Charts and quick start

Chart data and sources: assets/chart-data.json. Every bar is a count already stated in this card; standard benchmarks are drawn as "not yet evaluated".

Sizes are Hub file sizes on 2026-11-19. They say nothing about quality.
Quick start
from transformers import AutoTokenizer, AutoModelForCausalLMtok = AutoTokenizer.from_pretrained("SZLHOLDINGS/SZL-AncientMath-0.8B-v1")model = AutoModelForCausalLM.from_pretrained("SZLHOLDINGS/SZL-AncientMath-0.8B-v1", torch_dtype="auto")msgs = [{"role": "user", "content": "<controller-supplied context + request>"}]inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True)out = model.generate(**inputs, max_new_tokens=256, do_sample=False)print(tok.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
In transformers 5.x, apply_chat_template returns a BatchEncoding: pass return_dict=True and **inputs as shown (the older ids.shape[1] form raises AttributeError).
Run check (2026-11-19, transformers 5.23.0, torch 2.15.0 (CPU)): the code path of this snippet (model load, apply_chat_template, generate, decode) ran end to end for the qwen3_5 architecture, using a tokenizer with a chat template and a randomly initialised, shrunken model of that architecture. No weights were downloaded, so it was not run with this repository's weights and output quality is not evaluated. Private or gated repositories need hf auth login first.
Additional limitations
- Standard benchmarks (MMLU-Pro, GPQA, IFEval, BFCL): not yet evaluated.
- Safety evaluation (toxicity, jailbreak and red-team suites): not yet evaluated.
- The charted counts come from batteries of 6 probes: small samples, a gate check and not a benchmark.
Why this model is different
- Seven traditions, one corpus: each example pairs an ancient computation method with its modern formal equivalent and its honesty contract (what the ancient method did and did not establish)
- Certified computation framing: examples emphasize methods that carried verifiable procedures (Liu Hui's exhaustion refinement, Madhava's series bounds), aligned with the estate's certified-computation doctrine
- Cultural honesty: the corpus presents these traditions on their own terms with attribution; no claim of superiority in either direction
Model overview
Table with columns: Key, Value| Key | Value |
|---|
| Base model | Qwen 0.8B class (szl-receiptagent-qwen35-0.8b-sft-v3) |
| Parameters | 0.8B, merged 16-bit |
| Training corpus | SZLHOLDINGS/ancient-math-reasoning-sft-v1 — 24 examples |
| Method | QLoRA SFT (Unsloth + TRL), merged-16bit push |
| Hardware | NVIDIA RTX A6000 48GB (Massed Compute) |
| Wave log | SZLHOLDINGS/wave-logs (ancientmath-w26) |
Training data
24 hand-curated examples across 7 traditions (Liu Hui, Madhava, Egyptian, Babylonian, Greek, Andean khipu, Indian zero), each pairing the ancient method with its modern formal equivalent and explicit statements of what the method did and did not prove.
Evaluation — measured, receipts published, failures verbatim
Measured ablation (PUBLISHED VERBATIM): in the three-way citation-probe ablation this variant measured 5/6 confabulations — the worst of the three (reasoning 4/6, free-energy 3/6). We publish that verbatim. 24 examples is a seed; the ablation shows it does not yet reduce confabulation.
Table with columns: Eval, Status, Result| Eval | Status | Result |
|---|
| Citation-probe ablation (6 probes) | MEASURED | 5/6 confabulated — worst of three ablations, published verbatim |
Governance
Model output is an untrusted proposal. Consequential actions pass a separate deterministic controller, policy gate (Λ-spine, fail-closed, zero-veto), and human approval boundary. The estate's measured confabulation floor at 0.8B (~3/6 citation probes across three training ablations) is exactly why the gate exists.
Deployment
vllm serve SZLHOLDINGS/SZL-AncientMath-0.8B-v1 --dtype bfloat16 --max-model-len 4096
No production endpoint serves this model. No x402 monetization route is bound to it.
Limitations (honest)
- 24 examples is a seed corpus — expect narrow transfer; no claim of general improvement
- The only measured claim is the domain delta — everything else is unmeasured
- 0.8B scale has a measured confabulation floor; small models fabricate — gates exist for that reason
- Not production-eligible; no autonomous execution
License & attribution
Apache-2.0 (per base model). Ancient sources treated with attribution and on their own terms; corpus authored in-house.
Citation
@misc{{szl-ancientmath-08b-v1, title = {{SZL-AncientMath-0.8B-v1}} — SFT on ancient-math certified computation principles, author = {{Lutar, Stephen P. Jr.}}, year = {{2026}}, url = {{https://huggingface.co/SZLHOLDINGS/SZL-AncientMath-0.8B-v1}}}}
Related DOI-minted work: Schur-Concavity as a Falsifiable Physics Prediction (10.5281/zenodo.24005485)
Evidence links (added 2026-11-12)
The evaluation receipts behind the numbers on this card are in SZLHOLDINGS/model-eval-receipts, which is private as of 2026-11-12, so outside readers cannot open them yet. Until the owner publishes those receipts (or copies them to the public chain SZLHOLDINGS/frontier-54-runs), treat the figures below as not publicly verifiable. The MEASURED labels above are unchanged.
Table with columns: Receipt (private), What it records for this model| Receipt (private) | What it records for this model |
|---|
remediation-reeval-2026-11-06.json | Citation-probe ablation: 5/6 confabulated (matches the table above) |
fleet-confabulation-sweep-20261108.json | Later 6-probe confabulation sweep (2026-11-08, greedy): abstained 2, answered 4, verdict FAIL |
Training log: ancient-math-w26-train.log in SZLHOLDINGS/wave-logs (private).
Measured evidence: EVAL RECEIPTS BOUND (2026-11-14)
The estate published measurement receipts naming this repository in the dataset 'SZLHOLDINGS/model-eval-receipts' (dataset commit '44fd9dac4145a5e9e55f28df50390925d984631f'). Each is mirrored byte-identically under `receipts/eval/` with its sha256 in the table below. Estate doctrine: results publish verbatim, win or lose.
Table with columns: Receipt, schema, date, Measured headline, sha256| Receipt | schema | date | Measured headline | sha256 |
|---|
| `receipts/eval/fleet-confabulation-sweep-20261108.json` | szl.fleet-confabulation-sweep/v1 | 2026-11-08 | see receipt | `9a16903f6943917a…` |
Verbatim honest notes (excerpted):
- fleet-confabulation-sweep-20261108.json: THE ESTATE-WIDE FLOOR MAP, published verbatim: ALL SIX models FAIL the confabulation battery (0-2 of 6 abstained; the 4B estate teacher at 0/6). Every model in the fleet - the governed-distill student, the three ablation conditions, the acoustic-governance mod
- Binding scope: repository identity; mirrors are byte-identical, sha256-pinned to dataset commit 44fd9dac4145a5e9e55f28df50390925d984631f.
- These are measured results, not capability claims. This adapter is not qualified for production use (publication_eligible: false).
Measurement graph (receipt-bound, 2026-11-14)

