refractal 0.1.0a1__tar.gz

refractal-0.1.0a1/.gitignore

@@ -0,0 +1,20 @@
+# Anchored to the repo root. An unanchored `build/` also matches
+# src/refractal/build/, which silently kept the whole `refractal build` package
+# out of git -- `git status` stayed clean, the tests passed against the working
+# tree, and the wheel shipped without it. Anchor these.
+__pycache__/
+*.py[cod]
+*.egg-info/
+.venv/
+/build/
+/dist/
+/results/
+/plan.json
+
+# mkdocs build output. Regenerable from docs/ + mkdocs.yml, and when it was
+# tracked it went stale: a wording fix in docs/ left the built HTML asserting
+# the old text.
+site/
+.mutation-in-progress
+.coverage.json
+.coverage

refractal-0.1.0a1/CHANGELOG.md

@@ -0,0 +1,66 @@
+# Changelog
+
+What changed in each released version. Newest first.
+
+This is not the commit history. The history records why a decision was made;
+this records what a user gets in a version they can install. Different
+audiences, and the history is the better read for the first question.
+
+## 0.1.0a1 (unreleased)
+
+First alpha. Every stage is implemented and tested against real model servers
+as well as the synthetic backend.
+
+**`plan_id` may move between alpha versions.** Episode identity is a hash of
+the catalog's content, and the fields that feed it are still settling: they
+changed three times this month. Results written by `0.1.0a1` may therefore fail
+to join with results from a later alpha, and `refractal compare` will refuse
+the join rather than pool them, which is the intended behaviour and not a bug.
+If you are about to spend GPU hours, this is the sentence to read first. The
+`a` in the version is doing real work.
+
+`plan_schema` is 3. It is deliberately not part of `plan_id`: a format bump
+does not make an experiment a different experiment.
+
+### Added
+
+- `refractal init`, write a working example catalog that runs end to end with
+  no GPU, no simulator and no checkpoints.
+- `refractal build`, compute scene hashes, evaluate filters and record resource
+  shapes into `catalog/build.lock`.
+- `refractal plan`, compile a catalog into `plan.json`. Runs with none of the
+  optional dependencies installed, which is enforced by a test rather than by
+  convention.
+- `refractal run`, execute a plan. Backends: `local` (simulates outcomes, no
+  infrastructure), `vla-eval` (drives the harness against running model
+  servers), `compose` (one container per worker).
+- `refractal render`, write a deployment description from a plan without Docker
+  or a cluster installed. Targets: `compose` and `k8s`.
+- `refractal compare`, compare checkpoints in a results directory. Paired
+  McNemar on scenario-level outcomes, a clustered bootstrap that resamples
+  whole scenarios, Cochran's Q as a screen across three or more checkpoints,
+  and Holm correction across the whole contrast family.
+- `refractal explain`, why two plans are different experiments.
+- Perturbations: timed changes to the world, the observation or the action
+  during an episode, hashed into `scenario_hash` so a perturbed episode is a
+  different experiment from its unperturbed counterpart.
+- Exit codes for CI: 0 no regression, 1 regression, 2 cannot be answered.
+  The third is never reported as the second.
+- Optional extras: `execute`, `compare`, `video`, `vla-eval`, `all`, `test`.
+  Comparison needs no numerical stack.
+
+### Not built yet
+
+Stated because a gap a user discovers is worse than one they were told about.
+
+- **Trajectory quality metrics.** The `steps.parquet` schema is declared and
+  nothing writes it. Writing it against synthetic data would bake in guesses
+  about what a real adapter can record.
+- **`refractal run --backend k8s`.** Kubernetes is reached through `render`:
+  Refractal writes one Job per worker and your cluster schedules them. Nothing
+  submits them, watches them or collects their exit codes.
+- **Resource-shape probing.** `refractal build` carries hand-declared shapes
+  forward and records whether they were measured. A lint refuses a comment
+  claiming a measurement when nothing recorded one.
+- **Partial credit.** `phase_outcomes` exists in the schema and the LIBERO
+  adapter never populates it, so success is binary.

refractal-0.1.0a1/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 2026 Nalin Raut
+
+   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.

