Skip to content

Core Claims — Viability and Identity Branches

Status: Synthesis index
Scope: A minimal claim set for the Viability Arc and identity branch — not the organizing theory of the whole repository; the model-identification arc is registered separately.
Purpose: Keep these hypotheses precise, challengeable, and linked to artifacts.
See also: Identification Claims — the register for the model-identification arc, proposed 2026-09-02.

Relation to the Project

The Foundations Reconstruction supplies the shared process language and its limits. The repository then separates a bounded model-identification programme from a viability arc. The older Generator Question no longer supplies a universal spine; its useful residue is the measured underdetermination of candidate process models under declared evidence and interventions. The claims below belong to the viability arc and to one proposed identity test family; they are not consequences of the foundation.

Their status is mixed. The Viable Corridor has a conditional necessity result and two synthetic demonstrations of capability loading; the wider applicability of the claims below remains hypothesized. Each claim therefore carries an artifact and a failure condition.

Claim 1: Substrate Veto

Claim: Every implemented optimizer has finite physical resource limits. In a declared model, an independently enforced capacity constraint can throttle selected actions before a monitored substrate variable leaves its operating envelope. Physical exhaustion alone is failure, not a safe veto, and no universal ecological scalar \(D_{\max}\) has been established.

Not obvious because: Better local prediction or higher efficiency can make the global system less viable if the optimization process consumes the physical, ecological, or institutional substrate it depends on.

Artifacts:

Current status: two runnable toy models, no measurement outside them. Both put the coupling into the equations rather than deriving it: the veto simulation stipulates a substrate-stress penalty, and the breathing model compares two declared compute schedules. No universal ecological scalar \(D_{\max}\) has been established, and neither toy shows the coupling is unbreakable.

Failure condition: The design claim is weakened if the controller bypasses or manipulates the constraint, exports harm to an unmeasured substrate, fails before enforcement acts, or gains no robustness over a simpler resource budget under matched tests.

Claim 2: Impedance Matching

Claim: High-speed silicon proposal generation becomes unsafe when it exceeds the absorption bandwidth of slower biological or institutional regulators. Action budgets and latency are safety mechanisms, not merely inefficiencies.

Not obvious because: The usual engineering instinct is to remove latency. This claim says some latency is structurally necessary when fast agents are coupled to slow legitimacy, review, or care systems.

Artifacts:

Current status: the weakest evidential base in this set. All three artifacts are [SPECULATIVE] architecture logs; none is runnable and nothing here is measured. The claim rests on an Ashby's-law argument in temporal form plus one cited external episode (high-frequency-trading flash crashes). Log 002 names its own promotion condition — latency budgets formalized as measurable regulator-capacity constraints — and that work has not been done.

Failure condition: If a shared human-AI governance layer can preserve legitimate human review while operating at silicon-speed delegation and proposal volume, the impedance claim is weakened.

Claim 3: Identity Persistence

Claim: A system can score high on prediction and adaptation while lacking stable identity if its goals, constraints, and values are consulted without being jointly satisfied at commitment. The current hypothesis is functional: persistent identity requires commit-time composition under perturbation, not physical simultaneity.

Not obvious because: Fluent self-description can be mistaken for stable identity. This claim separates rhetorical continuity from dynamical persistence.

Artifacts:

Current status: Exp5–7 measure a first toy dissociation, locate a passive binding signature, and show that two hand-built mimics still leak. They also expose blind metrics: Δ-Kohärenz misses the binding at one level, and IP can be fooled by consultation without composition.

Failure condition: If an optimized mimic preserves the same held-out constraints under adversarial perturbation without commit-time composition, or if the distinction disappears on real models under matched controls, this claim is weakened.

Claim 4: Vital Floors in Governance

Claim: Public AI governance cannot be evaluated only by model behavior or aggregate efficiency. It must preserve measurable floors for human vital systems: food, heat, care, housing, utilities, trust, and crisis response.

Not obvious because: A proposal can be safe at the model-output layer and efficient in aggregate while degrading the people whose viability gives the system legitimacy.

Artifacts:

Current status: one runnable toy compares two planners under a shared shock sequence and operationalizes the log. The fiction entry is where the framing came from and carries no evidential weight — see What This Project Does NOT Claim §16. Nothing here is calibrated against a real vital system.

Failure condition: If vital-floor governance does not reduce red-line violations compared with naive efficiency optimization, or reduces them only by hiding harm in unmeasured indicators, the architecture claim is weakened.

Maintenance Rule

Every new synthesis claim should add:

  1. a precise statement,
  2. a non-obvious implication,
  3. at least one artifact link, each marked with its evidential kind — measured, runnable toy, argument, speculative design note, open questions, or narrative origin, not evidence,
  4. a current-status line saying what the artifacts actually establish,
  5. a failure condition.

Rule 3 exists because an unmarked list flattens the distinction the rest of the repository works to keep. A [SPECULATIVE] architecture log, an executable simulation, and a short story are all legitimate artifacts and are not the same evidence. Reading the bullets should be enough to tell them apart; clicking through should not be required.