ambral-evals 0.1.0__tar.gz

ambral_evals-0.1.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

ambral_evals-0.1.0/PKG-INFO

@@ -0,0 +1,218 @@
+Metadata-Version: 2.4
+Name: ambral-evals
+Version: 0.1.0
+Summary: Reports a suite's scenarios to Ambral: opens an offline run, hands each scenario's agent its headers, reports what the suite scored, and closes the run.
+License-Expression: Apache-2.0
+License-File: LICENSE
+Classifier: Programming Language :: Python :: 3
+Classifier: Typing :: Typed
+Requires-Dist: httpx>=0.27,<1
+Requires-Python: >=3.10
+Project-URL: Repository, https://github.com/AmbralAI/Cerberus
+Description-Content-Type: text/markdown
+
+# ambral-evals
+
+Reports an eval suite's scenarios to Ambral. Your harness runs each scenario and scores it; its agent's model calls go
+through Ambral's Gateway, which serves them with the model the run chose and records them. The SDK opens a run, hands
+each scenario's agent its headers, reports what the suite scored, and closes the run, which then reads with
+`ambral runs show <run>`.
+
+A run holds two things for each scenario it ran: its **result** (what the suite reported: whether it passed, its
+rewards, the reasons for them) and its **calls** (the agent's model calls the Gateway recorded). A result can have no
+calls (its agent called no model through the Gateway), and a session can have made calls but have no result (its report
+was refused, or never sent): `ambral runs` reads both.
+
+```sh
+pip install ambral-evals
+```
+
+Python 3.10 or later. Its one dependency is [httpx](https://www.python-httpx.org/).
+
+## A verifiers eval
+
+Run the eval with `ambral-verifiers`, naming the suite, then `vf-eval`'s own arguments:
+
+```sh
+ambral-verifiers support-agent/refunds @ eval.toml
+```
+
+It opens the run, sets `AMBRAL_EVAL_HEADER` to its header, runs `vf-eval` in the same process, reports each episode as
+it finishes, and closes the run. The eval's agent seat calls the Gateway, naming the header in its client:
+
+```toml
+model = "ambral/auto"
+
+[env.agent.client]
+type = "eval"
+base_url = "https://app.ambral.ai/v1"
+api_key_var = "AMBRAL_API_KEY"
+headers_from_env = { x-ambral-eval = "AMBRAL_EVAL_HEADER" }
+```
+
+- The trace of the seat whose client names the header is one run of its episode's scenario. A trace of another seat (a
+  simulated user, a judge) isn't reported. Give only the agent's seat the header, in its own client
+  (`[env.agent.client]`): an episode in which more than one seat sends it (as each does when the eval's own client
+  names it) isn't reported, and the run's close names it.
+- verifiers sends each rollout's trace id as `X-Session-ID`, which is the session its calls are recorded under.
+- A scenario's id is its task's `id`, or else its `name`, or else its index in the taskset, which holds only while the
+  taskset's order does: give each task an `id`.
+- Its rewards, metrics, judge calls and errors are the trace's, and its task the episode's.
+- `ambral-verifiers` reports what this invocation runs: a resumed eval (`--resume`) reports the rollouts it runs again,
+  into a run of its own unless `AMBRAL_EVAL_KEY` names the first one's key. With no arguments after the suite, `--help`
+  or `--dry-run`, `vf-eval` runs no rollout, and no run is opened.
+- An episode that can't be reported is printed and named when the run closes; the eval goes on. It exits with
+  `vf-eval`'s code, or 1 when a scenario has no result, or when `vf-eval` wrote episodes `ambral-verifiers` wasn't
+  handed (a verifiers other than the one it was built against).
+
+## A Harbor job
+
+Run the job with the plugin, naming the suite:
+
+```sh
+harbor run --plugin ambral_evals.harbor:Report \
+  --plugin-kwarg workload=support-agent --plugin-kwarg taskset=terminal-tasks ...
