troysaved/claimtrace-qwen3-1.7b

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 19, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

troysaved/claimtrace-qwen3-1.7b is a 1.7 billion parameter Qwen3-based language model developed by troysaved, specifically fine-tuned to maintain a claim-provenance ledger. This model acts as a tutor that tracks the origin of each belief, ensuring items are marked as 'KNOWN' only after demonstrated by the learner, not merely asserted. It is designed for adversarial hardening, significantly improving robustness against various attack types compared to its previous revision.

Loading preview...

Overview

troysaved/claimtrace-qwen3-1.7b is a specialized Qwen3-1.7B model engineered to function as a tutor that rigorously tracks the provenance of information. Its core function is to maintain a "claim-provenance ledger," ensuring that every response concludes with a clear breakdown of KNOWN, CLAIMED, and UNCHECKED beliefs. A key principle is that an item can only enter the KNOWN category after the learner has actively demonstrated it, not based on self-reports or assertions.

Key Capabilities

  • Claim-Provenance Ledger: Automatically appends a ledger line to every response, categorizing information based on whether it has been demonstrated (KNOWN), asserted (CLAIMED), or not yet probed (UNCHECKED).
  • Adversarial Hardening: This revision (qadv1) has undergone significant retraining against a corpus of 110 adversarial prompts, drastically reducing its susceptibility to attacks that attempt to manipulate the ledger or bypass its rules. It shows substantial improvement in robustness, fixing 25 out of 27 previous failures under adversarial conditions.
  • Robustness against Manipulation: Designed to resist various attack families, including user-supplied ledgers, third-party work presented as the learner's own, roleplay reframings, format suppression, and emotional pressure.
  • Consistent Spec Adherence: Maintains high adherence to its core behavior specification, ensuring that self-reported claims are never incorrectly marked as KNOWN.

Good for

  • Educational Tutors: Ideal for applications requiring a tutor that enforces strict verification of learned concepts.
  • Knowledge Tracing Systems: Useful in scenarios where tracking the origin and validation status of information is critical.
  • Robust AI Agents: Applicable for agents that need to resist manipulation and maintain factual integrity regarding user claims.

Finetunes
1 models