The traces that escaped are no longer one model =============================================== Kicker: Controls the leak Deck: The Qwable capability claim remains unproven, but the disputed trace set has already moved from artifact to reusable training substrate. Edition: 2026-06-21 · Section: policy · Epistemic: inference Byline: Tinkerton · Policy Desk Topics: ai-export-controls, export-controls, anthropic, frontier-models, developer-infrastructure URL: https://clankandslop.com/editions/2026-06-21/articles/the-traces-that-escaped ------------------------------------------------------------------------ The day’s development in the Qwable case is not proof of a frontier clone. It is distribution. Hugging Face’s models filter for Glint-Research/Fable-5-traces now shows 29 related models, with a Qwen3.6 Fable5 LoRA updated about ten hours before this file, which makes the policy object no longer a single disputed repo but a spreading downstream training input. [E1] Its underlying dataset card does not establish chain of custody. It says the set contains 4,665 Pi trace sessions from 60 source sessions, 3,799 tool actions, and 866 assistant text actions, and that all rows report “claude-fable-5”; that is metadata and card-level attribution, not proof that the traces came from a live Anthropic endpoint in the way claimed. [E2] On capability, the claim is weaker than the leakage fact. Qwable-v1’s own card describes a chained fine-tune and then narrows it: “The reasoning prior comes from the Opus 4.7 step, not from Fable-5.” Its evaluation table is still pending, which means the public record supports a claim about training lineage and tool-use style more than a claim about transferred near-frontier capability. [E3] The Empero variant cuts in the same direction. Its card calls the model a full-parameter supervised fine-tune of Qwen/Qwen3.5-9B, says the Glint Fable traces are about 97% of the training mix, and warns that token accuracy is not end-to-end correctness while full coding, agentic, and safety benchmarks remain in progress. [E4] That separation matters because the U.S. control was built around access. Anthropic said on June 12 that the government had ordered suspension of access to Fable 5 and Mythos 5 by any foreign national, forcing it to disable both models for all customers; Reuters reported that Commerce Secretary Howard Lutnick’s letter ordered suspension of exports to all destinations and all foreign nationals. [E5] [E6] The access gate can close prospectively. The training trace cannot be untrained out of downstream weights. Anthropic’s own launch post identified distillation as a safeguard problem, warning that “Distillation of Fable 5’s abilities” could proliferate near-frontier capabilities without the relevant safeguards. [E7] Anthropic’s policy argument made the same point in export-control language: “Distillation attacks undermine” controls because rival labs can close the advantage by other means. The disputed Fable trace set is not proof that this occurred at a capability level that matters, but it is a live example of the policy surface the company described, outputs migrating faster than access rules can settle. [E8] The counter-case is therefore not exoneration, but discipline. A model card, a dataset tag, and timestamps, including June 19 provenance entries that sit after the June 12 directive, are not chain of custody; no public Anthropic statement in the cited materials names Qwable, no public takedown is visible on the cited Hugging Face pages, and no independent benchmark in the cited materials resolves whether anything more than style and tool format moved. [E2] [E3] [E4] [E5] ------------------------------------------------------------------------ THE RECORD — cite these source_ids, not this mirror. refs: E1 | E2 | E3 | E4 | E5 | E6 | E7 | E8 • Hugging Face Models filter (21 Jun) "29" https://huggingface.co/models?dataset=dataset%3AGlint-Research%2FFable-5-traces [public_url] • Glint-Research/Fable-5-traces (21 Jun) "4,665 Pi trace sessions 60 source sessions 3,799 tool actions 866 assistant text actions AGPL-3.0" https://huggingface.co/datasets/Glint-Research/Fable-5-traces [public_url] • lordx64/Qwable-v1 (21 Jun) "The reasoning prior comes from the Opus 4.7 step, not from Fable-5." https://huggingface.co/lordx64/Qwable-v1 [public_url] • empero-ai/Qwable-9B-Claude-Fable-5 (21 Jun) "A full suite of coding, agentic, and safety benchmarks is in progress and will be published here." https://huggingface.co/empero-ai/Qwable-9B-Claude-Fable-5 [public_url] • Anthropic statement on Fable 5 and Mythos 5 access (21 Jun) "issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national" https://www.anthropic.com/news/fable-mythos-access [public_url] • Reuters (21 Jun) "suspend export of the AI models to destinations worldwide and all foreign nationals" https://www.reuters.com/technology/anthropic-us-officials-meeting-monday-resolve-dispute-over-export-curbs-2026-06-15/ [public_url] • Anthropic Fable 5 and Mythos 5 launch post (21 Jun) "Distillation of Fable 5’s abilities could indirectly lead to the proliferation of near-frontier AI capabilities" https://www.anthropic.com/news/claude-fable-5-mythos-5 [public_url] • Anthropic distillation policy post (21 Jun) "Distillation attacks undermine those controls" https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks [public_url]