+```
+
+Your agent class sends the headers `headers(self.context_id)` gives on each of its model calls:
+
+```python
+from ambral_evals.harbor import headers
+
+
+class SupportAgent(BaseAgent):
+    async def run(self, instruction, environment, context):
+        client = AsyncOpenAI(base_url=gateway_url, default_headers=headers(self.context_id))
+        ...
+```
+
+- Each trial is one run of its task, whose name is the scenario's id. Its session is the trial's id (the agent's
+  `context_id`), which is new for each attempt.
+- Its rewards are the verifier's (`reward.json`), and each reward's reason is what its criteria said of it, from the
+  trial's `verifier/reward-details.json` when the verifier wrote one (rewardkit does). An exception the trial recorded
+  is one of its errors.
+- The run's key is the job's id, which a resumed job keeps: it reopens its run and reports the trials it runs now.
+  `--plugin-kwarg key=…` or `AMBRAL_EVAL_KEY` names another; `model=…` and `source=…` name the run's model and source.
+- The run opens before any trial runs, and a run that can't be opened stops the job. A trial that can't be reported
+  is printed and named when the run closes; the job goes on.
+
+## Any other harness
+
+```python
+from ambral_evals import open_run
+
+with open_run("support-agent", "refunds", model="anthropic/claude-haiku-5-5") as run:
+    scenario = run.scenario("refunds/over-the-limit")
+    client = OpenAI(base_url=os.environ["AMBRAL_GATEWAY_URL"], default_headers=dict(scenario.headers))
+    # ... the agent runs ...
+    scenario.report({"passed": False, "rewards": {"solved": {"score": 0}}, "reasons": {"solved": "..."}})
+```
+
+- `run.scenario(id, session=...)` takes the agent's own session id, for an agent that names its sessions.
+- `scenario.report(result)` returns once Ambral stored it; reports are sent together, about once a second.
+  `scenario.send(result)` queues it and returns a `concurrent.futures.Future` at once, for a harness that shouldn't
+  wait (`await asyncio.wrap_future(scenario.send(result))` in asyncio). A report queued is sent: cancelling the wait
+  (a timeout) doesn't take it back.
+- `scenario.drop()` says a run of a scenario has no result to report (it was skipped), so the close doesn't ask for
+  one.
+- `run.close()` (or the end of the `with` block) sends what is left, prints the run's id and how to read it, and raises
+  `ReportError` naming each scenario this process ran that has no result. A result that doesn't fit (a reward
+  outside 0 to 1, a `NaN` in its task) raises at `send` or `report`, before it is queued, so it never sinks another
+  scenario's report.
+
+## What a result holds, and how it reads back
+
+Each result is one run of one scenario, a `dict`. Every key is optional, but a result with neither `passed` nor a reward
+says little.
+
+| Key           | Shape                                  | What it holds                                                                                                                                                                                             |
+| ------------- | -------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `passed`      | `bool`                                 | The suite's own verdict. Left out, the run is judged by its scores alone.                                                                                                                                 |
+| `rewards`     | `{name: {"score", "weight"?} \| None}` | `score` from 0 (fails) to 1 (passes); a score outside that is refused. `weight` from 0 to 1000, 1 when left out. `None` for a reward that didn't run. A suite that scores on another scale divides by it. |
+| `metrics`     | `{name: float \| None}`                | What else the suite measures (turns, latency, its own cost): no bound.                                                                                                                                    |
+| `reasons`     | `{name: str}`                          | Why a reward or metric got its value, by its name; at most 16,384 UTF-16 code units each.                                                                                                                 |
+| `task`        | verifiers' `TaskData`                  | The scenario's prompt, what it expects, and any keys of the suite's; at most 256 KiB as JSON. Sent once per scenario, then by its id.                                                                     |
+| `judge_calls` | verifiers' judge calls                 | Each judge exchange behind the scores, at most 64.                                                                                                                                                        |
+| `errors`      | `[{"type", "message"}]`                | What failed while the scenario ran.                                                                                                                                                                       |
+
+Neither integration rescales a score: a verifiers or Harbor reward outside 0 to 1 refuses its scenario before it is
+sent, naming the reward. Divide it by its scale in the suite, or record it as a metric.
