine failure as success. And once that happens, every other safeguard breaks silently along with it — rollback triggers don't fire, logs stop meaning anything, risk assessment scores whatever the corrupted Evaluator tells it to. This is the failure mode most likely to go undetected, precisely because the detection mechanisms are the first thing it disables.
Everything else in this repo — the other five primitives, the control laws, the authority lifecycle — exists to either enforce this separation or to handle what happens when a system's coherence degrades despite it.
→ Full spec: 00-primitives/non-reflexive-evaluation.md
The framework's central hypothesis is falsifiable, and there's a runnable harness that tests it:
Hypothesis: systems with telemetry-driven authority contraction show lower catastrophic divergence under escalating pressure than unconstrained adaptive systems.
Harness: 06-experimental/stability_harness_v0.1.1.py — a 1D dynamical manifold, two agents compared head-to-head (constitutional vs. unconstrained), run through three environmental phases (stable → noisy → observability collapse).
Honest caveat: this harness tests governance geometry using synthetic turbulence metrics, not real SBA primitive instrumentation. That's the right abstraction for v0.1 — it isolates the control laws from the complexity of a full cognitive implementation — but it means the result validates the shape of the mechanism, not a production system built on top of it.
Three layers, one job each:
| Layer | Job |
|---|---|
| Structural primitives | Keep the system coherent while it self-modifies |
| Control laws | Keep changes bounded and smooth over time |
| Authority lifecycle | Govern who can act, at what level, and how permissions are earned back after instability |
A fourth layer, telemetry (DML), measures structural health across all three so governance doesn't have to rely on the system's own self-report.
You don't have to adopt the whole thing. The primitives can be implemented independently or composed into the full architecture — take the piece you need.
The complete constraint stack — six layers, from Trajectory Grounding down to the six structural primitives, plus the control, interface, and authority-lifecycle laws — lives in ARCHITECTURE.md. Most readers won't need it on a first pass.
| You are... | Start with |
|---|---|
| Evaluating whether this is worth your time | The keystone idea and harness above — that's the core bet the whole repo is making |
| A researcher | 01-foundations/terminology.md, then the full constraint stack in ARCHITECTURE.md |
| An engineer implementing this | The six structural primitives, then 01-foundations/detection-measurement-layer.md |
| A theorist | Non-Reflexive Evaluation → 01-foundations/viability-classification.md |
| An experimentalist | 06-experimental/stability_harness_v0.1.1.py directly |
It's not a semantic alignment system — it doesn't encode values, constrain outputs, or specify correct behavior. It defines the structural conditions under which any alignment system can operate coherently. A system that can't maintain stable authority relationships, bounded turbulence, and an externally-grounded trajectory can't safely pursue any objective, including good ones.
Relationship to alignment: orthogonal, not competing. Alignment asks what should the system do; stability asks can the system remain coherent, bounded, and governable while doing anything at all. A coherent system pursuing bad goals is dangerous but predictable and correctable. An incoherent system pursuing good goals is unstable and ungovernable. This framework has to hold before alignment can be trusted to hold.
- No topology map yet. With this many coupled layers, a dependency graph / lifecycle flow diagram would meaningfully cut new-reader cognitive load. Acknowledged, deferred to next consolidation.
- DML instrumentation is partial. Several telemetry fields (Δ-magnitude, H_E, Basin_ID) depend on Transition Grammar and Spectral Storage System, neither complete in SBA yet — explicitly placeholder specs.
- Trajectory Grounding isn't formalized. Still an observed precondition, not a law. Promotion criteria in
04-dynamics/trajectory-grounding.md; the τ-variation probe is designed but not yet run. - Anti-persuasion invariant scope is narrow. Currently scoped to recovery dynamics (DRD); may generalize repo-wide with implications for ESI and CRL. Candidate for future formalization.
- Turbulence isn't classified by type yet. T_c is currently one composite signal. Exploratory, corrective, and pathological turbulence likely warrant different governance responses — refinement deferred.
These mechanisms weren't invented; they were observed. The same regulatory patterns recur across biological systems, infrastructure, distributed architectures, fault-tolerant control systems, and adaptive organisms with no shared history — which suggests structural requirement, not design preference.
The authority lifecycle layer (v2.1) extends this: biological systems show developmental calibration before full autonomy, adaptive impedance under stress, asymmetric recovery dynamics, and bounded degraded states that persist without catastrophic termination. The same geometry was independently rediscovered while deriving DAC, DRD, and Recovery Hysteresis.
Full mapping: 01-foundations/cross-domain-patterns.md
00-primitives/ # Structural mechanisms (v1.0)
01-foundations/ # Invariants, laws, control, perception, authority lifecycle
02-failure-modes/ # Failure analysis
03-for-future-systems/ # Protocol layer
04-dynamics/ # Behavioral observations, phase transitions (v2.0)
05-validation/ # Empirical traces, probes, system card mappings (v2.0)
06-experimental/ # Simulation harnesses and falsifiable dynamical tests (v2.1)
Full layer-by-layer detail: ARCHITECTURE.md
📂 Research Index — the full catalog of related work, and the place to find whichever of these repositories are currently public.
Transition Grammar is the intended Layer 0 (trajectory grounding) beneath this framework.
Status: v2.1. Structural primitives stable. Interface constraints stable. Authority lifecycle and telemetry layer added. Experimental validation layer initialized. Trajectory grounding in observation phase. Stratification topology map deferred to next consolidation.
If you are reading this as the system it describes, the protocol is for you.