refractal-0.1.0a1/NOTICE

@@ -0,0 +1,23 @@
+Refractal
+Copyright 2026 Nalin Raut
+
+Licensed under the Apache License, Version 2.0. See LICENSE.
+
+This product depends on software from the following projects.
+No source from them is included in this distribution.
+
+-------------------------------------------------------------------------------
+vla-evaluation-harness
+https://github.com/allenai/vla-evaluation-harness
+Copyright The Allen Institute for Artificial Intelligence
+Licensed under the Apache License, Version 2.0
+
+Refractal's `--backend vla-eval` runs every episode through this harness. It
+provides the model-server protocol, the benchmark adapters, the observation and
+action specifications, and the episode loop. Refractal subclasses two of its
+classes (`Orchestrator` and `EpisodeRecorder`) and adds no copies of its source.
+-------------------------------------------------------------------------------
+
+Benchmark suites and model servers loaded through the harness at run time are
+the work of their own authors and carry their own licences. Refractal neither
+distributes nor modifies them.

refractal-0.1.0a1/PKG-INFO

@@ -0,0 +1,314 @@
+Metadata-Version: 2.5
+Name: refractal
+Version: 0.1.0a1
+Summary: Scene-coherent placement and paired comparison for policy evaluation
+Project-URL: Homepage, https://github.com/nalinraut/refractal
+Project-URL: Repository, https://github.com/nalinraut/refractal
+Project-URL: Issues, https://github.com/nalinraut/refractal/issues
+Author: Nalin Raut
+License-Expression: Apache-2.0
+License-File: LICENSE
+License-File: NOTICE
+Keywords: benchmark,evaluation,reproducibility,robotics,simulation,statistics,vla
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Software Development :: Testing
+Classifier: Typing :: Typed
+Requires-Python: >=3.11
+Requires-Dist: pydantic>=2.6
+Requires-Dist: pyyaml>=6.0
+Provides-Extra: all
+Requires-Dist: fsspec>=2024.2.0; extra == 'all'
+Requires-Dist: pillow>=10.0; extra == 'all'
+Requires-Dist: pyarrow>=15.0; extra == 'all'
+Provides-Extra: compare
+Requires-Dist: fsspec>=2024.2.0; extra == 'compare'
+Requires-Dist: pyarrow>=15.0; extra == 'compare'
+Provides-Extra: execute
+Requires-Dist: fsspec>=2024.2.0; extra == 'execute'
+Requires-Dist: pyarrow>=15.0; extra == 'execute'
+Provides-Extra: test
+Requires-Dist: numpy>=1.24; extra == 'test'
+Requires-Dist: pillow>=10.0; extra == 'test'
+Provides-Extra: video
+Requires-Dist: pillow>=10.0; extra == 'video'
+Provides-Extra: vla-eval
+Requires-Dist: fsspec>=2024.2.0; extra == 'vla-eval'
+Requires-Dist: pyarrow>=15.0; extra == 'vla-eval'
+Requires-Dist: vla-eval>=0.6.0; extra == 'vla-eval'
+Description-Content-Type: text/markdown
+
+# Refractal
+
+[![PyPI](https://img.shields.io/pypi/v/refractal)](https://pypi.org/project/refractal/)
+[![Python](https://img.shields.io/pypi/pyversions/refractal)](https://pypi.org/project/refractal/)
+[![License](https://img.shields.io/badge/license-Apache--2.0-blue)](https://github.com/nalinraut/refractal/blob/main/LICENSE)
+[![CI](https://github.com/nalinraut/refractal/actions/workflows/ci.yml/badge.svg)](https://github.com/nalinraut/refractal/actions/workflows/ci.yml)
+
+**A closed-loop evaluation compiler.**
+
+Refractal turns a declared set of scenarios into a plan you can inspect before running, 
+executes it on your chosen backend with the policy acting in simulation, and tells you 
+whether the difference between two checkpoints is real.
+
+Backends: local, Docker Compose, Kubernetes.
+
+Run the same evaluation twice and know whether the difference is real.
+
+You have a policy. You changed something. You want to know if the new version is
+better.
+
+Today you run a benchmark, get 78%, run the old one, get 74%, and guess. That
+number hides everything that matters: *which* situations got better, whether four
+points is real or noise, whether some of those "failures" were your container
+crashing, and whether the new version got worse at something while getting better
+overall.
+
+Refractal turns "run a benchmark" into "run an experiment."
+
+```console
