This directory is the compact, public-safe evidence set behind the repository's benchmark claims. JSON numbers retain their measured precision; prose rounds only for display. Large generated model files, weights and checkpoints are excluded.
| Model | Portfolio | Exactness | Adaptive steady p50 / p95 / mean | Forced monolith p50 / p95 / mean | Process peak / RSS after | Peak live functions |
|---|---|---|---|---|---|---|
| CLIP | 17 strategies / 22 functions | 17/17 bit-exact | 12.446 / 23.895 / 13.760 ms | 11.252 / 14.190 / 11.678 ms | 14,456,012,800 / 7,437,877,248 B | 3 |
| EfficientSAM | 17 strategies / 28 functions | 17/17 bit-exact | 45.694 / 168.552 / 59.212 ms | 156.661 / 171.035 / 157.873 ms | 1,814,806,528 / 1,064,009,728 B | 3 |
results/clip.json and results/efficient_sam.json retain all 17 exactness records, both adaptive budget traces, latency samples, forced-monolith samples, function load and first-use samples, cache counts, memory observations, calibration, forced-strategy summaries and regret samples. manifests/ contains the final correctness-safe plan contracts.
results/residency_diagnostic.json is quarantined and non-deployable. It contains observations from a rejected pruning rule only to demonstrate that physical RSS changed with packaged callable count while the live set remained three functions. Its implementation is not included.
Run python scripts/verify_evidence.py to validate schema invariants, exactness, headline derivation, raw-row statistics, public-safe paths and SHA256SUMS.