TrickyRex/mini-neuralhorner-v02

MiniNeuralHorner v0.2

This is a development snapshot of NeuralHorner with 126,603 trainable parameters. It uses the same learned transition and fixed Horner schedule as the larger v8 model, but the recurrent hidden width is 61 instead of 128.

The learned cell approximates one transition:

s_next = (2*s + d*x) mod p

The cell is reused to reduce a, reduce b, and multiply the two residues. The surrounding code sequences binary digits and carries the predicted binary state between calls. It does not compute the modular product as a Python integer operation.

Status

This artifact has now been measured on the SAIR Playground. It is not exact and is not presented as a replacement for TrickyRex/bitserial-modmul-v8.

Hosted SAIR Playground result

The hosted run evaluated repository revision d9d611833d340c72d90a97d995a94031b798cf7c and completed successfully.

Field	Hosted result
Frontier	T9
Displayed overall	99%
Scored tiers 1-10	985/1,000 = 98.5%
All tiers, including the unscored tier-0 diagnostic	1,025/1,100
Runtime	156.602 seconds
Artifact size	520 KB
Completion time shown by the UI	2026-08-16 20:44:20

Tier	Correct
T0, unscored pure-multiplication diagnostic	40/100
T1-T9	100/100 on every tier
T10	85/100

The UI's 99% is the rounded display of the scored result, 985/1,000 = 98.5%. The 1,025/1,100 count includes tier 0 and is 93.18% across all generated cases; tier 0 does not enter the scored overall accuracy. The frontier is T9 because T10 scored 85%, below the frontier threshold.

These hosted numbers were transcribed by the submitter from the completed SAIR Playground UI. No public run ID or receipt URL was available. Revision 16acb08ea46295ac7c45be697890fd99dbe8e5b5 was a later wording correction with identical model.py and weights.pt; it was not the evaluated revision.

The weights were trained through state width L=512. The same tensor values then passed an update-zero L=1024 qualification:

Check	Result
L=1024 screen, tiers 6-9, fixed and dynamic width	64/64 for every tier and mode
L=1024 confirmation, tiers 6-9, fixed and dynamic width	256/256 for every tier and mode
Primes below 64, fixed L=1024	40,954/40,954
Primes below 64, dynamic L=32	40,954/40,954

A separate zero-training width screen used the same tensor values:

Inference width	Tier	Fixed	Dynamic
1024	8	64/64	64/64
1024	9	64/64	64/64
2048	9	64/64	64/64
2048	10	63/64	63/64

The tier-10 miss matters. These counts are development evidence from fixed local case sets, not a proof of exactness.

Before publication, the packaged float32 inference path was also run through the current official interface on 110 separately generated cases. It was deterministic and scored 99/100 across tiers 1-10: 10/10 on tiers 1-9 and 9/10 on tier 10. Tier 0, which is an unscored pure-multiplication diagnostic, was 6/10. This was a local interface check, not a SAIR Playground result.

Artifact identity

Field	Value
Parameters	126,603
Packaged inference width	2,048 bits
Qualified source checkpoint	d296b711bb6a7faaa1dd81e05478cfa75f11071c42a8c36fbf60e758ee7eb407
Qualification receipt	80948748a41183a809faf282600cc0f8343691b6cdb4bebaac4f1f468df95651
Tensor digest	7d1768ae1260f750e0a80ec93d98f86a80d441e479ce29e0a8e21fd098c742a3

weights.pt contains only the state dictionary, architecture fields, and provenance. Optimizer, scheduler, and random-number-generator state were removed. Full machine-readable identities and counts are in provenance.json.

CUDA inference uses float32. On a 22-case non-edge precision check, BF16 and float32 agreed on 21 outputs but differed on one tier-10 failure. Both versions were wrong on that case, so this package keeps the precision used by the local qualification instead of claiming BF16 decision safety.

Interface

The SAIR entry class is model.MiniNeuralHorner, with output base 2. Inputs outside the packaged width and operand limits return [0] rather than invoking an untested fallback.

Code and research record: https://github.com/Robby955/neural-horner

License: MIT. Copyright (c) 2026 Robert Sneiderman.