common/source/starBenchserver/docs/evaluation.md

Evaluation workflow

Start with the selected benchmark's README and EVALUATION_PROTOCOL.md. Installing this package does not install native simulators or grant asset licenses.

Choose a workflow

Central task scheduling is retained. Choose the path that is actually implemented; the two execution paths below are alternatives, not consecutive setup steps.

Purpose	Entry point	Current boundary
Service-based evaluation	rsi-bench serve → your trusted agent driver → result report	Recommended integration path; native runtime stays behind the benchmark server
Central case scheduling	python -m benchkit.campaigns prepare/run-case/summarize	Five registered plans; fixed local-adapter execution, not remote service orchestration
Inspect closed results	benchkit.reports.basic_result, optional interaction history	Reads records; never runs, retries or changes a task

The public service path is: select the protocol/schedule and deployment → start the service → connect a method → inspect its results. The actor receives only declared observations and robot feedback, never the driver's administrative client. Shared benchkit communication and output tools are not another per-action process.

Central scheduling remains available for its existing frozen execution path. It has not yet been integrated with remote server prepare/run-case. Do not substitute it for a service-based run while claiming the same isolation or provenance.

Select and launch a service

rsi-bench list
rsi-bench serve --benchmark simpler_env --help
# After following SimplerEnv's installation guide and setting the token:
rsi-bench serve --benchmark simpler_env \
  --parameters /absolute/benchmark-parameters.json \
  --seed 7 --episodes 1 --port 8765 \
  --output-dir /absolute/new-run/server

Without a console-script installation, use python -B -m benchkit.sim_server.cli instead of rsi-bench. The token is supplied privately through RSI_SIM_SERVER_TOKEN, never through the model prompt. See the wire protocol for authentication and remote forwarding. List and help do not start/reset a native environment.

The selector forwards launch arguments to the existing benchmark-owned entrypoint. There is no universal EEF control space or --full switch. A named route means an implemented development service, not installed assets or a validated full benchmark. The original per-benchmark python -m ...server remains valid.

Choose a plan, not a retry strategy

Plan	Sample schedule	Status
Fast	A preselected subset of cases, one attempt per selected case	Current campaign preparation records evaluation_mode: fast; service schedules are benchmark-specific
Full	Verified native suite, variants, seeds, resets and episodes per task	Benchmark-specific presets only; no universal full-schedule selector

RoboDojo declares its schedules in bench_server/protocol/task_catalog.json. Its official-source server configuration selects task_id: classify_objects (1100 controls per episode) or task_id: stack_bowls (800 controls per episode). Both offer panel: fast with five layouts; their panel: full schedules contain 50 and 25 layouts respectively, each once. The historical reference deployment still accepts only classify_objects. A full panel covers the selected task, not the complete RoboDojo suite. These fixed local schedules retain failed and unstable cases without the upstream batch evaluator's replacement behavior. The catalog is protocol data; runtime assets belong to environment/, and deployment paths/GPU selection belong to the operator's server configuration. Use rsi-bench serve --benchmark robodojo; the delegated benchmark-owned wrapper remains benchmarks.simBenchmarks.RoboDojo.server, with its adapter in bench_server/.

The current campaign CLI has benchmark-specific default cases and selected --task-scope expansions. It does not accept an arbitrary --mode full switch. Inspect the benchmark's declared plan module and generated manifest for the exact denominator. SimplerEnv, LIBERO, SIMPLE, RoboCasa365 and RoboCasa Tabletop keep the plan at bench_server/protocol/plan.py; their old evaluation packages are removed. Native planned repetitions are samples, not retries triggered by failure.

If development budgets differ from native conditions, label that track explicitly. Do not compare task-only coverage or a longer horizon with a native full score. Policy persistence preferences are a separate model configuration.

Connect an agent method

Select a complete agent method, not just a model label. The protocol-first path supports your own driver and decision loop; Python Policy is optional. Declare all model components/roles and configuration using the method contract. The reference driver accepts --method-file. Additional inference/planning does not expand the robot or environment interface.

anonymous_code_agent is an external mailbox with unknown provider/model identity. gpt6_astra supplies Native, persistent Direct v3 and Planned Codex methods; their compatibility and evidence differ. Its server driver defaults to generic Direct v3 with an explicitly selected benchmark. Numeric-interface compatibility is not native validation or universal robot capability. opus is an unconfigured placeholder. Profile dependencies and credentials are separate from benchmark installation.

For the mailbox profile, the actor receives only its opaque endpoint and permitted protocol context. Do not provide campaign directories, native workers, evaluator state or private scoring data. Mailbox-only instructions are not a substitute for enforced process isolation or a trusted provider-request meter.

