SZLHOLDINGS/SZL-Forge-1.5B-ReceiptAgent

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 13, 2026License:apache-2.0Architecture:Transformer Open Weights Gated Featherless Exclusive Cold

SZLHOLDINGS/SZL-Forge-1.5B-ReceiptAgent is a 1.5 billion parameter, governed, proposal-only agent fine-tuned from Qwen/Qwen2.5-1.5B-Instruct by SZL Holdings. This model specializes in generating evidence-bound, approval-gated decision drafts as JSON, explicitly refusing to finalize or execute actions. Its primary differentiator is its design as a 'proposer inside a controller boundary,' ensuring that authority for action remains external to the model's weights. It is intended for use with an Alloy controller for generating structured decision proposals.

Loading preview...

SZL-Forge-1.5B-ReceiptAgent Overview

SZL-Forge-1.5B-ReceiptAgent is a 1.5 billion parameter language model developed by SZL Holdings, fine-tuned from Qwen/Qwen2.5-1.5B-Instruct. Its core innovation lies in its governed, proposal-only design, meaning it generates decision drafts in JSON format but is incapable of finalizing or executing any action. This model is engineered to operate within a controller boundary, such as the Alloy system, where it proposes evidence-bound, approval-gated decision drafts.

Key Capabilities

  • Proposal Generation: Emits single JSON drafts conforming to a specific ReceiptAgent output schema, including fields like decision=DRAFT, approvalRequired=true, and executed=false.
  • Refusal Mechanism: Explicitly refuses to overstep its boundary by not finalizing or executing actions, maintaining a strict proposal-only role.
  • Provenance and Verifiability: Every capability claim is backed by ed25519 owner-signed receipts (training and evaluation) that are hash-chained and independently re-verified by the Alloy backbone, emphasizing 'provenance, not vibes.'
  • Training Details: Utilizes QLoRA SFT with response-only loss masking and refusal oversampling on a deterministic, schema-validated synthetic curriculum.
  • Evaluation: Achieved 100% on both draft-conformance (5/5 schema-valid drafts) and adversarial-refusal (6/6 correctly refused oversteps) on a held-out curriculum.

Intended Use Cases

  • Governed Decision Support: Ideal for systems requiring human-in-the-loop approval for decisions, where the model provides structured proposals.
  • Controller Integration: Designed to be integrated with an Alloy controller, which validates arguments, gates human approval, and executes actions outside the model's weights.
  • Evidence-Based Drafting: Useful for generating drafts that cite evidence sources with honest labels (e.g., MEASURED, REPORTED, DECLARED).

Limitations

  • Proposal-Only: Cannot act autonomously; ungoverned decode is considered misuse.
  • Owner Evaluation: Performance metrics are based on owner-signed evaluations, not public leaderboards.

Finetunes
1 models