physical-ai-data-factory git:20260906.4539e93A

Immutable. This exact content is served forever at /api/v1/blob/fb5f2dd5bf91f38b.

---
name: physical-ai-data-factory
description: Use when authoring, running, submitting, or viewing the NVIDIA Physical AI Data Factory blueprint on Nebius + SkyPilot (no OSMO) — annotate → Cosmos Transfer augment → Cosmos Evaluator gate → re-label → Cosmos Curator + FiftyOne curate → Rerun visualize — implemented as an npa.workflow that composes existing workbench tools.
---

# Physical AI Data Factory (NPA-native, no OSMO)

## Source And Attribution

NPA-native re-implementation of the NVIDIA Physical AI Data Factory / Video Data
Augmentation workflow. Design adapted from NVIDIA agent skills
(https://github.com/NVIDIA/skills), primarily `physical-ai-video-data-augmentation`.
Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. Upstream licenses:
Apache-2.0 and CC-BY-4.0. See `skills/NOTICE-NVIDIA-SKILLS`. NPA orchestrates on
SkyPilot (not OSMO) and composes existing workbench tools.

Three NVIDIA components in the pipeline are the real open-source projects, not
NPA look-alikes: **Cosmos Transfer 2.5** augments, **Cosmos Evaluator**
(https://github.com/nvidia-cosmos/cosmos-evaluator, Apache-2.0) grades, and
**Cosmos Curator** (https://github.com/nvidia-cosmos/cosmos-curate, Apache-2.0)
curates. See `skills/NOTICE-NVIDIA-COSMOS-OSS` for exactly which upstream code
runs and where NPA substitutes its own endpoint.

## When To Use

Load this skill when the user wants to author, validate, submit, run, or view the
`physical-ai-data-factory.yaml` blueprint, adapt it to a new dataset, run it on
GPUs, or troubleshoot why a run's Rerun panel / augmented output looks wrong.

Do NOT invent an `npa workbench data-factory` tool — there is none. The blueprint
is pure composition of existing toolRefs; only add real tools with tests.

## What It Is

The independent `paidf-cosmos3.yaml` variant is documented at
`docs/workbench/guides/paidf-cosmos3.md`. It uses real source-video-conditioned
Cosmos 3 `video2video` generation and does not replace or silently change this
skill's Cosmos Transfer 2.5 blueprint.

`npa/workflows/workbench/npa-workflows/physical-ai-data-factory.yaml` — one
`npa.workflow/v0.0.1` spec. Blueprint → NPA stage mapping:

| NVIDIA stage | NPA state | Tool (all REAL — no stubs) | Runtime |
| --- | --- | --- | --- |
| Config Generation | `generate-configs` | `data_factory_stages.generate_configs` (run.shell) | CPU |
| Understand & Annotate | `annotate-original` | `workbench.token_factory.caption` | Token Factory (zero-GPU) |
| Augment & Multiply | `augment` | `workbench.cosmos2.transfer_execute` (real Cosmos Transfer 2.5 `--execute`; uploads video+frames to S3) | GPU |
| Evaluate & Validate | `grade` loop (`evaluate` + `quality-gate`) | `workbench.cosmos_evaluator.evaluate` (real Cosmos Evaluator: hallucination + attribute verification) + `data_factory_stages.grade_gate` | Token Factory + CPU |
| Pseudo-Label Augmented | `annotate-augmented` | `npa workbench token-factory caption` (run.shell) | Token Factory |
| Curation | `cosmos-curate` | `workbench.cosmos_curate.curate` (real Cosmos Curator stages → `clips/` + `metas/v0/`) | CPU |
| Curation review | `curate` | `workbench.fiftyone.curate_augmented` (real FiftyOne Brain, fail closed, merges the curator report) | CPU |
| Visualize | `visualize` (accepted) / `visualize-rejected` (rejected evidence) | `workbench.nurec.visualize` → `data_factory_viz.build_run_rrd` → `reports/sim2real.rrd` | CPU, prebuilt `npa-rerun-viewer` image |
| Finalize | `finalize` | `data_factory_stages.finalize` (real aggregate report) | CPU |

Every stage invokes a real component (enforced by `test_real_components.py` and
the `real-components` skill). The `augment` stage runs the real Cosmos Transfer
2.5 model on GPU via `--execute` and publishes the generated video + extracted
frames to `augment_uri`, which the grade / re-label / curate / visualize stages
consume.

