Ornith-1.5 moves the RL loop upstream, and the hard job becomes reward design

Ornith AI released Ornith-1.5 on August 19, 2026, extending the self-scaffolding framework from Ornith-1.0 into a closed loop in which the model proposes its own training tasks, generates the scaffold and evaluation harness for each one, and produces the rollouts it then learns from ^[1][2]. The consequence is a shift in where the scarce engineering effort sits: Ornith AI's own framing is that this reduces dependence on hand-built tasks while making reward design even more consequential ^[3]. Ornith-1.0, shipped in June, learned to write the scaffold around a coding task, meaning the instructions, tools, decomposition and orchestration that carry a model through a longer job ^[4]. Ornith-1.5 adds the assignment itself. Each training cycle begins with an environment or codebase, high-level instructions about the type of problem wanted, and the model's own history of solved problems ^[5]. The system proposes a harder task, generates or refines the scaffold, and produces a rollout conditioned on both, with reward propagating back across all three stages so that question, harness and answer improve together ^[2][6]. The scoring is where the design shows its hand. Task reward multiplies three signals: whether task and scaffold form a valid and verifiable environment, whether difficulty sits near the current capability frontier, and whether the task is novel against work already generated ^[7]. Frontier difficulty targets a 0.2 empirical success rate, so a task loses value to the generator once the model clears it reliably ^[8]. That implies roughly four rollouts in five are expected to fail ^[23], which is the point: hard enough to teach, with enough successes left to train on ^[9]. Validity acts as a hard gate that zeroes out malformed tasks, and all three stages are optimised with GRPO, according to Ornith ^[10]. The harness gets its own reward for matching the assignment, measuring solution quality accurately, and resisting reward hacking ^[11]. This is the load-bearing part. A system that writes both the exam and the grading script can manufacture easy points through malformed tests, hidden shortcuts or criteria that reward the wrong behaviour ^[12]. It is also the part with the least public detail: the official Ornith-1.0 repository documents benchmark harnesses, anti-hacking filters and resource settings, but runtimewire reports no substantiation of a more specific evaluation protocol involving measures such as Git-history removal or network blocking ^[20]. One word deserves deflating. Self-improvement here describes the training procedure, not the shipped artefact: the downloaded model does not continuously retrain itself on a user's phone or workstation, and Ornith runs the loop during development before publishing weights ^[21]. The release covers a 397B mixture-of-experts flagship, a 35B MoE model activating 3B parameters per token, and a 9B dense model with a quantized mobile build for iPhone and Android ^[13]. On the company's published tables, averaged over five independent runs, the 397B scores 85.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, which Ornith reports as on par with Claude Opus 4.8 at 85.0 and 59.0 ^[25]: ahead by 0.1 on one, behind by 3.0 on the other ^[26]. The 35B reaches 68.5 on Terminal-Bench 2.1 and 79.0 on SWE-Bench Verified, the 9B 47.0 and 70.6 ^[14], and the flagship posts 92.8 on GPQA Diamond and 86.6 on BrowseComp ^[22]. Ornith AI documents local deployment for the 9B and publishes an MLX build for Apple hardware ^[19]. The lineage is consistent with the bet. Ornith-1.0 shipped in June 2026 in four variants under an MIT license with weights on Hugging Face, post-trained on Gemma 4 and Qwen 3.5 checkpoints, and introduced the scaffold as a learnable object co-evolving with the policy ^[15][16]. DeepReinforce has published reinforcement learning optimisation work openly before, including CUDA-L1 and the IterX agent loop, with reward hacking a running concern ^[17]; runtimewire describes a related kernel-optimisation effort that used execution speed as the reward but names it CUDA-L2 ^[18], a discrepancy to keep in mind when tracing the research trail ^[24].