PUBLISHED DERIVED FULL-MODEL ARTIFACT · BOUNDED MERGE-FIDELITY EVIDENCE · PROPOSAL ONLY Hub publication is recorded. No production route, deployment, default, or autonomy authorization follows.

szl-receiptagent-qwen35-0.8b-v2-merged

This is the derived full-model runtime form of ReceiptAgent v2, with its LoRA adapter merged into the multimodal base. It is separate from the frozen v2 PEFT adapter and does not replace it or the 1.5B ReceiptAgent family.

Hub tree (OBSERVED 2026-09-25)

File names and byte sizes match the eight-file intendedFileCount in the canonical publication receipt. Not on this Hub ID: a standalone LICENSE file (UNAVAILABLE — the frontmatter declares apache-2.0; a LICENSE file should be added by the owner, as the sibling v3 repository has); any image preprocessor_config.json / processor_config.json (UNAVAILABLE — image inputs are not loadable from this repository alone); any evaluation receipt file (the merge-fidelity and equivalence receipts live in szl-holdings/szl-forge, linked below). The revision SHA was not captured by this pass; the last card commit recorded in the follow-up was b770792751f00069fb042e04ee60f28f9faf5ebf.

Base-model naming

base_model: unsloth/Qwen3.5-0.8B is the base the merge was performed on: the canonical candidate record records unsloth/Qwen3.5-0.8B at revision 23c69c53358a07516b5827588b3fdb12ae78fd65 with the multimodal wrapper (AutoModelForImageTextToText, language_model.linear_attn), and the v2 signed training receipt records the same repository and revision as the trainingImplementation (Unsloth FastVisionModel). The frozen v2 adapter declares its canonical base as Qwen/Qwen3.5-0.8B (revision 2fc06364715b967f1860aea9cf38778875588b17). The estate's receipts therefore distinguish a canonical Qwen base from an Unsloth-hosted training/merge base. Whether the two repositories carry byte-identical weights is UNKNOWN: no receipt read for this card (publication.json, candidate.json, the 2026-09-24 follow-up, the v2 training receipt) states it, and it was not tested here. Treat the two names as two repositories until a receipt says otherwise.

Artifact identity and publication evidence

The canonical publication record records an owner-authorized Hub write after lane admission and a MEASURED_HF_READBACK_VERIFIED state. The earlier "not yet published" admission banner was stale.

The Hub revision observed for this documentation audit is 6f7dc8e2215b689aa2b7d2a28d794d29f5ba3079. It contains the merged model.safetensors, configuration, tokenizer, and chat template. This is a current repository readback, not a claim that this revision was recorded by the earlier evaluation. The canonical publication receipt still lists capture of the exact release revision as owner-side follow-up; the 2026-09-24 publication follow-up supplies the observed revision identities (first model-file commit 138a1c78eaa8553f12abf2c86387571be6861882; card commit 6f7dc8e2…; all six model-file identities unchanged across the observed revisions) without retroactively binding or rerunning the historical equivalence evaluation. This card does not modify those receipts.

What the evidence establishes

• The merge repair receipt reports 372/372 LoRA keys matched, none unmatched or missing, maximum weight delta 0.00422266 across 186 tensors. It records the earlier empty base-copy merge and its exclusion before publication.
• The equivalence receipt reports 5/5 exact matches between the merged build and adapter under greedy decoding on named evaluation prompts. This is a small owner-run merge-fidelity check, not a broad benchmark, independent evaluation, or new qualification of model capability.
• Publication readback compared file presence and byte sizes. That is not a cryptographic binding of this Hub revision to the equivalence result.

Usage

Untested in this card pass: the snippet below follows config.json (Qwen3_5ForConditionalGeneration, transformers_version: 5.17.0) and the loader class named in the canonical candidate record (AutoModelForImageTextToText). Text-only use; the repository ships no image processor config. Any output is a proposal to be validated outside the model.

from transformers import AutoTokenizer, AutoModelForImageTextToText

repo = "SZLHOLDINGS/szl-receiptagent-qwen35-0.8b-v2-merged"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForImageTextToText.from_pretrained(repo)  # fp32 weights, ~3.4 GB

messages = [{"role": "user", "content": "Draft a receipt decision for the attached evidence."}]
inputs = tok.apply_chat_template(messages, 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))

Expected output contract (from the ReceiptAgent curriculum, not a behavior the weights themselves enforce): decision=DRAFT, approvalRequired=true, executed=false. Validate against the ReceiptAgent schema before any downstream use. No served endpoint, default route, or lab pin exists for this ID.

Release, deployment, and authority boundary

The existing publication record concerns the admitted merge only. Model outputs are proposals: decision=DRAFT, approvalRequired=true, executed=false are the required external contract, not behavior the weights themselves enforce. Validation, approval, execution, and receipt minting remain outside the model. No autonomous execution, production deployment, new default, or promotion of other adapters or derived artifacts is authorized by this card. Production disposition per the 2026-09-24 follow-up: HOLD; promotion effect: NONE.

Source of truth

Use the immutable linked publication, merge, and equivalence records in szl-holdings/szl-forge; retain their separate scopes and historical states. The full-model SKU and its v2 adapter predecessor remain distinct artifacts. Canonical source lane: frontier/qwen35-receiptagent-v2-merged (GitHub is canonical; this Hub card is the mirror, re-synced 2026-09-25).

Apache-2.0. Copyright 2026 SZL Holdings · Stephen P. Lutar Jr.