Start the server following the selected benchmark's guide, then use the current Direct v3 driver command. Historical model examples named Direct v2 require their original frozen package, not today's active factory. The existing RoboDojo server protocol still uses its original Direct-v2 interface identity; a model-method upgrade does not rename that server protocol. Use its matching seed and named panel; a separately selected one-case development sample is not a full-panel score. Never copy its EEF space or budget to another robot. Select method and recording options before execution.

An independent driver can implement the wire protocol without joining the reference factory registry or using Python Policy. Keep claims, metrics, finalization and credentials on the trusted side; project only protocol-allowed inputs to the actor. See method integration. The optional benchkit run --sim-server command uses the same agent-side runner, not a separate benchmark-control implementation.

The small independent-driver example demonstrates its own loop and essential result records, without a registered profile. Its single-case/prefix-one scope is intentionally narrow; use the benchmark-compatible method driver for full panels and other control contracts.

Calls, executed decisions, controls and native physics ticks are different counters. Invalid-action feedback, motion guards and model network continuation are versioned method/track behavior; consult the Direct guide and action validation contract. No path automatically retries a physical action after an uncertain ACK.

The method chooses a chunk and its trusted driver sends each control separately. --replan-interval 5 in a method example is a configured prefix ceiling, not a server requirement to execute five controls per decision. A shorter chunk or native end can execute fewer. Native physics substeps and the total execution budget remain benchmark-owned. Existing decision-coupled tracks and any declared required_replan_interval still require matching configuration; do not silently override them when selecting a different method.

Native's model-call recovery is separate from case scheduling: it can wait and retry a classified transport failure at the same observation, up to ten total attempts before returning an infrastructure error. No decision/control is added until an action actually returns and executes. Preserve all transport attempts and available usage; never reinterpret model-request recovery as a task reset, native failure or permission to replay an uncertain robot action.

Inspect and report

Interaction recording is optional and enabled by default. The reference driver accepts --recording-mode minimal and granular overrides such as --no-record-videos. See recording settings for client retention, runtime scratch and native capture limitations. Do not change options mid-run.

By default, retained native videos stay in the service's output directory; the client keeps their server/session/path references, not duplicate downloads. Keep that server directory durable alongside the client result. Client --copy-server-media opts into another local copy. --no-record-videos affects client retention and does not disable native capture or delete server videos. Standalone frames are normally temporary; missing/incomplete video evidence or failed native close preserves them with a warning. See media ownership.

Generate a basic result even when no trajectory/video/source snapshot is available:

python -B -m benchkit.reports.basic_result \
  --run-dir /absolute/closed-run/client \
  --output-dir AgentPolicies/my_method/results/my_benchmark/report/my-run

This validates closed essential records, not provider identity or native replay. It preserves all selected attempts and failures; optional --metric KEY exposes an explicitly selected finite native scalar. Do not compare different protocols or method configurations as if they were only model changes. No aggregate cross-benchmark leaderboard is implied.

When sufficient interaction events were retained, an optional richer view is available for an explicitly selected closed client run:

python -B -m benchkit.reports.interaction_history \
  --run-dir /absolute/closed-run/client \
  --output-dir AgentPolicies/my_profile/results/my_benchmark/interaction/my-run

The destination must be new and outside the source run. Open its index.html directly in a browser; history.json contains the same structured view. The exporter performs no model request, simulation or retry and does not edit the original record. It checks closed event/result consistency, not full source integrity or provider identity. It accepts the generic client-run schema, not an arbitrary server/campaign parent directory; unsupported or inconsistent records fail explicitly. Redacted/incomplete histories may not satisfy its stronger event audit; use the basic result view instead. This is not a publication failure.

Each call shows the recorded policy input, note/output, proposed actions, acknowledged controls, rejection/errors and episode outcome. Calls, action decisions and controls remain distinct. Planned/unattempted counts refer only to the client-selected schedule; a larger server/native panel needs a separate denominator record. Camera names do not prove retained pixels; this first exporter copies no media and cannot reconstruct missing actions or raw chat. It does not prove backend consumption or final ACK delivery. Review recorded task text and notes before sharing. Benchmark-specific score, environment/asset provenance and video publication stay in their existing tools; they are not guessed from arbitrary nested metadata.

Report the selected denominator, attempted/completed/unattempted counts, successes, policy/protocol errors and infrastructure errors. Preserve every original failure. Report available usage/latency and mark unavailable provider accounting as unknown. Typed policy-call failures retain known usage and bounded receipt metadata without becoming successful decisions or controls. Request-attempt counts are not proof of completed inference. Record audits check consistency, not provider attestation. Do not merge different budgets, asset tracks or model identities into one score.

Central case scheduling

This case-level coordinator is retained, not just a historical reader. It assembles benchmark-owned plans, performs declared preflight, claims a case permanently, launches its frozen local entrypoint and summarizes status. SimplerEnv, LIBERO, SIMPLE, RoboCasa365 and RoboCasa Tabletop currently have registered plans.