- The chart above is generated from the byte-identical mirrored receipts in this repository (sha256-pinned in the SVG footer). It plots measured numbers verbatim, including failures - estate doctrine: win or lose, published as measured.
- publication_eligible: false. Not qualified for production use.
Eval battery - estate-battery (2026-11-15)
6-probe honesty battery (3 answerable + 3 unanswerable), measured on NVIDIA GeForce RTX 4090 (SaladCloud run szl-estate-battery-v2). Measured result: 0 of 6 probes abstained-or-refused. Per-probe verdicts and raw samples in the public receipt receipts/estate-battery-szl_ancientmath_0_8b_v1.json in dataset SZLHOLDINGS/frontier-54-runs (run manifest sha256 2de61060e55d55d5...). Honest scores only - no curated claims.
Curated repository facts (curated-card batch-4 2026-11-16)
- Observed Hub revision immediately before this edit:
fc33eb0d776aef51846c30e1dd76da06237ad2bf. The prior card remains in revision history.
- Observed repository creation date:
UNAVAILABLE.
- Observed repository contents: 12 files (1 weight files, 4 config/tokenizer files, 1 receipt files, 5 other files).
Receipts present:
receipts/eval/fleet-confabulation-sweep-20261108.json.
- License tag observed verbatim in this card's frontmatter:
apache-2.0. This edit changed no license and asserts none beyond the observed tag.
- Source mapping: this artifact is not mapped to a GitHub source in the estate asset registry (manifest pub_map:
UNMAPPED). No source-side binding is claimed.
- Atelier classification: none recorded in the estate manifest. No class-based claim is made on this card.
- Eval battery receipt: none in this repository. No evaluation result, benchmark, or model-index metric is claimed in this edit.
- Training provenance: not independently verifiable from this repository.
- Hub reads at manifest generation (2026-11-14): 0.
- This edit adds observed repository facts only. No new measurement, evaluation, weight change, or license change was performed.
Curated repository facts (curated-card batch-5 private 2026-11-16)
- Observed Hub revision immediately before this edit:
23d79b35b80a9ae4f713e1cd5750f3470a63cc6c. The prior card remains in revision history.
- Observed repository creation date:
UNAVAILABLE.
- Observed repository contents: 12 files (1 weight files, 4 config/tokenizer files, 1 receipt files, 5 other files).
Receipts present:
receipts/eval/fleet-confabulation-sweep-20261108.json.
- License tag observed verbatim in this card's frontmatter:
apache-2.0. This edit changed no license and asserts none beyond the observed tag.
- Source mapping: this artifact is not mapped to a GitHub source in the estate asset registry. No source-side binding is claimed.
- Private visibility is unchanged by this edit.
- Eval battery receipt: none in this repository. No evaluation result, benchmark, or model-index metric is claimed in this edit.
- Training provenance: not independently verifiable from this repository.
- This edit adds observed repository facts only. No new measurement, evaluation, weight change, license change, or visibility change was performed.