Invariants Are Necessary but Not Sufficient

Structural checks can catch a malformed model before an expensive evaluation. Passing them does not show that the model still answers correctly.

Historical behavior: This note discusses the guard chain used before v0.13. The v0.12.1 guard contracts preserve the structural checks described here. Current InvarLock uses signed paired evidence; the distinction is explained below.

What the structural checks established

The historical invariants guard scanned for non-finite tensors, tokenizer and embedding mismatches, broken weight tying, incompatible shapes, and missing or drifting structural evidence such as LayerNorm or positional-encoding checks.

Fatal findings, including non-finite tensors and tokenizer alignment failures, blocked before evaluation. Other structural drift could remain a warning in default monitor mode unless strict mode or a blocking policy was configured. An invariants pass therefore meant that the fatal checks had cleared, not that every structural property was unchanged.

The old default chain ran invariants before and after the spectral, random-matrix, and variance stages. Both passes remained structural checks; the closing pass was not a quality verdict.

What a structurally valid edit can miss

Consider an illustrative edit that changes finite weight values while preserving every tensor shape, embedding dimension, and weight-tying relationship. The model can load successfully and still choose a different answer to the same prompt.

That example illustrates a logical gap, not a measured result. Structural compatibility says whether the artifact is well formed enough to evaluate. Measuring the changed answers requires behavioral evidence.

In the historical workflow, later primary-metric and guard checks supplied additional evidence under their own contracts. Structural checking remained valuable because it could stop fatal problems before spending compute on those stages.

How the current system separates the questions

Since v0.13, InvarLock has used a closed request, canonical signed evidence, and independent verification. The paired comparison retains the baseline and subject outputs for the same ordered schedule, then recomputes the declared metric, interval, and policy result. See pairing and replay.

Spectral, random-matrix, and variance diagnostics now live in an optional diagnostics package. Their authenticated observations can help explain a comparison, but they do not determine its policy verdict.

For a current review, check the artifact’s required structural properties, then inspect the separate paired behavioral evidence. Neither a successful load nor a structural check substitutes for that comparison.

Limitations

• The guard-chain details describe historical behavior, not current operator instructions.
• The example is illustrative; this note reports no malformed-model experiment.
• A behavioral policy pass remains limited to the exact comparison and schedule. It does not establish general quality, safety, or deployment suitability.

Sources

Website documentation explains the maintained workflow. The structural checks described here use the historical pre-v0.13 guard chain; the tagged v0.12.1 contracts preserve that evidence basis. The transition article explains the change to signed paired evidence.

• Pairing and replay
• Optional diagnostics
• InvarLock 0.13.0: One Request, One Signed Decision Trail
• Historical v0.12.1 guard contracts and structural coverage