khipu-r3

<!-- szl:artifact-identity-reconciled -->

Artifact identity (noted 2026-09-30). Besides the LoRA adapter, this repository's root carries a merged full-precision checkpoint (model.safetensors + config.json, loadable with transformers) produced by the receipted CPU merge of this adapter into its declared base (merge_receipt.json). The adapter remains the artifact of record for every figure on this card; the merged bytes carry no separate held-out receipt and add no claim. Metadata-only note.

SEPARATE RESEARCH ADAPTER · TRAINING RECORD · CANONICAL HELD-OUT GATE FAIL · NOT PROMOTABLE

SZLHOLDINGS/khipu-r3

What this is

A separate laptop Blackwell research adapter based on Qwen/Qwen3.5-0.8B. The training_receipt.laptop-blackwell.json records bf16 LoRA with r=16 and alpha=32, 23 training rows from train.jsonl and train.abstain.jsonl, and train loss 0.4397. That loss is a training metric, not an evaluation.

Evidence and artifact identity

Reviewed Hub revision: 89807a2680e2400bbbc7523772aa6ff4b86a7988. The repository contains adapter_model.safetensors and model.safetensors; the historical training receipt's weights=LOCAL and push_to_hub=false describe that training step, not current file absence.

There is no eval_report.json in this Hub snapshot, and no committed Hub held-out receipt supporting the former README claim “grounding 4/5, abstain 0/6.” That unsupported claim is withdrawn.

The separate post-publication gate receipt names Hub revision 89807a2680e2400bbbc7523772aa6ff4b86a7988 and reports 1/5 on the canonical held-out suite in transformers-cpu mode, with overall FAIL, heldout_passed=false, productionAuthorization=false, and publicationAuthorization=false. This recorded canonical result does not substantiate the former 4/5 + 0/6 claim. It records refusalPassed=false and a refusal regression: rate 0.2 against baseline 0.4. The receipt retains runtime warnings, and earlier execution failures remain in the canonical gate archive. Therefore “no held-out evaluation exists anywhere” would also be inaccurate: the canonical source contains this later failed evaluation.

The merge_receipt.json reports a 2026-09-14 repair of a prior empty base-copy merge: 96/96 adapter modules applied, maximum weight delta 0.0013806568458676338. A nonzero merge delta does not establish behavioral quality.

What the evidence does and does not establish

The training_receipt.laptop-blackwell.json retains evals=none-this-run, publication_eligible=false, and autonomy_eligible=false. Its training-run adapter digest is receipt-reported and is not asserted as a direct hash of the currently published adapter file. Exact current artifact qualification does not follow from a training receipt, a merge receipt, or file presence. The later canonical gate names the reviewed Hub revision and returns FAIL. The provenance.json also retains blocked artifact-lineage, consent, privacy-review, deployment, and served-revision boundaries.

Release, autonomy, and deployment boundary

Not promotable. publication_eligible=false and autonomy_eligible=false remain unchanged. No production, deployment authorization, autonomous action, approval authority, or house-lab pin follows from these records. Lambda uniqueness remains Conjecture 1: open and advisory, never a theorem.

Relationship and source of truth

This SKU does not overwrite SZL-Khipu-1.5B, SZL-Khipu-1.5B-GGUF, KHIPU-R2, brain-navigator-r2, or SZL-Khipu-1.5B-BrainNavigator.

Use the immutable Hub training and merge receipts and the canonical failed evaluation receipt above. The historical source card and Khipu curriculum source provide context; their prose is not a substitute for an artifact-scoped evaluation receipt.