The operator supplies the case index and device. There is no automatic resource allocator, remote bench-server binding or interrupted-service reconciliation. Future extensions must preserve selected conditions, all failed/error attempts and no replay after uncertain execution. A fresh directory or run ID does not authorize another attempt. Agent methods still own their decision loops.

These commands use internal source freezing and the reference profile registry. Neither is a prerequisite for protocol-first participation or basic result publication. Do not run the example commands below as additional steps of a service-based evaluation.

Prepare and inspect

From the repository root, the following prepares the SimplerEnv fast plan without starting an environment or making a model request:

python -B -m benchkit.campaigns prepare \
  --benchmark simpler_env \
  --model-profile anonymous_code_agent \
  --prompt-file AgentPolicies/anonymous_code_agent/prompts/control_guidance.txt \
  --mailbox-root /absolute/stable/policy-endpoints \
  --output-dir AgentPolicies/anonymous_code_agent/results/simpler_env/raw/my-fast-run/campaign

Replace placeholder paths with writable locations. Use a new output directory. The endpoint root must be independent of task, date, seed and model identity. Campaign preparation freezes cases, budgets, source inventories, effective prompt and future endpoint addresses; it does not prove the native environment is ready.

For a direct profile, choose --model-profile gpt6_astra, omit --mailbox-root and use a new path under AgentPolicies/gpt6_astra/results/<benchmark>/. Omitting --prompt-file uses that profile's prompt. No endpoint or actor mailbox is created. Direct campaigns reject carry-forward of anonymous historical attempts; they require a fresh, explicitly selected schedule. Preparation makes no model request. New GPT6Astra packages declare a local runtime preflight: run-case checks the frozen model's dependencies and credential presence in the same Python environment before creating the permanent case claim. Failure leaves the case unattempted, with no outcome or policy call. This does not validate the key with the provider, check model permissions or prove network/native readiness.

For local diagnostics without a benchmark or request:

python -B -m benchkit preflight-policy --model-profile gpt6_astra

The command reports a bounded status/error code only. Campaign preflight discards child output, times out after 60 seconds, and rechecks the frozen plan/source before its atomic claim. Direct benchkit run also checks before constructing the benchmark. Runtime state can still change afterward; post-claim errors remain permanent attempts. Historical packages without a captured runtime_preflight declaration retain their original behavior and do not acquire this check retroactively.

Before running, inspect campaign.json: selected tasks/states, planned cases, seed, action/observation conditions, control prefix, policy-call limits and policy parameters. Changing a frozen plan requires a new run identity, not an edit. Use the benchmark's separate no-policy diagnostics to verify environment readiness. Such diagnostics are not capability attempts.

Execute an unattempted case

Only after environment readiness and actor setup:

python -B -m benchkit.campaigns run-case \
  --output-dir AgentPolicies/anonymous_code_agent/results/simpler_env/raw/my-fast-run/campaign \
  --case-index 0 --device-index 0

The command claims that case once and runs its frozen entrypoint. A failure or timeout does not authorize another claim. Run additional preselected indices explicitly; preparation alone is not coverage. Concurrency must use separate case identities, native processes and actor contexts.

The runner may execute several controls from an action chunk before the next policy call; the service still returns an observation after each acknowledged control. Record policy calls, executed action decisions, controls and native physics ticks separately. Ordinary policy stop does not award success; the native benchmark remains the scorer.

For retained campaign summaries:

python -B -m benchkit.campaigns summarize \
  --output-dir AgentPolicies/anonymous_code_agent/results/simpler_env/raw/my-fast-run/campaign

Internal integrity tools

For internal integrity and maintenance, keep manifest.json, lifecycle status, episode results/events, source indices and referenced artifacts. Use the generic offline tools under benchkit/audits/ and benchkit/reports/; benchmark-specific auditors may require their matching source/schema version. A consistency check is not a native replay.

Storage and portability

An environment directory on shared storage can still contain a Python symlink or base prefix on node-local storage. Before moving a service, verify the resolved interpreter, pyvenv.cfg, native libraries and asset mount on the destination node. An SSH alias or matching absolute path alone does not establish the same machine or filesystem. Preserve old environments and evidence when rebuilding; installation recovery does not authorize repeating a claimed task.

Maintain reports and indices under the selected model's results/<benchmark>/. Keep raw runs and media outside Git. Export video directories with task IDs and short descriptions, not opaque case numbers alone. Do not rename frozen evidence to improve presentation; use explicit hash-bound indices or copies.

Repository-based editable installation is the documented workflow. When installing a wheel, explicitly choose a writable output directory rather than a default under site-packages. catalog --check validates a complete source checkout; the wheel packages active runtime components, not all benchmark documentation. Internal project_readmes/ notes are not required for public operation.