+
+How Ambral reads them back:
+
+- **A run of a scenario's score** is the weighted mean of its rewards: `solved` 0.25 (weight 1) and `polite` 1
+  (weight 0.5) score 0.50.
+- **`ambral runs show <run>`** gives the mean of those scores over the run, the pass tally from `passed`, and for each
+  reward its mean, median, how many runs scored it, how many runs and scenarios scored it below 1, and the reason
+  given for its lowest score.
+- **`ambral runs verdicts <run>`** lists each reward of each run of a scenario with its reason; `--failed` keeps the
+  runs `passed` says failed, `--below <n>` and `--check <name>` filter by reward.
+- **`ambral runs trace <run> <scenario>`** shows one run whole: its task, each reward with its reason, its metrics and
+  errors, and its calls.
+- **`ambral runs export <run>`** writes each run of a scenario as verifiers' `Episode`, which verifiers and prime-rl
+  load.
+
+Metrics are shown per run, in `runs trace` and `runs export`; `runs show` doesn't add them up.
+
+## Handing the agent its headers
+
+Every model call of a scenario's agent carries its headers, which the Gateway reads the run and the scenario from. Any
+of three ways:
+
+- as its model client's default headers (`default_headers=dict(scenario.headers)`);
+- on each request (`extra_headers=dict(scenario.headers)`);
+- through `ambral_evals.httpx`'s transport, for a client built once, at module scope, which adds the current
+  scenario's:
+
+  ```python
+  import httpx
+  from ambral_evals.httpx import Transport  # AsyncTransport for an httpx.AsyncClient
+
+  client = OpenAI(base_url=os.environ["AMBRAL_GATEWAY_URL"], http_client=httpx.Client(transport=Transport()))
+
+  with scenario.current():
+      agent(client, "Refund invoice inv_123")
+  ```
+
+  The current scenario is the one whose `current()` block the call is made within, in the same thread or asyncio task,
+  or one started within it. A call through it outside a scenario is refused, rather than recorded as anything but the
+  run's.
+
+Only the system under test's clients carry them: never a judge's, or a simulated user's.
+
+The agent sends model `ambral/auto`, as it does in production. The headers name the run's workload, so a call that
+names none is served the run's model. One that names another workload (`x-workload`, or `ambral/auto:<workload>`), such
+as a second agent the first calls, is served that workload's primary, and is recorded under the run too.
+
+## The environment
+
+| Variable             | What it says                                                                                         |
+| -------------------- | ---------------------------------------------------------------------------------------------------- |
+| `AMBRAL_API_KEY`     | the org's API key, the one its agent calls the Gateway with. Required                                |
+| `AMBRAL_URL`         | where Ambral is served: `https://app.ambral.ai` when unset                                           |
+| `AMBRAL_EVAL_MODEL`  | the model the run serves, any of the workload's pool, when the code names none: its primary if unset |
+| `AMBRAL_EVAL_KEY`    | the key the suite's processes share one run by (below)                                               |
+| `AMBRAL_EVAL_SOURCE` | `ci` or `local`, where the code names none                                                           |
+
+A CI matrix over models sets `AMBRAL_EVAL_MODEL` per job, with no change to the code.
+
+## One run per invocation, on CI
+
+Shards, workers and matrix jobs of one CI run report to one run: each opens it by the same key, taskset and model, and
+it completes once each has closed it. The key is the CI's own id for the workflow run:
+
+- GitHub Actions: the run's id and attempt, so re-running failed jobs is a new run, holding only the jobs run again;
+- GitLab CI: the pipeline's id; CircleCI: the workflow's; Buildkite: the build's. A retried job there keeps that id,
+  so it reports into the same run, and its scenarios gain repeats. Set `AMBRAL_EVAL_KEY` to tell its attempts apart;
+- Jenkins and Azure Pipelines don't set `CI`: set `AMBRAL_EVAL_SOURCE=ci`, and `AMBRAL_EVAL_KEY` to the pipeline run's
+  id.