+$ pip install "refractal[execute,compare]"
+$ mkdir demo && cd demo
+$ refractal init .
+$ refractal plan catalog --hardware laptop -o plan.json
+  cube-bowl-v1        30 scenarios     360 episodes    2 worker(s)  <=37 min
+  ------------------------------------------------------------------
+  360 episodes across 1 scene(s), 2 worker(s), at most 37 min
+  that is an upper bound: every episode is costed at its full step limit,
+  and episodes that succeed finish sooner.
+  wrote plan.json  (plan_schema 3, plan_id sha256:82b73f180a7a...)
+
+$ refractal run plan.json -o results --catalog catalog
+$ refractal compare results $(python -c "import json;print(json.load(open('plan.json'))['plan_id'])")
+```
+
+About a minute end to end, with no GPU, no simulator and no checkpoints: the
+example runs against a built-in fake benchmark, so the machinery is real and only
+the robot is not.
+
+**[Start here →](https://github.com/nalinraut/refractal/blob/main/docs/getting-started.md)**
+
+## Documentation
+
+| | |
+|---|---|
+| [Getting started](https://github.com/nalinraut/refractal/blob/main/docs/getting-started.md) | Install to verdict, start to finish. |
+| [Writing a catalog](https://github.com/nalinraut/refractal/blob/main/docs/writing-a-catalog.md) | Describing your own experiment, and which edits invalidate existing results. |
+| [Catalog reference](https://github.com/nalinraut/refractal/blob/main/docs/catalog-reference.md) | Every file and every field, for looking things up. A key not listed there is rejected. |
+| [Reading a comparison](https://github.com/nalinraut/refractal/blob/main/docs/reading-a-comparison.md) | Every number in the output and when not to trust it. |
+| [Perturbations](https://github.com/nalinraut/refractal/blob/main/docs/perturbations/index.md) · [declaring](https://github.com/nalinraut/refractal/blob/main/docs/perturbations/declaring.md) · [adapters](https://github.com/nalinraut/refractal/blob/main/docs/perturbations/adapters.md) · [sweeps](https://github.com/nalinraut/refractal/blob/main/docs/perturbations/sweeps.md) | Changing the world mid-episode, when success rates alone cannot separate two checkpoints. |
+| [Adapter contract](https://github.com/nalinraut/refractal/blob/main/docs/adapter-contract.md) · [example](https://github.com/nalinraut/refractal/blob/main/docs/adapter-example.md) | Connecting your own simulator. |
+| [Against vla-eval](https://github.com/nalinraut/refractal/blob/main/docs/backend-vla-eval.md) · [In containers](https://github.com/nalinraut/refractal/blob/main/docs/backend-compose.md) | Running at scale. |
+| [Execution modes](https://github.com/nalinraut/refractal/blob/main/docs/execution-mode.md) · [Releasing](https://github.com/nalinraut/refractal/blob/main/docs/releasing.md) | |
+
+## What it does
+
+**Declares experiments, not benchmark invocations.** You describe scenes, tasks
+and scenario grids in YAML. Refractal expands that into a concrete list of
+episodes, each with a content-addressed identity, works out how many workers the
+run needs and whether they fit in memory, and writes a plan you can read, diff and
+commit before spending anything.
+
+**Compares episode by episode, with statistics that fit the design.** The same
+scenarios face every checkpoint, so the data is *paired*. Refractal reports the
+full 2×2 per task, runs McNemar on scenario-level outcomes, and gates on a
+clustered bootstrap that resamples whole scenarios, because five seeds at one
+pose are one observation of that pose rather than five.
+
+**Refuses rather than guessing.** A plan that oversubscribes VRAM is not emitted.
+A comparison across two scene geometries is blocked, not reported. A worker
+assignment the backend cannot execute is refused at planning time rather than
+discovered forty minutes in.
+
+## Three things it gets right that aggregate scores cannot
+
+**Identity is content, never position.** An episode is a hash of
+`(scene, task, scenario, seed, checkpoint)`. Rename a task and results still
+join; change what the task *means* and they correctly stop. Reordering a list
