Learned Output-Mode Decision / V85–V85.2
A Mode Decision fine-tuned on 710 requests labelled blind by two Source Models reached 83.6% on held-out operations, against 44.2% for V81, but did not beat the rule control at 84.7% and failed five of seven gates. Composed with the rule it reached 85.5% with no precision misses. V85.1 and V85.2 also failed; the rule control stays in service.
SCOPE + LIMITATIONS
- Labels are Source Model consensus, not independent verification, and the table’s modes came from one of the same Source Models. Scored sets of 124 and 111 requests give intervals 12 to 14 points wide, from one fit and one seed. Generated requests are cleaner than real ones (94.2% Source Model agreement against 84.3%). The rule control was measured with each request’s own operation always offered. The run also found that a full fine-tune of the Reference Decision needs about 6.8 GB and had been spilling beyond the 4 GB card; later recipes train the top 12 of 28 layers to stay on the card. The measured division of labour, a rule where an operation matches and a learned Decision where none does, is a finding about this corpus.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Stage 0 runs checks that need no training: each V81 weight set answers with the three mode options in all six orders, and a rule control is measured in which the linked Stem Decision chooses the operation and the operation table supplies its mode. Stage 1 rebuilds the training ledger: two Source Models label 630 original and 839 contrastive requests blind, and only rows both support are admitted, under a 50% cap on rows supported by Source Model consensus alone. Stage 2 fine-tunes the Reference Decision once on the 710 admitted rows with options shuffled. V85.1 trains a two-stage Decision with only the top six of 28 encoder layers trainable so that it fits the 4 GB graphics card; V85.2 fits small word-feature units on the CPU. Sealed questions are never labelled or scored.
Stage 2 failed five of seven locked gates: fit 84.7% against 95%, accuracy 81.5% against 90%, precision miss 10.9% against at most 2%, open recall 66.7% against 75%, and option-order range 4.0 points against 3. The composition passed its precision gate at 0.0% and sits one request above the rule control, inside the sampling error. V85.1 failed all seven: the trainer’s own check showed that the partly frozen links had not learned, so it is a failed fit of a memory-safe recipe rather than a test of the two-stage design. V85.2 failed six of seven. The Stage 2 weights are the retained best among learned Mode Decisions; the rule control, composed with the learning node, stays in service. No gate was changed and the sealed theme remained unread.
- V81 head · held-out operations, six option orders
- 44.2% mean; first-listed option chosen 62.0% of the time
- Rule control · held-out operations (124)
- 84.7%; 2.2% precision miss
- Stage 2 fine-tune · held-out, mean / worst order
- 83.6% / 81.5%; precision miss up to 10.9%
- Composition, exact if either says exact
- 85.5%; 0.0% precision miss
- Trained operations’ dev requests (111), mostly matching no registered operation
- Rule 39.6%; fine-tune 87.1%; CPU unit 79.3%
- V85.1 two-stage, top six layers trained
- 42.1%; links did not learn
- V85.2 CPU word-feature unit · held-out
- 67.7%; 32.6% precision miss
- Source Model agreement · 630 original requests
- 84.3%; 99 disputed rows excluded
Question and method
Stage 0 runs checks that need no training: each V81 weight set answers with the three mode options in all six orders, and a rule control is measured in which the linked Stem Decision chooses the operation and the operation table supplies its mode. Stage 1 rebuilds the training ledger: two Source Models label 630 original and 839 contrastive requests blind, and only rows both support are admitted, under a 50% cap on rows supported by Source Model consensus alone. Stage 2 fine-tunes the Reference Decision once on the 710 admitted rows with options shuffled. V85.1 trains a two-stage Decision with only the top six of 28 encoder layers trainable so that it fits the 4 GB graphics card; V85.2 fits small word-feature units on the CPU. Sealed questions are never labelled or scored.
Recorded results
| Condition | Recorded result |
|---|---|
| V81 head · held-out operations, six option orders | 44.2% mean; first-listed option chosen 62.0% of the time |
| Rule control · held-out operations (124) | 84.7%; 2.2% precision miss |
| Stage 2 fine-tune · held-out, mean / worst order | 83.6% / 81.5%; precision miss up to 10.9% |
| Composition, exact if either says exact | 85.5%; 0.0% precision miss |
| Trained operations’ dev requests (111), mostly matching no registered operation | Rule 39.6%; fine-tune 87.1%; CPU unit 79.3% |
| V85.1 two-stage, top six layers trained | 42.1%; links did not learn |
| V85.2 CPU word-feature unit · held-out | 67.7%; 32.6% precision miss |
| Source Model agreement · 630 original requests | 84.3%; 99 disputed rows excluded |
Comparative standing and locked gates
Stage 2 failed five of seven locked gates: fit 84.7% against 95%, accuracy 81.5% against 90%, precision miss 10.9% against at most 2%, open recall 66.7% against 75%, and option-order range 4.0 points against 3. The composition passed its precision gate at 0.0% and sits one request above the rule control, inside the sampling error. V85.1 failed all seven: the trainer’s own check showed that the partly frozen links had not learned, so it is a failed fit of a memory-safe recipe rather than a test of the two-stage design. V85.2 failed six of seven. The Stage 2 weights are the retained best among learned Mode Decisions; the rule control, composed with the learning node, stays in service. No gate was changed and the sealed theme remained unread.
Interpretation and limitations
Labels are Source Model consensus, not independent verification, and the table’s modes came from one of the same Source Models. Scored sets of 124 and 111 requests give intervals 12 to 14 points wide, from one fit and one seed. Generated requests are cleaner than real ones (94.2% Source Model agreement against 84.3%). The rule control was measured with each request’s own operation always offered. The run also found that a full fine-tune of the Reference Decision needs about 6.8 GB and had been spilling beyond the 4 GB card; later recipes train the top 12 of 28 layers to stay on the card. The measured division of labour, a rule where an operation matches and a learned Decision where none does, is a finding about this corpus.
SOURCE PROVENANCE
Mode Decision V85, Stage 0: free checks before any new fit
LABORATORY REPORT / 2026-10-11SOURCE CHECKSUM / SHA-256
9bcb8f5fb19d6f0286cc80b9b1cc7ca67045246eae957c9ba789ffc99040a6c1Public journal edition reviewed 2026-10-19. 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.