+
+Elsewhere, each invocation opens a run of its own. A fixed `AMBRAL_EVAL_KEY` opens the same run every time, again once
+it completed, so make it unique per invocation. A Harbor job's run is keyed by the job's id instead (above).
+
+A run records where it ran: `ci` when `CI` is set (to anything but `0` or `false`), and `local` otherwise. A suite's
+history leaves local runs out unless asked for them.
+
+## License
+
+Apache-2.0.

ambral_evals-0.1.0/README.md

@@ -0,0 +1,205 @@
+# ambral-evals
+
+Reports an eval suite's scenarios to Ambral. Your harness runs each scenario and scores it; its agent's model calls go
+through Ambral's Gateway, which serves them with the model the run chose and records them. The SDK opens a run, hands
+each scenario's agent its headers, reports what the suite scored, and closes the run, which then reads with
+`ambral runs show <run>`.
+
+A run holds two things for each scenario it ran: its **result** (what the suite reported: whether it passed, its
+rewards, the reasons for them) and its **calls** (the agent's model calls the Gateway recorded). A result can have no
+calls (its agent called no model through the Gateway), and a session can have made calls but have no result (its report
+was refused, or never sent): `ambral runs` reads both.
+
+```sh
+pip install ambral-evals
+```
+
+Python 3.10 or later. Its one dependency is [httpx](https://www.python-httpx.org/).
+
+## A verifiers eval
+
+Run the eval with `ambral-verifiers`, naming the suite, then `vf-eval`'s own arguments:
+
+```sh
+ambral-verifiers support-agent/refunds @ eval.toml
+```
+
+It opens the run, sets `AMBRAL_EVAL_HEADER` to its header, runs `vf-eval` in the same process, reports each episode as
+it finishes, and closes the run. The eval's agent seat calls the Gateway, naming the header in its client:
+
+```toml
+model = "ambral/auto"
+
+[env.agent.client]
+type = "eval"
+base_url = "https://app.ambral.ai/v1"
+api_key_var = "AMBRAL_API_KEY"
+headers_from_env = { x-ambral-eval = "AMBRAL_EVAL_HEADER" }
+```
+
+- The trace of the seat whose client names the header is one run of its episode's scenario. A trace of another seat (a
+  simulated user, a judge) isn't reported. Give only the agent's seat the header, in its own client
+  (`[env.agent.client]`): an episode in which more than one seat sends it (as each does when the eval's own client
+  names it) isn't reported, and the run's close names it.
+- verifiers sends each rollout's trace id as `X-Session-ID`, which is the session its calls are recorded under.
+- A scenario's id is its task's `id`, or else its `name`, or else its index in the taskset, which holds only while the
+  taskset's order does: give each task an `id`.
+- Its rewards, metrics, judge calls and errors are the trace's, and its task the episode's.
+- `ambral-verifiers` reports what this invocation runs: a resumed eval (`--resume`) reports the rollouts it runs again,
+  into a run of its own unless `AMBRAL_EVAL_KEY` names the first one's key. With no arguments after the suite, `--help`
+  or `--dry-run`, `vf-eval` runs no rollout, and no run is opened.
+- An episode that can't be reported is printed and named when the run closes; the eval goes on. It exits with
+  `vf-eval`'s code, or 1 when a scenario has no result, or when `vf-eval` wrote episodes `ambral-verifiers` wasn't
+  handed (a verifiers other than the one it was built against).
+
+## A Harbor job
+
+Run the job with the plugin, naming the suite:
+
+```sh
+harbor run --plugin ambral_evals.harbor:Report \
+  --plugin-kwarg workload=support-agent --plugin-kwarg taskset=terminal-tasks ...
+```
+
+Your agent class sends the headers `headers(self.context_id)` gives on each of its model calls:
+
+```python
+from ambral_evals.harbor import headers
+
+
+class SupportAgent(BaseAgent):
+    async def run(self, instruction, environment, context):
+        client = AsyncOpenAI(base_url=gateway_url, default_headers=headers(self.context_id))
+        ...
+```
+
+- Each trial is one run of its task, whose name is the scenario's id. Its session is the trial's id (the agent's
+  `context_id`), which is new for each attempt.
