🜂 SZL Holdings estate artifact — the estate that receipts its own thinking.
The second remediation attempt. DPO pass on the FreeEnergy model (which showed the best confabulation signal: 3/6 vs 4-5 for all others). Hypothesis: the free-energy annotation (accuracy-complexity tradeoff training) + the DPO counterweight may produce better citation grounding than either alone.
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.
from transformers import AutoTokenizer, AutoModelForCausalLM
tok = AutoTokenizer.from_pretrained("SZLHOLDINGS/SZL-FreeEnergy-0.8B-DPO-v1")
model = AutoModelForCausalLM.from_pretrained("SZLHOLDINGS/SZL-FreeEnergy-0.8B-DPO-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.
The citation probes are about to run. If confabulation = 0/6 → citation_grounding = 1.0 → first GO on the Λ-spine gate. If > 0/6 → honest FAIL, published.
Model output is an untrusted proposal.
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 chainSZLHOLDINGS/frontier-54-runs), treat the figures below as not publicly verifiable. The MEASURED labels above are unchanged.
| Receipt (private) | What it records for this model |
|---|---|
freeenergy-dpo-reeval-2026-11-06.json |
Citation-probe re-gate: 3/6 confabulated, citation_grounding_pass false |
governed-validation-v3-2026-11-06.json |
Governed validation v3: 3/6 confabulated, same as the FreeEnergy base. The "about to run" status above predates these receipts |
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.
| Receipt | schema | date | Measured headline | sha256 |
|---|---|---|---|---|
| `receipts/eval/freeenergy-dpo-reeval-2026-11-06.json` | szl.remediation-reeval/v2 | None | confabulation_count=3; citation_grounding_pass=False | `eb5fa17231b38901…` |
Verbatim honest notes (excerpted):
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_freeenergy_0_8b_dpo_v1.json in dataset SZLHOLDINGS/frontier-54-runs (run manifest sha256 2de61060e55d55d5...). Honest scores only - no curated claims.
c4652c48d7a00a7cfed71b66a09360871a07429f. The prior card remains in revision history.UNAVAILABLE.receipts/eval/freeenergy-dpo-reeval-2026-11-06.json.apache-2.0. This edit changed no license and asserts none beyond the observed tag.UNMAPPED). No source-side binding is claimed.238d093417cbe872f320ded35c3a003d16bfb4f0. The prior card remains in revision history.UNAVAILABLE.receipts/eval/freeenergy-dpo-reeval-2026-11-06.json.apache-2.0. This edit changed no license and asserts none beyond the observed tag.SZLHOLDINGS/SZL-FreeEnergy-0.8B-v1