common/source/starBenchserver/benchmarks/simBenchmarks/RoboDojo/ACTION_VALIDATION.md

Action validation and rejection feedback

Validation is not task scoring. Native RoboDojo still decides success. Input checks use only submitted values and public contract constants, never object truth, hidden predicates or hypothetical simulator outcomes.

Requirement ownership

Requirement	Owner/source	Configurable?
EEF16: two [x,y,z,qw,qx,qy,qz,opening] targets, finite numbers, unit wxyz quaternions, opening [0,1]	Benchmark Direct v2 EEF interface; norm tolerance 1e-4	Mandatory, versioned input validity
Fixed IK and per-control numerical bounds	Benchmark robot-only low-level conversion	Another declared benchmark track is required to change it
Per-decision 5 cm / 0.35 rad, binary opening 0 or 1	Current model's AgentPolicies/gpt6_astra/ action guard	Optional, enabled by default
Response JSON fields, repeated-target count, memory/note lengths, safety-stop representation	Selected model's output/memory protocol	Declared in model configuration, not native scoring

The 5 cm / 0.35 rad guard and 1–5 repeated controls were settings of the first Codex EEF method, not universal RoboDojo limits. Another model may declare its own preferences while respecting the same mandatory benchmark action contract. The benchmark accepts continuous opening commands; binary opening is only a model preference. Direct v2 aligns its target tolerance with GPT-as-Policy Direct at commit 8f3d362b077d8efb77e2a7274d5b2c20e2243846.

The accepted submitted quaternion is retained unchanged in action records; conversion to a rotation uses SciPy's quaternion interpretation, as upstream does. This is not repair of a target rejected by its declared tolerance. The v2 fixed DLS clips in float64 and casts the resulting command to float32, exactly as the reference does. It does not apply v1's extra one-ULP inward adjustment or revalidate converted joint targets against the old RGB/joint14 decimal bounds. Native soft-limit intersection and per-control DLS bounds still apply before that documented numerical cast.

Reject a command without ending the episode

New services advertise action_rejection: starvla-rsi.action-rejection/v1. A deterministic pre-execution rejection returns a closed issue code, declared field, finite scalar actual/limit, executed: false, and the unchanged next control sequence. The session remains open. No native control, new camera capture, reward evaluation, reset or automatic replay occurs.

A corrected command can use that same control sequence. The model's driver returns the issue through the next normal policy call with the last acknowledged observation. That call and its available usage are counted; an invalid target adds zero executed action decisions, controls or physics ticks. Any previously acknowledged valid prefix remains real execution. Invalid output does not become an executed history entry or a successful memory update.

Lost acknowledgements, native exceptions, corrupt responses, authentication failures and uncertain execution still fault the stream. Native end, valid safety stops and public resource limits remain terminal. The service never repairs, clips, normalizes or substitutes a target, calls a model, or supplies task-solving feedback. Public rejection data excludes raw provider/native error text, arbitrary JSON keys, credentials and hidden state.

Model configuration and budgets

The Codex v2 client defaults to feedback handling with enabled extra guards. --no-action-guard disables only extra motion/binary-gripper preferences. Mandatory validity checks remain. --invalid-action-handling terminate explicitly selects terminal rejection. Choices are frozen before execution and recorded with the method version.

Feedback mode requires a finite policy-call budget, since invalid outputs do not advance the simulator horizon. The model driver declares this budget with --max-policy-calls; the current RoboDojo service does not assign a model-call allowance. Rejections consume that same call budget, with no hidden extra messages. The native 1100-control horizon is a separate benchmark limit.

Frozen codex-eef/v1 attempts retain their original terminal-error handling. Keep v2 and repair-round results separate; never replace first-panel failures. The model-method version codex-eef/v2 and the fixed benchmark protocol are independent version namespaces. Bind both sides before any attempt. CPU fixtures validate interface/accounting mechanics, not native capability or SR.