common/source/starBenchserver/docs/result-format.md

Closed result format for any-language drivers

A trusted evaluation driver can produce these JSON files without implementing Python Policy, using our SDK, or routing model requests through a central proxy. The driver must still use the benchmark service protocol: claims, native execution, scoring and no-replay rules are unchanged. Model actors must not receive the administrative client or private scoring data.

This is the existing starvla-rsi.eval/v1 client-run contract, with optional method and recording declarations. A conforming file is a reported result, not independent proof of native execution or model/provider identity. The basic exporter checks closed-record consistency; full integrity audits are separate.

Files

closed-run/
├── manifest.json
├── status.json
├── summary.json
├── episodes.jsonl
└── episodes/
    └── task-a-seed7/
        └── result.json

Each line of episodes.jsonl is one complete episode result object, in selected schedule order. Each must have an identical object in exactly one episodes/<directory>/result.json; matching uses episode.id, not the directory name. Use safe descriptive directory names. Symlinked source files are rejected.

Interaction events, model/tool transcripts, videos, source snapshots and hashes are not required for basic result export. Disabling them does not reduce publication eligibility. Extra private files are neither scanned nor copied.

Manifest

Required fields and their meanings:

Field	Value
schema_version	starvla-rsi.eval/v1
run_id	Nonempty run identity, also used by the summary
benchmark	name, version, protocol_version, source_revision; copy recorded service declarations, do not infer from today's checkout
policy	At least a declared name; legacy kind, model, provider, version may also be present
action_spec	Service-declared mode, positive size, names, optional low/high, control_hz, frame, rotation, gripper
observation_spec	Service-declared cameras, state_names, modalities, privileged: false, optional control_feedback_schema
protocol	Selected episode_ids, positive max_steps, max_policy_calls (positive integer or null), replan_interval (positive integer or null), fail_fast (boolean); retain actual budget/track declarations

Optional public method declaration:

{
  "method": {
    "schema_version": "starvla-rsi.method/v1",
    "id": "my-planner-reviewer",
    "version": "v1",
    "models": [
      {"id": "planner", "role": "planning", "provider": null, "model": null, "version": null},
      {"id": "reviewer", "role": "action review", "provider": null, "model": null, "version": null}
    ],
    "configuration": {"memory": "rolling", "parallel": false}
  },
  "method_origin": "participant_declaration"
}

These fields belong inside manifest.json. Component IDs are unique ASCII letters/digits/dot/underscore/dash, beginning with a letter or digit. Unknown provider/model/version stays null; unknown composition can be an empty list. Configuration is bounded, explicitly public JSON, not a dump of runtime settings. Do not include credentials, private paths or sensitive prompt content. Credential-named fields and absolute paths are rejected, but validation is not general secret detection: review before publication.

The reference runner projects old PolicyDescriptor identity when no method file is supplied, using method_origin: legacy_policy_projection. Arbitrary legacy parameters are not copied. The report labels that configuration as not supplied, rather than claiming an empty object is a complete method definition.

For multi-call method accounting, the reference runner records protocol.usage_accounting_version: starvla-rsi.method-usage/v1. This separates robot decision-loop calls from model inference requests; it does not change a benchmark's native execution budget.

Episode result

One synthetic example, with no inference usage claimed:

{
  "schema_version": "starvla-rsi.eval/v1",
  "episode": {"id": "task-a/seed7", "task": "task-a", "seed": 7,
              "instruction": "Perform the declared task.", "metadata": {}},
  "status": "completed",
  "success": false,
  "termination_reason": "environment:truncated",
  "policy_calls": 2,
  "action_decisions": 2,
  "environment_steps": 10,
  "rejected_actions": 0,
  "duration_s": 3.5,
  "reward_sum": 0.0,
  "usage": {"calls": null, "input_tokens": null, "output_tokens": null,
            "cached_input_tokens": null, "estimated_cost_usd": null, "latency_s": null},
  "metrics": {"success": false, "score": 40},
  "artifacts": {},
  "error": null
}

completed means an ordinary native outcome or policy stop, not success. Its success must be boolean and match metrics.success. An errored episode uses status: error, success: null, and an error object with a nonempty type plus category equal to policy, protocol, infrastructure or evaluation. The report omits error messages/tracebacks, arbitrary metrics, metadata and paths. Never turn an infrastructure interruption into a completed task failure.

Counts are distinct: policy_calls counts driver decision calls; action_decisions counts decisions with at least one acknowledged control; environment_steps counts acknowledged native controls. usage.calls instead counts reported inference requests and may exceed policy_calls. The example's two decisions and ten controls illustrate a possible method-selected chunk schedule, not a five-controls-per-decision requirement. Native physics ticks are not added to environment_steps, and an unacknowledged in-flight control must not be silently counted as either a confirmed action or zero motion. Unknown usage fields remain null. Failed calls retain known accounting; do not infer cost, tokens or component identity from missing evidence.

Optional component_usage maps declared model component IDs to the same six usage fields. Missing components/fields are unknown, never zero or a share of aggregate usage. The reference runner sums per-field values with unknown propagation. latency_s is reported usage latency, while duration_s is episode wall time; parallel models can have different latency and wall-time totals.

Optional recording uses the reference settings:

{"schema_version":"starvla-rsi.recording/v2","mode":"minimal",
 "interactions":false,"model_io":false,"tool_traces":false,
 "images":false,"videos":false,"copy_server_media":false}

The manifest declares requested settings; each episode can additionally declare agent_events_written and bounded warning codes. A setting being enabled does not prove every provider exposed a transcript. See recording options.

Closure and summary

status.json contains status (completed or completed_with_errors) and episodes_written. All selected episodes must be present unless the declared fail_fast setting stopped after an errored episode. Unattempted cases stay in the manifest schedule and the report denominator; they are never silently deleted or retried.

summary.json repeats run_id, benchmark (name), policy (name), and totals for attempted, completed, errors, successes, policy_calls, rejected_actions, and the six-field aggregate usage object. Optional component_usage must agree with episode totals. Unknown propagates when summing usage. When both are known, cached input tokens cannot exceed input tokens. Optional success_rate, attempt_success_rate, selected_success_rate and completion_rate must agree with the recorded counts; boolean rates are invalid.

The basic report prominently shows successes / selected cases (selected_success_rate), retaining errors and unattempted selected cases in that denominator. It also provides successes/completed (success_rate), successes/attempted (attempt_success_rate) and explicit denominator labels. For example, one success followed by one infrastructure error and one fail-fast-unattempted case is 1/3 over selected, 1/2 over attempted and 1/1 over completed—not an unqualified 100% result. None establishes that a client-selected schedule covers the entire native benchmark.

Export

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

The destination must be new and outside the source run. The exporter writes result.json and standalone index.html, without changing source files or calling a model/simulator. Repeat --metric KEY to explicitly include additional native numeric/boolean metrics; missing metrics are null. It does not select arbitrary nested metric objects or copy media automatically. This publication choice does not make private metrics available to the acting model.

Inputs reject duplicate JSON keys and nonfinite numbers, are bounded to 10,000 episodes and 128 MiB of files read, and are checked for changes during export. The optional interaction viewer adds a timeline when compatible detailed events exist. Its absence does not prevent the basic result above.