+- Its rewards are the verifier's (`reward.json`), and each reward's reason is what its criteria said of it, from the
+  trial's `verifier/reward-details.json` when the verifier wrote one (rewardkit does). An exception the trial recorded
+  is one of its errors.
+- The run's key is the job's id, which a resumed job keeps: it reopens its run and reports the trials it runs now.
+  `--plugin-kwarg key=…` or `AMBRAL_EVAL_KEY` names another; `model=…` and `source=…` name the run's model and source.
+- The run opens before any trial runs, and a run that can't be opened stops the job. A trial that can't be reported
+  is printed and named when the run closes; the job goes on.
+
+## Any other harness
+
+```python
+from ambral_evals import open_run
+
+with open_run("support-agent", "refunds", model="anthropic/claude-haiku-5-5") as run:
+    scenario = run.scenario("refunds/over-the-limit")
+    client = OpenAI(base_url=os.environ["AMBRAL_GATEWAY_URL"], default_headers=dict(scenario.headers))
+    # ... the agent runs ...
+    scenario.report({"passed": False, "rewards": {"solved": {"score": 0}}, "reasons": {"solved": "..."}})
+```
+
+- `run.scenario(id, session=...)` takes the agent's own session id, for an agent that names its sessions.
+- `scenario.report(result)` returns once Ambral stored it; reports are sent together, about once a second.
+  `scenario.send(result)` queues it and returns a `concurrent.futures.Future` at once, for a harness that shouldn't
+  wait (`await asyncio.wrap_future(scenario.send(result))` in asyncio). A report queued is sent: cancelling the wait
+  (a timeout) doesn't take it back.
+- `scenario.drop()` says a run of a scenario has no result to report (it was skipped), so the close doesn't ask for
+  one.
+- `run.close()` (or the end of the `with` block) sends what is left, prints the run's id and how to read it, and raises
+  `ReportError` naming each scenario this process ran that has no result. A result that doesn't fit (a reward
+  outside 0 to 1, a `NaN` in its task) raises at `send` or `report`, before it is queued, so it never sinks another
+  scenario's report.
+
+## What a result holds, and how it reads back
+
+Each result is one run of one scenario, a `dict`. Every key is optional, but a result with neither `passed` nor a reward
+says little.
+
+| Key           | Shape                                  | What it holds                                                                                                                                                                                             |
+| ------------- | -------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `passed`      | `bool`                                 | The suite's own verdict. Left out, the run is judged by its scores alone.                                                                                                                                 |
+| `rewards`     | `{name: {"score", "weight"?} \| None}` | `score` from 0 (fails) to 1 (passes); a score outside that is refused. `weight` from 0 to 1000, 1 when left out. `None` for a reward that didn't run. A suite that scores on another scale divides by it. |
+| `metrics`     | `{name: float \| None}`                | What else the suite measures (turns, latency, its own cost): no bound.                                                                                                                                    |
+| `reasons`     | `{name: str}`                          | Why a reward or metric got its value, by its name; at most 16,384 UTF-16 code units each.                                                                                                                 |
+| `task`        | verifiers' `TaskData`                  | The scenario's prompt, what it expects, and any keys of the suite's; at most 256 KiB as JSON. Sent once per scenario, then by its id.                                                                     |
+| `judge_calls` | verifiers' judge calls                 | Each judge exchange behind the scores, at most 64.                                                                                                                                                        |
+| `errors`      | `[{"type", "message"}]`                | What failed while the scenario ran.                                                                                                                                                                       |
+
+Neither integration rescales a score: a verifiers or Harbor reward outside 0 to 1 refuses its scenario before it is
+sent, naming the reward. Divide it by its scale in the suite, or record it as a metric.
+
+How Ambral reads them back:
+
+- **A run of a scenario's score** is the weighted mean of its rewards: `solved` 0.25 (weight 1) and `polite` 1
+  (weight 0.5) score 0.50.
+- **`ambral runs show <run>`** gives the mean of those scores over the run, the pass tally from `passed`, and for each
+  reward its mean, median, how many runs scored it, how many runs and scenarios scored it below 1, and the reason
+  given for its lowest score.
