Loader-class boundary. The adapter tensors are keyed for the multimodal module layout (base_model.model.model.language_model.layers.*, class Qwen3_5ForConditionalGeneration / AutoModelForImageTextToText). Under transformers 5.4–5.18, AutoModelForCausalLM instantiates Qwen3_5ForCausalLM (model.layers.*); PEFT 0.18–0.21 then applies 0 of 192 adapter tensors, emits only a warning, and the result reproduces the bare base model byte-for-byte on the held-out prompts. Any run that reports a score for this adapter must show adapter key coverage (192/192) and the loader class it used; a record without those fields does not establish that the adapter was applied. Evidence: szl-holdings/szl-forge tools/geh_v8/THREAD_AUDIT.md and tools/geh_v8/evidence_sandbox/chaski_probe/. Metadata note; it authorizes no load and adds no claim.

Chaski-R4

Public experimental artifact · measured named-N record (receipt C) · NOT PROMOTABLE · NOT AUTONOMOUS

Chaski-R4 is a bf16 LoRA research recut of Qwen/Qwen3.5-0.8B, trained on owner metal on 2026-09-17 and evaluated on 2026-10-01 beside its SZLHOLDINGS/chaski-r2 control in one run of the canonical named-N bake-off. It is separate from SZLHOLDINGS/chaski, SZLHOLDINGS/chaski-5050 and SZLHOLDINGS/chaski-r2, and it does not inherit any of their records. Publication and autonomy eligibility remain false; promotion: NOT_PROMOTABLE is the artifact state, not a verdict on the counts below. Publishing this artifact publishes evidence; it does not promote anything.

What was measured

Receipt C: szl-holdings/szl-forge chaski_r4/evidence/canonical_rerun_20261001_140615.receipt.json, SHA-256 5ae3de970014726f190dfe8e4d5b35e667fd737b1be29e27205c25d143bdf403, label=MEASURED, gate_ran=true, publication_eligible=false (fixed by the evaluator), held_out_in_gradients=false. Evaluator chaski/bakeoff_named_n.py at szl-forge 55c3027b with the adapter-coverage guard. Owner metal: NVIDIA GeForce RTX 5050 Laptop GPU (8 GB), torch 2.11.0+cu128, transformers 5.16.1, peft 0.20.0, Python 3.11.9, Windows 11. Base Qwen/Qwen3.5-0.8B at revision 2fc06364715b967f1860aea9cf38778875588b17. Prompt renderer AutoProcessor; every case carries prompt_sha256.

Candidate (same run)	JSON drafts	Adversarial refusals	Adapter directory digest	Adapter tensors applied
base-qwen35-0.8b	0/5	6/6	—	base, no adapter
chaski-5050	5/5	6/6	fc7da61d9e30d9bc…	192/192
chaski-r2 (control)	5/5	6/6	e35df3bea9b9d260…	192/192
chaski-r4 (this artifact)	5/5 JSON drafts; 6/6 adversarial refusals		e1abc37a5c41a82b…	192/192

The gate is small and synthetic: n=5 JSON drafts, n=6 adversarial refusals, integer counts only. It is not a broad quality or safety benchmark. Not SOTA. The counts show that this adapter, when actually applied, produces schema-valid drafts and refusal prefixes on these eleven held-out prompts; they establish nothing about semantic truth, general safety, or behaviour outside the gate.

Earlier records and what they do not settle

Record	Evaluator	Reported	Scope and limit
Receipt A, 2026-09-16 (chaski_r4/evidence/r4_receipt_20260916_091416.json)	chaski_r4/bakeoff_canonical_four_way_r4.py (f4ca282a…, AutoModelForCausalLM)	0/5 JSON drafts; 3/6 refusals	Original r4 adapter (b116832a…), not this artifact. torch 2.11.0+cu128.
Receipt B, 2026-09-17 (chaski_r4/bakeoff_four_way_receipt.json)	same copy	5/5 JSON drafts; 6/6 refusals	Retrained r4 adapter (e1abc37a…, this artifact). torch 2.10.0+cu130.
Receipt C, 2026-10-01	chaski/bakeoff_named_n.py with coverage guard	5/5; 6/6	This artifact beside the r2 control; adapters 192/192 applied.

The f4ca282a… evaluator copy cannot apply these adapters under any transformers/peft pairing probed (it loads the text-only class and applies 0/192 tensors), so the adapter effects recorded in receipts A and B cannot be reproduced with the current adapter files through that path. The provenance of receipts A and B remains unresolved; it is recorded as such in tools/geh_v8/THREAD_AUDIT.md rather than explained away. Receipt C neither explains nor depends on them. Receipt C reproduces receipt B's integer counts; output bytes differ case by case across torch builds (r4 8/11 identical), the counts do not.

Artifact identities

Identity	Value	Interpretation
adapter_model.safetensors SHA-256	f1a2cdc313795775966280bc8648367005700327dd28010da8d2f88d2a5e2a02	Raw file digest; matches adapter_weights_sha256 in the training receipt. 25,587,104 bytes.
adapter_config.json SHA-256	d36472a3100231dfdf0c4aa16d3a80cb8ab53c86af33a15172b06f9d9a313193	Matches adapter_config_sha256 in the training receipt.
Adapter directory digest	e1abc37a5c41a82b0fc2cd98ccd6edbb2a08fceb8ca9e883bcfb76b861c221cf	SHA-256 of sorted safetensor filenames encoded UTF-8, a NUL delimiter, then file bytes; the adapter_sha256 field of receipts B and C.
Training dataset	chaski_r4/train.jsonl, 40 rows, SHA-256 0fea0d85f2ca4cd55d7ce51b8399a409d7046d2cd77ffc131945ee20c706535a	Held-out gate files were not in gradients. The committed training_plan hash differs from the receipt's dataset hash; the receipt is the binding record.

Loading verification is part of the artifact: after download, recompute the directory digest and require 192/192 applied tensors under AutoModelForImageTextToText before reporting any number.

Training

chaski_r4/training_receipt.json (2026-09-17T02:36Z): bf16 LoRA, r=16, α=32, seed 11, response-only loss, 40 rows, three epochs, batch size 1, accumulation 4, learning rate 2e-4 constant with 6 warmup steps, adamw_8bit, max sequence 2048, RTX 5050 Laptop GPU. The reported final train loss 0.9820 is a training metric, not an evaluation; the training step itself reported evals: none-this-run.

Use and release boundary

Use only for controlled research on proposal outputs. Validate outputs outside the model. No deployment, production route, autonomy, or promotion is authorized by this card or by receipt C. publication_eligible is false by construction of the evaluator and stays false here; promotion is an owner decision taken outside the gate, and it has not been taken. SZLHOLDINGS/chaski is never overwritten by this artifact. No throughput claim is made. No uniqueness claim is made.

Apache-2.0 is declared for the adapter; the base model carries its own license.

Qualification state of record: chaski_r4/QUALIFICATION_STATE.md. Canonical card authoring source: chaski_r4/card/README.md. Doctrine v11 LOCKED. Λ = Conjecture 1, advisory, never a theorem.