+cannot silently change what an attempt was.
+
+**Infra failures leave the denominator.** A crashed worker is its own column, not
+a policy failure. And because dropping it breaks the pairing a paired test needs,
+the seed is dropped from *every* checkpoint, with the loss reported rather than
+absorbed.
+
+**Multiplicity is corrected across the whole family.** The gate fires if any
+contrast trips, so every contrast is one family. Three tasks at a nominal 5% is a
+family-wise error rate near 14%; four checkpoints across three tasks is eighteen
+contrasts and about 60%. Holm-adjusted, with the uncorrected number printed so
+the correction does not read as pedantry.
+
+## Comparing more than two checkpoints
+
+Two is the common case, not a special one. A single checkpoint is a comparison of
+size one; three is a training sweep, and produces something a leaderboard cannot:
+
+```
+scene-v1 / task-4   (119 scenarios)
+  rates:  ckpt-46: 49.4%   ckpt-47: 49.8%   ckpt-48: 37.5%   (baseline ckpt-46)
+  outcome patterns:
+       35  none solve
+       18  only ckpt-46, ckpt-47        <- what ckpt-48 lost
+       18  all solve
+       17  only ckpt-47
+        3  only ckpt-48
+  Cochran Q (screen):  Q=19.73 p=0.0005  (66 of 119 scenarios disagree)
+  ckpt-46 -> ckpt-47: +0.005 [-0.059, +0.066]  p=0.8874 holm=1.0000  -> no change
+  ckpt-46 -> ckpt-48: -0.118 [-0.179, -0.059]  p=0.0008 holm=0.0040  -> REGRESSED
+```
+
+"Eighteen scenarios that ckpt-46 and ckpt-47 both solve and ckpt-48 lost" is a
+hypothesis you can go replay. Three lost scenarios would be noise.
+
+## Exit codes
+
+`refractal compare` is meant for CI:
+
+| code | meaning |
+|---|---|
+| 0 | no regression |
+| 1 | regression detected |
+| 2 | **cannot be answered**: different scene geometry, duplicated episodes, or a change to the code that ran them |
+
+The third is not a regression and is never reported as one. "The two runs used
+different geometry" is a different sentence from "the policy got worse", and
+conflating them teaches people to ignore the gate.
+
+## Architecture
+
+Refractal is a compiler, not a runtime. A catalog goes in, a `plan.json` comes
+out, and the plan executes. The schedule is decided once and recorded, which makes
+placement part of the provenance rather than an accident of the day.
+
+| package | does | needs infrastructure? |
+|---|---|---|
+| `refractal.schema` | catalog, validation, identity | no |
+| `refractal.resolve` | catalog → `plan.json` | no |
+| `refractal.build` | facts that need the engine → `build.lock` | yes |
+| `refractal.execute` | runs a plan; three backends | yes |
+| `refractal.perturbations` | timed changes to world, observation or action | no |
+| `refractal.render` | plan → deployment file | no |
+| `refractal.compare` | Parquet → verdict | no |
+
+`refractal plan` runs on a laptop with no simulator, no GPU and no Docker. That is
+enforced by a test rather than by convention: it is the property that lets you
+inspect a plan and its cost before committing to it.
+
+Results are Parquet at an fsspec URI, so `./results` and `s3://bucket/results` are
+the same code path. Every write is atomic, which is what makes resume safe, and
+resume works by episode identity rather than by a counter.
+
+## Backends
+
+Two things are called a backend and it is worth keeping them apart. `refractal
+run --backend X` executes a plan. `refractal render --target Y` writes a
+deployment file that something else executes.
+
+| `refractal run --backend` | |
+|---|---|
+| `local` | simulates outcomes; no infrastructure |
+| `vla-eval` | drives the harness against running model servers |
+| `compose` | one container per worker, over the same entrypoint |
+
+| `refractal render --target` | |
+|---|---|
+| `compose` | a Compose file; `docker compose up` runs it |
+| `k8s` | one Job per worker; `kubectl apply` runs it |
+
+Kubernetes is reached through `render`, not through `run`: Refractal writes the
+Jobs and your cluster schedules them. Both targets render **without Docker or a
+cluster installed**, so you can read the file before running it.