+- **`ambral runs verdicts <run>`** lists each reward of each run of a scenario with its reason; `--failed` keeps the
+  runs `passed` says failed, `--below <n>` and `--check <name>` filter by reward.
+- **`ambral runs trace <run> <scenario>`** shows one run whole: its task, each reward with its reason, its metrics and
+  errors, and its calls.
+- **`ambral runs export <run>`** writes each run of a scenario as verifiers' `Episode`, which verifiers and prime-rl
+  load.
+
+Metrics are shown per run, in `runs trace` and `runs export`; `runs show` doesn't add them up.
+
+## Handing the agent its headers
+
+Every model call of a scenario's agent carries its headers, which the Gateway reads the run and the scenario from. Any
+of three ways:
+
+- as its model client's default headers (`default_headers=dict(scenario.headers)`);
+- on each request (`extra_headers=dict(scenario.headers)`);
+- through `ambral_evals.httpx`'s transport, for a client built once, at module scope, which adds the current
+  scenario's:
+
+  ```python
+  import httpx
+  from ambral_evals.httpx import Transport  # AsyncTransport for an httpx.AsyncClient
+
+  client = OpenAI(base_url=os.environ["AMBRAL_GATEWAY_URL"], http_client=httpx.Client(transport=Transport()))
+
+  with scenario.current():
+      agent(client, "Refund invoice inv_123")
+  ```
+
+  The current scenario is the one whose `current()` block the call is made within, in the same thread or asyncio task,
+  or one started within it. A call through it outside a scenario is refused, rather than recorded as anything but the
+  run's.
+
+Only the system under test's clients carry them: never a judge's, or a simulated user's.
+
+The agent sends model `ambral/auto`, as it does in production. The headers name the run's workload, so a call that
+names none is served the run's model. One that names another workload (`x-workload`, or `ambral/auto:<workload>`), such
+as a second agent the first calls, is served that workload's primary, and is recorded under the run too.
+
+## The environment
+
+| Variable             | What it says                                                                                         |
+| -------------------- | ---------------------------------------------------------------------------------------------------- |
+| `AMBRAL_API_KEY`     | the org's API key, the one its agent calls the Gateway with. Required                                |
+| `AMBRAL_URL`         | where Ambral is served: `https://app.ambral.ai` when unset                                           |
+| `AMBRAL_EVAL_MODEL`  | the model the run serves, any of the workload's pool, when the code names none: its primary if unset |
+| `AMBRAL_EVAL_KEY`    | the key the suite's processes share one run by (below)                                               |
+| `AMBRAL_EVAL_SOURCE` | `ci` or `local`, where the code names none                                                           |
+
+A CI matrix over models sets `AMBRAL_EVAL_MODEL` per job, with no change to the code.
+
+## One run per invocation, on CI
+
+Shards, workers and matrix jobs of one CI run report to one run: each opens it by the same key, taskset and model, and
+it completes once each has closed it. The key is the CI's own id for the workflow run:
+
+- GitHub Actions: the run's id and attempt, so re-running failed jobs is a new run, holding only the jobs run again;
+- GitLab CI: the pipeline's id; CircleCI: the workflow's; Buildkite: the build's. A retried job there keeps that id,
+  so it reports into the same run, and its scenarios gain repeats. Set `AMBRAL_EVAL_KEY` to tell its attempts apart;
+- Jenkins and Azure Pipelines don't set `CI`: set `AMBRAL_EVAL_SOURCE=ci`, and `AMBRAL_EVAL_KEY` to the pipeline run's
+  id.
+
+Elsewhere, each invocation opens a run of its own. A fixed `AMBRAL_EVAL_KEY` opens the same run every time, again once
+it completed, so make it unique per invocation. A Harbor job's run is keyed by the job's id instead (above).
+
+A run records where it ran: `ci` when `CI` is set (to anything but `0` or `false`), and `local` otherwise. A suite's
+history leaves local runs out unless asked for them.
+
+## License
+
+Apache-2.0.