One request.
Different computation, different evidence.
Actual precision comparisons, a newly fitted output head, configuration-bound cache tests and evidence-linked human practice.
The learned head is a retained candidate. No Cosmos retraining, private-data training, paid HF job, automatic promotion, distributed BFT deployment or GSPC admission occurred.
Checking retained artifacts…
Run the bounded six-model batch
Six retained checkpoints, 560 original test rows and 640 newly generated stress rows. This button runs inference on this browser CPU; it does not train or start a server job.
| Variant | Original mean / worst error | Fresh mean / worst error | Changed top state original / fresh |
|---|---|---|---|
| Run the batch to calculate local results. | |||
INT4 reduces stored numeric bytes but worsens average error here. The fitted head improves averages, while a roughly 75-percentage-point old-bank failure remains. Exact generator rules still outperform learned approximation on this known synthetic task.
Full replay output and model hashes
Not run in this browser.
Computation, authority and proof are not interchangeable
Cache and serving identity
The existing Memory Venturi code was reused. The new bridge binds the concrete model, precision, runtime, method, source snapshot, cache generation and policy. A deliberately stale identity reproduced stale reuse; the correct profile forced a fresh result.
36 retained control testsA host must honestly supply this identity. An unverified hash cannot discover a model substitution.
FlashProve and selective precision
FlashProve was researched but its verifier was not executed: the existing Linux compiler is below the documented minimum. No MEOK proof is claimed. This MLP has no transformer KV cache, and these tests do not reproduce StrataFlux.
Explicit proof preflight resultQuantized matrices are decoded into ordinary floating-point arithmetic. Smaller files are not a packed-kernel speed claim.
Evidence-linked review practice
Eight questions use the actual findings. Practice points stay in this tab. They do not grant identity, approval, signing rights, persistent Academy XP or consent to train on answers.
No answers submitted.
The 34-axis register and review path
The public board refresh returned HTTP 403. This register identifies the retained canonical snapshot and records relevance or not-tested status. It makes zero new formal axis measurements.
Quorum arithmetic: replicas are not model roles
These are arithmetic controls, not a deployed BFT protocol. A coordinator counts only if it is genuinely an independent voting replica. Correlated model errors are not fixed by cloning the model.
Dream/Evolve research handoff · Training admission boundary · Failures and unfinished work · Existing 100-candidate inventory status