What we can change — and what we train on
The submission-controls and dataset investigation, brought into the visualizer. Checked November 12, 2026. This page explains saved evidence; it does not start experiments or connect to Keith’s running jobs.
Submission controls
We control more than the prompt and an adapter, but we do not replace the competition’s execution system.
| Submission item | Supported changes | Boundary |
|---|---|---|
| Agent and helper configuration | Choose a provided model, instructions, enabled tools and adapter. Arrange sequential, parallel or looping agents; call a helper as a tool; pass recorded outputs through state. | Models, tools and callbacks resolve through organizer-controlled registries. A configuration is not an arbitrary replacement runner. |
| Evaluation budgets | Set per-task time, tool-call count, model-turn count and command timeout through the supported evaluation configuration. | These settings do not override the competition’s outer resource limits. |
| Prompts, skills and sampling | Change strategies, examples, checklists and helper roles. Configure temperature, top-p, top-k, presence/frequency penalties, seed, stop strings, output tokens and thinking settings. | The inspected schema does not include min-p. Generation constraints and the serving backend still apply. |
| Adapters | Supply compatible low-rank adaptation (LoRA) weights and select a discovered adapter for an agent. | The supplied base model remains the loading foundation; arbitrary full-checkpoint replacement is not a verified submission route. |
Technical evidence: installed official adk-submission 0.2.13 (schema, compiler and server) and swegemma 0.2.11 (evaluation configuration). These are verified installed versions, not a guarantee that the hosted evaluator has never changed. Evidence record and source locations.
Can we change or zero the weights?
Change the effective model through a compatible adapter: yes. Overwrite every supplied base tensor: no supported route verified. Zero adapter tensors remove that adapter’s contribution; they do not zero the base. A low-rank adapter generally cannot exactly cancel an entire full-rank base matrix. A hand-built adapter and a replacement model checkpoint are different proposals.
Our documented training recipe updates a rank-16 adapter over a frozen, four-bit E4B base. This describes the current documented recipe, not a universal mathematical limit on every possible adapter.
Can we put prompt injections in our own prompt?
We can change our agent’s instruction text, including role framing, examples and unusual attempts to redirect its behavior. In our own instruction field, that is prompt design. Saying “ignore the budget” cannot itself change the externally enforced budget, tool registry or grader. No successful bypass was demonstrated by this investigation.
Repository-specific caution: the team’s packager rejects curly braces in system prompts because the agent framework interprets them as session-state substitutions. Do not assume arbitrary pasted text will survive compilation unchanged.
Keith’s training data
The task catalog contains genuine archived post-training issues. It is not the current reinforcement-learning wave. Keith said he plans to upload his traces and dataset; today’s exact issue list and planned 31B training manifest have not been verified here.
The documented sources are SWE-rebench, SWE-rebench V2 and repaired SWE-rebench V2.1 RL: real Python repository issue-repair tasks. Each task pairs an issue with a repository snapshot, execution environment, reference fix and tests used to grade attempts. This is not simply a collection of question/ideal-answer pairs.
Curation checks that the bug reproduces, the reference fix passes, alternate correct fixes can pass, tests are useful and repeatable, and grading agrees with the competition harness. Evaluation tasks and competition repositories are kept outside the training pool.
Training generates repair attempts and rewards a resolved task with one; otherwise zero. General wave selection keeps tasks the current policy sometimes solves rather than always or never solves.
Latest committed wave report — historical snapshot
The November 10 wave-six report describes 300 fresh tasks, each selected at 40–60% success under the wave-five policy, without reused tasks. Its recorded configuration is 200 updates with 32 tasks × 8 attempts per update. A report of a configured wave is not proof that all planned updates completed, nor a live job status.
Open the archived training-task catalog → Search original issue text, filter repositories and recorded training-dataset membership, and open an issue to inspect its provenance. Current wave tasks can be imported once their authoritative manifest is available, without replacing the archive.