# The context we keep. The evidence we need.

Cavumo · Research notebook · Internal engineering report · Not peer reviewed

## Selecting from permitted evidence

Context 1 investigates compact learned selection over local embeddings and retrieval features. Permission checks happen before selection; a scoring model cannot broaden access or rewrite organizational policy. The MVP already exposes a working scoped hybrid context resolver independently of this research candidate.

## Strong baselines before promotion

The clause-selection study used original task-family-separated data: 320 training, 80 validation, and 320 held-out queries across two corpora, plus 40 permission scenarios. Validation triage compared dependency-preserving clause rules with the learned candidate. A replicated export produced an identical ONNX artifact.

## The decision

On 80 validation queries, the candidate covered the annotated evidence or correct empty selection in 78 cases; the strongest eligible clause baseline covered 80. Estimated complete input tokens totaled 22,391 for the candidate versus 18,238 for the baseline: 22.77% more. These are coverage proxies and tokenizer estimates, not generated-answer accuracy. The candidate was not promoted; full native quality and latency gates were not run for this failing candidate.

## The target remains unchanged

Context 1 must deliver at least 20% fewer input tokens than the strongest eligible baseline, at most 2 percentage points of quality degradation, and at most 5% p95 latency regression including specialist overhead. The next iteration needs better selection and full-pipeline evidence, including memory and cross-agent scenarios. No learned specialist is active today.