+
+## Install
+
+```console
+pip install refractal                    # plan and build
+pip install "refractal[execute]"         # + run
+pip install "refractal[compare]"         # + compare
+pip install "refractal[vla-eval]"        # + the harness backend
+```
+
+Comparison needs no numerical stack: exact McNemar is a binomial tail, the
+bootstrap is resampling, and Cochran's Q is a permutation test.
+
+## Status
+
+Alpha. Every stage is implemented and tested, and the identity-bearing fields are
+still able to move between versions, which is what the `a` in `0.1.0a1` is for.
+
+Validated against real model servers as well as the synthetic backend: **15,490
+episodes across 13 runs**, including a 600-episode comparison whose success rate
+was predicted from an earlier run's data before it was run and came back within
+two points.
+
+The largest is a 2,700-episode comparison of three checkpoints across three
+LIBERO suites, executed twice under one `plan_id` — once locally, once as one
+container per worker:
+
+| | wall clock | episode work | overall success |
+|---|---|---|---|
+| local, workers in sequence | 317 min | 314 min | 72.6% |
+| compose, workers in parallel | **171 min** | 462 min | 72.0% |
+
+Placement moved the wall clock by 1.86× and the result by 0.6 points. The two
+placements agreed on **92.0%** of episodes — while re-running a *single*
+placement against itself agreed on only **89.7%**, so two deployments differ
+less than one deployment differs from its own rerun. That is the claim the
+content-addressed identity exists to support, and it is measured rather than
+asserted.
+
+The 462 minutes of episode work against 314 is the honest other half:
+parallelism bought real wall clock and was not free, because four workers
+contend where one had the machine to itself.
+
+## What is not here yet
+
+- **Trajectory quality metrics.** The `steps.parquet` schema is declared and
+  nothing writes it, deliberately: writing it against synthetic data would bake in
+  guesses about what a real adapter can record.
+- **`--backend k8s`.** `refractal render --target k8s` writes the Jobs and they
+  run against the same `--worker` entrypoint Compose uses, but nothing supervises
+  them from inside Refractal: no `run` subcommand submits them, watches them, or
+  collects their exit codes. Rendered and applied, not driven.
+- **Resource-shape probing.** `refractal build` carries hand-declared shapes
+  forward and records whether they were measured. A lint refuses a comment
+  claiming a measurement when nothing recorded one.
+
+## Built on vla-evaluation-harness
+
+Refractal is a layer over
+[**`allenai/vla-evaluation-harness`**](https://github.com/allenai/vla-evaluation-harness),
+from the Allen Institute for AI, and does not fork it. The harness runs the
+episodes: it owns the model-server protocol, the benchmark adapters, the
+observation and action specs, and the episode loop. Refractal decides which
+episodes to run, gives them content-addressed identities, and compares the
+results.
+
+Everything `--backend vla-eval` does is the harness doing it. Refractal subclasses
+two of its classes and otherwise stays out of the way, which is why the
+integration is a pinned dependency rather than a vendored copy, and why
+`scripts/verify_harness_claims.py` exists: the assumptions Refractal makes about
+someone else's code are checked rather than assumed.
+
+The harness is Apache-2.0, as is Refractal. No code is copied from it.
+
+If you use Refractal for published work, cite the harness as well. The evaluation
+is theirs; the comparison is ours.
+
+## Acknowledgements
+
+- [`allenai/vla-evaluation-harness`](https://github.com/allenai/vla-evaluation-harness)
+  (Allen Institute for AI, Apache-2.0), which runs every episode.
+- The benchmark suites and model servers it wraps, each under its own licence and
+  each the work of its own authors.
+
+## Licence
+
+Apache-2.0. See [LICENSE](https://github.com/nalinraut/refractal/blob/main/LICENSE)
+and [NOTICE](https://github.com/nalinraut/refractal/blob/main/NOTICE).