Invariants Are Necessary but Not Sufficient

Invariants stop structural corruption, not all meaningful regressions

Highlights

• The invariants guard appears at both ends of the default chain: the first pass fails fast on fatal structural issues, and the closing pass stays structural too.
• In default monitor mode, an invariants pass means fatal checks cleared; some structural drift can still be recorded as warnings.
• Quality preservation requires later evidence from primary metrics and downstream guards.

What Invariants Actually Check

The public guard documentation is explicit that invariants are structural. They scan for non-finite tensors, tokenizer and embedding mismatches, broken weight tying, shape compatibility, and missing or drifting structural evidence such as LayerNorm or positional-encoding checks.

Early structural checking has a narrow job. Fatal invariant types such as non-finite tensors or tokenizer alignment failures block before evaluation. Other structural drift remains visible as warnings in default monitor mode unless strict mode or a blocking policy is configured.

Why A Pass Is Not A Quality Claim

Readers often blur structural sanity and acceptable behavior together. They are not the same thing.

An edit can preserve tensor shapes, avoid NaNs, and keep tied weights intact while still degrading the primary metric, tripping spectral limits, or violating activation stability. The guard contract primer makes that separation visible by listing invariants next to separate primary-metric, spectral, RMT, and variance gates.

Why Invariants Belong At The Front Of The Chain

The guard chain works because its stages do different jobs in sequence. Invariants clear the path for everything else. They answer: is this edit even well-formed enough to evaluate?

Putting invariants first is therefore not cosmetic. It is a resource and interpretation boundary. Fatal invariant failures should abort before later evidence is generated, because later evidence would only create noise around a structurally invalid artifact.

The default chain also runs invariants again after the spectral, RMT, and variance stages. That closing pass is still a structural check. It can catch post-stage structural drift, but it is not a quality verdict.

What The Later Guards Add

Once fatal invariants clear, the rest of the pipeline starts to measure behavior rather than shape. Spectral monitoring checks baseline-relative weight drift. RMT checks activation edge-risk movement. The primary metric and paired evaluation windows check whether the edit preserved measured behavior on the configured evaluation windows.

A reviewer should therefore read an invariants pass as a necessary structural condition, not as the conclusion of the review.

Claim Map

The narrow claim is straightforward:

• pre- and post-stage invariants protect against structurally invalid edits
• invariants do not certify quality preservation
• later guards and primary metrics carry that burden

No single green check should carry the whole assurance story.

Limitations

• No new malformed-model case study is included.
• The claim is limited to the current public guard behavior: invariants are structural checks before and after the behavioral guard stages, while later guards and metrics carry the quality-preservation evidence.

Sources

• Guards
• Guard Contracts & Primer
• Assurance Case
• Reports Reference