Selective Weight Loading in the Operating Path / V87
The learned reader’s encoder was loaded through EMMA’s block store: fetched by byte range, checked tensor by tensor against a pinned manifest, and held in a bounded cache. Encoder states and readings were bit-identical to a whole-file load in every condition, a damaged cache entry was detected and fetched again, and every gate passed.
SCOPE + LIMITATIONS
- A cold remote load takes about six minutes at this connection speed, so the demonstration reads the local pinned copy through the store and keeps remote storage for machines without it. The byte saving for this file is small; the gains are verified residency and a bounded local copy. Measured on the CPU with one encoder; parallel range requests are not built.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Load 134 of the encoder’s 137 tensors (596 MB) through the block store from pinned remote storage and from a local copy, cold, warm, under a 268 MB cache budget, and with one cache entry deliberately damaged. Compare every condition bit for bit with the whole file loaded locally on 24 fresh configurations only the learned reader can read. Answers list every executed component with its size.
Bit-identical states and readings, cache within budget after every load, every tensor checked, and a size listing for every executed component all passed. Once the cache is warm the store costs nothing against a local load.
- Whole file, local · reference
- 3.0 s
- Warm cache
- 2.6 s; identical
- Cold, remote storage
- 351.9 s; 596 MB fetched; identical
- Bounded cache, second load
- 230.0 s; 34 evictions; cache within budget; identical
- One cache entry damaged
- Detected by checksum, discarded and fetched again; identical
Question and method
Load 134 of the encoder’s 137 tensors (596 MB) through the block store from pinned remote storage and from a local copy, cold, warm, under a 268 MB cache budget, and with one cache entry deliberately damaged. Compare every condition bit for bit with the whole file loaded locally on 24 fresh configurations only the learned reader can read. Answers list every executed component with its size.
Recorded results
| Condition | Recorded result |
|---|---|
| Whole file, local · reference | 3.0 s |
| Warm cache | 2.6 s; identical |
| Cold, remote storage | 351.9 s; 596 MB fetched; identical |
| Bounded cache, second load | 230.0 s; 34 evictions; cache within budget; identical |
| One cache entry damaged | Detected by checksum, discarded and fetched again; identical |
Comparative standing and locked gates
Bit-identical states and readings, cache within budget after every load, every tensor checked, and a size listing for every executed component all passed. Once the cache is warm the store costs nothing against a local load.
Interpretation and limitations
A cold remote load takes about six minutes at this connection speed, so the demonstration reads the local pinned copy through the store and keeps remote storage for machines without it. The byte saving for this file is small; the gains are verified residency and a bounded local copy. Measured on the CPU with one encoder; parallel range requests are not built.
SOURCE PROVENANCE
V87: selective loading in the operating path (FR-G)
LABORATORY REPORT / 2026-10-20SOURCE CHECKSUM / SHA-256
9a7a27d123cfa8a9398924cfc587a06f496442c9faf90221ce46fd9944e404bcPublic journal edition reviewed 2026-10-22. Source documents and saved evidence were inspected; experiments were not rerun for this edition. Proprietary implementation code, model binaries, private infrastructure, and detailed machine records are not published here. Journal identifiers are editorial references. Catalog inclusion does not imply qualification or runtime promotion.