What makes an AI memory reduction evidence-valid?

An implementation can demonstrate that a transformation is technically possible. A simulation can estimate how that transformation might behave. A calibration run can help set thresholds. None of those observations automatically establishes the final claim.

VAMANIR treats candidate generation and result authority as different stages. A candidate may be promising, efficient or internally consistent and still fail to qualify as a result because the denominator is incomplete, the workload changed or the validation evidence was used during selection.

A defensible reduction should travel with the information required to interpret and challenge it. Without that record, identical ratios can describe physically different achievements.

The passport is not administrative decoration. It is the minimum context required to understand what became smaller, under which conditions and with what degree of authority.

• Target: model, system and version.
• Denominator: the exact memory quantity being reduced.
• Baseline: measurement procedure and runtime conditions.
• Candidate: complete composition of applied changes.
• Quality boundary: metrics, tolerances and workload coverage.
• Evidence: calibration, selection and final validation sources.
• Limitations: known exclusions and unresolved conditions.

When the same evidence repeatedly guides candidate selection and certifies the final result, the search can adapt to the test rather than to the underlying requirement. The reported boundary then becomes less independent than it appears.

The research system should distinguish evidence used to explore, evidence used to calibrate decisions and evidence reserved for final confirmation. Exact partitions will depend on the target and data regime, but the separation itself is a core protection.

Evidence that chooses the candidate should not be the only evidence that certifies it.

Predicted, simulated, measured and independently reproduced statements do not carry the same authority. They can all be useful, but they should not be presented as interchangeable.

A public note should identify whether a statement is a research position, a modelled expectation, an internal measurement, a controlled validation or an independently reproduced result. The language of the claim should never outrun that level.

An evidence system must be able to return less than the ambition asks for. If quality cannot be defended, if the physical measurement is unstable or if system interactions remain unresolved, the correct output is a blocker.

That discipline is central to zero-quality-loss language. It does not mean every dimension of intelligence can be compressed without consequence. It means a reduction does not count when the defined quality boundary is crossed.