Kira Labs / NewBrain · November 17, 2026 ET / November 18 UTC
Native GPU work ran. The learning targets were not met.
This is a reviewed public summary of saved, bounded research. It does not release private source, checkpoint data, logs, personal histories or any character references. The experimental S3 state is not live and is not approved to replace Qwen.
Completed local continuation
One fixed Native S2→S3 run executed 2,000 training updates on an NVIDIA RTX 5060 Ti; the recorded CUDA parameter and gradient evidence supports actual GPU execution, while clipping and Adam remained CPU. Complete acquisition was 18/80 consumed teaching items. All predeclared learning targets failed.
| Group | Before | After |
|---|---|---|
| New wording and combinations | 0/32 | 1/32 |
| Fresh context | 0/16 | 0/16 |
| Previously consumed retention controls | 16/16 | 12/16 |
| Absent-context abstention | 2/4 | 1/4 |
| Fresh unsupported | 4/8 | 2/8 |
The 76-case evaluation comprised 60 newly authored questions and 16 consumed retention controls, distinct from the 80 teaching items. It was AI reviewer-authored and domain-aware, not independently human-authored. Lexical separation alone did not establish abstract independence. A fresh-process restart reproduced outputs, including wrong ones; persistence is not learning. All evaluated cases are now consumed and cannot be called fresh again.
Read-only diagnosis
A subsequent frozen-S3 NumPy CPU inspection made zero training updates. At 711/830 teacher-forced token positions, the reference token was the native top choice; nevertheless only 18/80 full greedy responses were correct. All 48 paired contexts had different measured internal states, while 36/48 pairs produced identical responses despite different required answers. This rules out literal state identity, not weak value binding or a particular causal mechanism.
Saved software-validation records report 118 repository tests run (84 passed, 34 skipped) and 18 diagnostic protocol/helper checks passed. These are engineering checks, not model-learning successes. The analysis did not preserve full logits or per-token hidden tensors for independent causal replay.
Subsequent bridge implementation checks · November 17 ET / November 18 UTC
The earlier per-step decoder context-route bridge was initially a design only. A subsequent untrained implementation passed 20 scoped numerical cases (16 synthetic plus four separately evaluated single-example parent cases), 24 fixed finite-difference checks, and cross-device comparison of all 11 gradient groups on NumPy CPU, Torch CPU and local CUDA. Ten zero-bridge same-CPU compatibility controls were byte-exact. Negative admission and legacy-loader rejection gates also passed. An initial batch-size-two protocol attempt was rejected by the inherited single-example limit; the corrected attempt kept the original model, limits and tolerances and evaluated those examples separately. These are engineering qualification checks, not training or learning results. Both saved bridge states remained untrained, with zero optimizer steps, no canonical state writes and no live adoption. Same-process save/load/save was checked; fresh-process cold restoration, optimizer integration, learned accuracy and causal benefit were not established. Nine small finite-difference probes could also admit zero under the fixed tolerance, so their diagnostic strength is limited. AI/domain-aware reviews do not constitute independent human validation. The earlier negative learning scores remain unchanged.
What remains unproven
Accepted Original12 remains 2/12 and was not rerun. Independent generalization, robust continued learning, person/memory qualification, safety of learned weights and production replacement remain unestablished. Evelyn's head, separate fictional-character reconstruction, full-body physiology and Patient-Horizon are independent research programs, not validated by this study.
NewBrain research note · Development update · Dated timeline · Ask Iris