A rank-32 LoRA on one Mac Studio takes Forge manifests from 0 of 25 to 14 of 25

Two gates produced the app-writing numbers, and they check different things. Atlassian's validator checks a manifest against the platform's schema. The TypeScript compiler checks that the code builds. The untouched base got 0 of 25 briefs past the validator; tuned, it got 14 ^[4]. Complete apps that also compile went from 0 of 25 to 12 ^[5]. If the compiling apps are a subset of the accepted manifests, two briefs yielded a manifest Atlassian would take and code that would not build ^[2]. Eleven yielded no manifest the validator would accept at all ^[3]. "It is a strong assistant, not a compiler," the author wrote ^[8].

Before tuning, the base invented a manifest format and imported packages that do not exist ^[11]. That failure mode is what makes a schema validator a useful scorer: it catches confident invention. It grades the manifest against the schema; whether the app does what the brief asked for goes ungraded.

For 12 of 25, which is 48 percent ^[1], to mean anything against your own backlog, your briefs have to resemble his 25. And you have to accept schema acceptance plus a clean build as the definition of done. The training mix also included the model's own generated apps, kept only when they passed the validator, the allow-list and the TypeScript compiler ^[19]. The post does not disclose whether the 25 briefs were held out of that mix ^[5]. Because that filtering used the same gate that later scores the model, 14 of 25 is the best case for briefs like his.

The training cost is the part any team can check against its own hardware. Rank-32 LoRA, frozen 8-bit base, 26 hours across three kept rounds, on a single Mac Studio M3 Ultra with 96 GB ^[6]. That averages under nine hours a round ^[4]. Only the adapter trains; the base stays frozen and quantised. Every round, including the one that was rejected, is published in the ledger ^[7].

The author said he wanted to call it the first ever Atlassian-trained model in the title, searched before he did, and found that it is not true ^[13]. Since late 2025 there have been several small Jira fine-tunes, including a 0.6B LoRA for tool-calling trained on synthetic conversations and an adapter on Qwen2.5-7B named for JQL with an undocumented training set ^[15]. Atlassian says it hosts fine-tuned open models inside its own VPC, and the author found no released weights ^[16]. What he will defend is narrower than the title he wanted: first open-weights model trained to write Forge apps, and first Atlassian-tuned model released with base-versus-tuned evidence on Atlassian's own validator ^[12]. "If someone shows me a prior Forge model, I will update this paragraph," he wrote ^[17].