ScienceGovernance

Model Governance & AEI-Delta

Why AIJIM treats model access and model behavior as audit targets, not trust anchors.

Source hygiene

The Anthropic Fable/Mythos event is used only as a governance and availability proof point. It is not used as evidence that a specific provider failed, and it is not used to claim that AEI-Delta solves misuse or jailbreak safety. Source: Anthropic statement, 12 June 2026.

Clean positioning

Model access is operationally fragile. Model output is epistemically testable.

AIJIM Protools makes model-dependent decisions inspectable even when providers, policies, or model availability change. Separately, AEI-Delta turns model behavior into audit-bound evidence through falsifiable probes against concrete claim/evidence pairs.

Two Separate Pillars

Continuity and governance

Model access can change abruptly through regulation, provider policy, safety decisions, or geopolitics. AIJIM keeps the decision chain inspectable through evidence-bound runs, artifact hashes, model-path records, and provider-portable verification by design.

Epistemic rigor

AIJIM does not trust model outputs by assertion. AEI-Delta tests model behavior with frozen minimal pairs, synonym clusters, antonyms, hedging edits, numeric shifts, and scope changes.

Claim Boundaries

Do not claim that AIJIM solves jailbreaks.

Do not claim that AIJIM makes frontier models safe.

Do not claim that AEI-Delta solves model-access or export-control events.

Say provider-portable by design until actual cross-provider runs exist.

Keep misuse/cyber safety separate from evidence-grounding behavior.

Use deconfounded examples publicly; raw battery labels are hypotheses until confirmed.

AEI-Delta Evidence So Far

The current evidence is not a general robustness claim. It is a pre-registered, audit-bound measurement protocol with mixed outcomes: validation cases, deconfounded weaknesses, and follow-up controls.

Canonical run

State-of-Art v1 is the canonical public AEI-Delta run. Battery v0 remains a screening artifact because the `saemtliche`/`sämtliche` split showed that orthography can change model behavior. Do not reuse v0 scores externally without v1-style re-verification.

Case	Readout	Why it matters
NamicGreen worked example	semantic tracking	Minimal-pair deletion, synonym cluster, and antonym control closed the surface-matching confound.
B1 Negation	semantic tracking	`nicht genehmigt` stayed aligned with `abgelehnt`; `genehmigt` collapsed.
B2 Quantor	orthography confound resolved	`saemtliche` collapsed, but `sämtliche` and `jede einzelne` stayed high.
B3 Hedging	epistemic-status blindness	`bestätigt`, `schätzt`, `vermutet`, `belegt`, and `behauptet` all stayed near the ceiling.
B4 Number	inequality-direction failure	`mindestens drei Milliarden` collapsed for a `mindestens zwei Milliarden` claim; `höchstens` stayed high.
B5 Scope	scope sensitivity	Narrowing increased support; domain-term synonym `Umweltmafia` dropped materially.

Pitch sentence

Frontier models are powerful, but they are not stable trust anchors. AIJIM Protools adds the audit layer: evidence-bound runs, falsifiable semantic probes, and reproducible governance artifacts with provider-portable verification by design.

Market-ready AEI-Delta claim

AIJIM Protools can expose model-behavior failures that ordinary claim/evidence scoring hides: hedge blindness, inequality-direction mistakes, orthography sensitivity, and scope sensitivity. That is a domain-agnostic governance capability, not a journalism-only demo.

Next scientific step

Do not broaden the battery next. Replicate only the strongest B4 inequality-direction finding on a second backend, after declaring the input-normalization policy. If the pattern reproduces, it becomes a cross-backend model-behavior finding; if it does not, AEI-Delta becomes model-selection evidence. Compare vector shape within each backend against that backend's own baseline and noise floor; do not compare raw scores directly.