1. Connect your data
In Data, create a source and copy its capture URL and capture key. Keep your existing provider API key. Send the capture key inx-reasonblocks-key, alongside
your normal provider authentication. Use the exact URL returned for your source:
- OpenAI Chat Completions:
/capture/SOURCE_ID/openai/v1as the SDK base URL. - Anthropic Messages:
/capture/SOURCE_ID/anthropicas the SDK base URL.
x-rb-run for each task and reuse it on every model call in that
task. Use 1–128 letters, numbers, underscores, periods, colons or hyphens. Keep
the full conversation, policy, tool definitions, and actual tool results in each
request. Append the returned assistant message and its tool result before the
next call; your application continues executing its tools.
The snapshot ID refers to a repeatable starting state in your test environment.
Freeze it before the workflow changes that state and keep it unchanged for the
whole execution. Capture does not create the snapshot for you.
You can instead upload JSONL containing id, group_id, OpenAI-shaped text
messages, function tools, and metadata.environment_snapshot_id. Use the
same group_id for related attempts so they cannot cross the training,
calibration and evaluation split. The initial history, tool contract, policy
version and snapshot must agree within a group. A missing snapshot ID defaults
to the group ID, which must still exist in your environment.
Provide one complete starting conversation per task, with no pending tool calls.
Do not include hidden expected outcomes in the conversation, tool definitions or
snapshot names. Multiple rows from a long conversation are not independent tasks.
The default recipe needs at least 32 independent training tasks after separate
calibration and evaluation groups have been reserved.
The current source capture limit is 200 requests. For longer workflows or a
larger collection, upload curated initial-task JSONL instead; the upload limit is
2 MiB.
2. Connect a test environment once
Ask your administrator to register an isolated copy of your agent’s tools and state. It appears in Training for your organization. For each attempt, it must restore the task’s snapshot into a fresh disposable session, execute newly proposed tool actions, return their real results, and preserve state until the task ends. Training writes must stay inside this environment. An independent outcome evaluator is also required. It may be your deterministic checker or an identified judge such as Fable. Judge assessments are reported as judge assessments. The evaluator inspects completed tasks for calibration and evaluation; its answers are not inputs to distillation or the serving guard. Your administrator can follow the platform’s sandbox connection guide, including an executable local example, and deployment guide. Repository access may be required. Provider credentials, server URLs and registry files belong in that administrator setup, not in your uploaded tasks.3. Prepare and approve training
In Training, select your data source and Full agent distillation. A single connected environment is selected automatically. Choose Prepare training to freeze the data and check the snapshots, tools, evaluator and local model setup. Preparation makes no model calls. If it fails, fix the reported issue and choose Recheck setup. Remote provider access and funding are verified during the authorized provider stages, not by these local checks. The current preset usesQwen/Qwen3.6-35B-A3B as the student and
Qwen/Qwen3.5-397B-A17B as both the teacher and serving fallback. The student
attempts tasks and calls tools; the open teacher scores the student’s own sampled
tokens. Those token scores drive token-level on-policy distillation. This
recipe does not require golden-trace filtering or switch to teacher-text
supervised fine-tuning when token scores are unavailable.
The preset runs 75 updates, with 32 task groups and two fresh trajectories per
group, rank 16, learning rate 0.00003 and training temperature 1. Review the plan
and cost estimate, set a run spending cap, and approve paid work. The cap and
conservative reservation are limits, not predicted invoices. Baselines, training
and evaluation share the run budget. A later budget pause can be resumed after
adding credits, once any uncertain provider reservations are resolved.
API setup
For API integrations, create a run with an authenticatedPOST /api/pipelines.
Use your organization and populated data source IDs and the connected environment
ID. Set config.mode explicitly: omitting it selects the legacy request-response
method.
POST /api/pipelines/PIPELINE_ID/run queues free
preparation; review its checks and estimate in Training before approval.
API approval uses /approve-eval with reviewed criteria and notes; it
authorizes the paid baseline, training and evaluation stages together.
4. Review complete-task results
The report compares the stock student, teacher, distilled student, and distilled student with its guard on separate evaluation tasks. It includes task success, mean and p95 latency, and inference cost, with upfront training and evaluation spending reported separately. It identifies the evaluator and retains incomplete or unknown outcomes instead of treating them as successful tasks. Inference includes student calls, discarded student proposals, teacher fallback and embeddings. Token-based costs are tariff estimates until invoices are reconciled. Paid judge evaluation is separate from serving cost. Sandbox hosting and business-tool charges are external costs unless separately metered. Missing measurements are unavailable, not zero.5. Enable the reviewed release
Select a traffic share and confirm inference spending in Training. Keep the same capture URL, full history and stablex-rb-run. The rollout share selects
which tasks enter the candidate system; the guard’s teacher escalation rate is
a separate measurement.
The guard checks customer coverage, tool validity, repeated calls, observed
errors and calibrated limits. It can escalate unfamiliar tasks to the frozen
teacher. The teacher continues the actual conversation and tool state, without
replaying earlier actions. Your application still executes the tool calls.
Out-of-distribution detection is present, but these task evaluations do not
establish its detection accuracy; that requires a separate labeled test.
Use non-streaming OpenAI Chat Completions or Anthropic Messages with text and
ordinary function tools. Omit temperature and output-token overrides to use the
reviewed release defaults: 0.7 and 1,536 tokens. The Anthropic SDK requires
max_tokens; set it to 1536. Explicit smaller output caps are accepted.
Unsupported initial requests, including images, forced tool selection, strict
tool schemas, response-format controls or an incompatible temperature, remain
with your original provider. A missing run ID also keeps a new request upstream.
Once a guarded episode starts, changing its history or request contract causes
an error rather than silently moving it to another provider.
Ask your administrator to keep each run on the same API worker for its entire
task. An interrupted or different worker cannot transparently resume its live
runtime. Reconcile uncertain tool state before starting a new execution; do not
blindly retry a write under a new run ID. Use the application’s own tool timeout
and idempotency handling.
Return to original provider stops admission of new candidate tasks. It cannot
undo an action already executed, and it does not restart an active task. Keep the
reviewed checkpoint available and maintain sufficient inference credits while
the release is enabled.