Consolidation: prior answers retained, sorting gate failed

The three-shard experiment completed on 14 September 2026. Its selected 1.7B checkpoint improves fresh task answers from 375/512 to 451/512 and preserves every previously correct answer in a separate 256-case prior-task set. It fails the complete frozen gate because sorting falls from 118/128 to 116/128 on the primary set. The complete result publishes all 768 final answer pairs and every evaluated development pair.

Method and selection

The frozen method combines the accepted phase-A weights with the rejected calculation-step checkpoint. Each owner computes the declared FP32 interpolation for its own parameters and retains the fast checkpoint's complete Adam state. There are no additional gradient updates, routing between models, calculator calls or answer repairs. The result is one fixed-size model with 1,711,376,384 parameters.

Development candidate	Prior correct / 512	Lost prior answers	New correct / 128	Decision
Parent	378	0	97	Baseline
75% learned update	378	1	Not evaluated	Reject
50% learned update	378	0	118	Select

The 75% candidate gains one prior arithmetic answer and loses another. The zero-loss rule rejects that exchange despite the unchanged total. The 50% candidate preserves the entire correct prior set and gains 21 new arithmetic answers; conversation loss changes by +0.001206 nats. It is the first complete development pass. The 25% and 12.5% candidates are therefore not attempted. The actual checkpoint is committed before final evaluation and is not replaced after the failed final result.

Final measurements

Parent and candidate receive identical prompts and a 256-token greedy limit. The earlier strict JSON scoring rules still apply. Fresh cases and conversation documents exclude every role from all three preceding experiments by identity and normalized content. Instruction families remain public; this measures new examples of those families, not general assistant quality.

Measurement	Parent	Candidate	Decision
New-task correct answers, 512 cases	375	451	Gain passes; sorting floor fails
New-task invoice totals, 128 cases	4	82	80 wins, two losses
New-task sorting, 128 cases	118	116	Two losses
Prior-task correct answers, 256 cases	194	194	No individual wins or losses
Conversation response loss, 256 documents	0.539504	0.540975	Retention bound passes

New-task lookup remains 128/128 and filtering 125/128. All prior families retain their complete correct-answer sets: lookup 64/64, filtering 63/64, sorting 64/64 and invoice totals 3/64. The primary set has 80 wins and four losses, with exact one-sided McNemar p = 1.049e-19. Conversation loss changes by +0.001470 nats; its 95% bootstrap upper bound is +0.002425, below the frozen +0.02 limit.

One lost sorting response uses top_two instead of ids while selecting the right items. The other selects the wrong second item. They occur under two different instruction wordings, including one already used in training. The two arithmetic losses contain incorrect additions in generated calculation steps. All remain errors. The large arithmetic gain cannot override the separate sorting floor.

The result supports retaining the new arithmetic skill with less interference on the measured prior sample. A larger development probe and smaller update still do not establish preservation on every fresh task set. The final data are now exposed and cannot approve a replacement candidate.

Execution and preserved state

Three A10G hosts in separate availability zones ran under one operator. Each held only its own parent, learned and optimizer partitions; boundaries were [0, 6, 15, 24]. Owners held 503,343,104 / 604,016,640 / 604,016,640 model parameters. The complete screen took 2,149.92 seconds, with peak allocated GPU memory of 8.33 / 10.21 / 10.21 GiB. All owners agreed on the selected checkpoint and the complete evaluation measurements.

The initial parent evaluation hit a 180-second transport timeout before any answer-generation event or completed report. One worker was observed blocked on I/O. The failed logs and exit codes are retained. Preloading and SHA-256 checking both owned input checkpoints preceded a successful retry of the unchanged numerical evaluation. This is an operational recovery, not a second candidate selection; the first attempt has no completed report to compare.

Completed parent/candidate evaluations took 770.56 / 1,996.71 seconds. New-task generation used 6,408 / 22,358 tokens and 450.46 / 1,682.22 seconds. The additional calculation tokens are part of the method's cost.

• Numerical source and plan freeze: 89ea9edc4fc43b0588b3609731b2e30c3e261554.
• Prepared-input commit: acb0c0a; prepared identity af7526b07a8046ecbb0337875030eed5e2d43d48ab1b6069697ea9bee42da205.
• Actual selection commit: 59b1f0fcdb836a005f7d979afd12be999cc28858.
• Selected checkpoint: 4f5fcde935dba361a83a65f63c1db48c479796debd574d8ea9f7e10dc959bddb.
• Selected learned-state root: 917895b27de4ebaf9b418ca268314e70cb62e2f1c78bc2cefed7b5be020511b5.

Validation passed 435 tests, repository and distribution checks. Both CI checks passed at the selection commit. Reproduction must use the pinned numerical source bytes. Both attempted checkpoints, including full Adam state, have verified S3 readbacks. The full evidence archive has SHA-256 322a6003c2e25f0a2116ebc0757d76155f0a04182dcd63eb444002218d9f234d. Large tensors and full execution evidence remain in operator-controlled storage.

All three temporary instances, root disks and the dedicated security group were confirmed deleted at 19:06:13 UTC; both CPU network hosts remained running. Estimated compute is $6.23. The conservative planning total including transfer and a $10 allowance is $23.42, below the $100 cap. These are estimates rather than an invoice; retained S3 storage accrues separately. No NEURO was issued and no serving promotion occurred. Useful growth, broad assistant quality and public network integration remain open requirements.