**Config → augment MULTIPLY.** `generate-configs` samples N appearance combos
(from `config.n_augmentations`); the `augment` toolRef passes `--configs-uri`, and
the augment stage runs **one real Cosmos Transfer 2.5 inference per sampled combo**
— each combo's prompt drives a distinct appearance, and each is published as its
own per-clip dir under `cosmos_augmented/<clip>/` with its own `metadata.json`
`variables` (which drives that clip's Rerun label). So an N-augmentation config
yields **N scenario variants**, not one image. The fan-out is surfaced in the
machine-readable artifacts: `variant_count` / `multiply_mode` / `variant_parallelism`
in the augment run-level `manifest.json`, `multiply` (mode + `variant_count`) in the
curation report, and `multiply_mode` / `variant_count` in the finalize report. (A
config with a single combo still emits one variant — `multiply_mode: single-variant`.)

**Multi-GPU fan-out (use ≥4 GPUs).** The multiply loop fans the N GPU-bound
diffusions **across the augment pod's GPUs**, one variant per GPU (pinned via
`CUDA_VISIBLE_DEVICES`), then publishes sequentially in combo order. Concurrency =
`NPA_COSMOS_VARIANT_PARALLELISM` if set, else the auto-detected visible-GPU count,
capped at the variant count (so it is safe on 1 GPU and never pins a variant to a
GPU the pod lacks). Request the GPUs in the spec: `resources.gpu.accelerators:
RTXPRO6000:4` runs 4 variants at once (~one variant's wall-clock instead of 4×).
When multiple model processes need more host RAM than the profile default, set
`NPA_WORKFLOW_GPU_MEMORY=<N>Gi` alongside the accelerator override; it changes
only accelerator-backed profiles, never the CPU stages. Failed variants publish
typed, non-promoting records below the scheduler-attempt prefix while successful
siblings continue into the run manifest. The manifest reports attempted,
successful, and failed counts; an all-failed wave preserves those records but
never publishes an empty candidate batch.
Verified live: a 4-variant run on `RTXPRO6000:4` drove all 4 GPUs to 100%
(4 distinct compute PIDs) and finished in ~14 min end-to-end; the manifest recorded
`variant_parallelism: 4`.

**Multi-node fan-out (`--var augment_nodes=N`).** GPUs-per-pod is the first axis;
nodes are the second, and only the `augment` stage uses it. `resources.gpu` declares
`num_nodes: "{{config.augment_nodes}}"` (default `1`), so submit chooses the block
size without editing the blueprint — concurrent renders = `augment_nodes` × GPUs per
node. Validation requires `augment_nodes <= n_augmentations`, so surplus GPU workers
fail before provisioning. `num_nodes` accepts a `{{config.*}}` token on any profile,
resolved against the `--var`-merged config. Existing clusters also receive a read-only
submit-time snapshot check for enough distinct, Ready, schedulable, product-compatible
nodes after active pod GPU, CPU, memory, init-container, and pod-overhead requests
are subtracted. An active unbound GPU pod makes shared placement indeterminate and
fails this check; task-profile node selectors and required node affinity are applied.

SkyPilot runs the *same* augment command in every pod of the gang, so the stage
shards: node `k` of `N` renders variants `k, k+N, …` (striding keeps the nodes within
one variant of each other) with node-local GPU pins and publishes clips plus
`manifest-rank-<k>.json` under
`cosmos_augmented/_attempts/<attempt-id>/`. **Rank 0 is the join**: it waits for all N
current-attempt shards and conditionally commits the usual
`cosmos_augmented/manifest.json` in sampled
combo order, adding `node_count` and a per-rank `shards` block; a rank that never
reports fails the stage by name instead of publishing an understated fan-out. That
wait has no default deadline — a sibling's remaining work is however long its
diffusions take. It periodically reports elapsed time and missing/received ranks;
`NPA_COSMOS_SHARD_JOIN_TIMEOUT_S` opts into a visible deterministic deadline for a
live-but-hung sibling. The rank-0 identity rendezvous is also unbounded by default;
`NPA_COSMOS_IDENTITY_TIMEOUT_S` is its separate explicit opt-in bound. SkyPilot
0.12.2 intentionally preserves its task id across managed recovery and exposes no
globally ordered recovery epoch to the workload. The durable NPA runtime therefore
pre-issues an ordered wave-sequence/explicit-attempt fence. Rank 0 may claim only
that token and shares its fresh attempt id with the exact ordered gang. An inner
SkyPilot recovery retains the token and cannot supersede an existing same-token
claim; it may safely become the first claimant if the prior worker died before
claiming. A configured NPA retry gets a higher token after the prior job is terminal. Final publication is
compare-and-swap fenced by that claim, so a late old worker stays beneath its old
attempt prefix and an escaped old leader cannot replace the newer canonical claim.
Downstream consumers follow only the executed canonical manifest, never enumerate
`_attempts/`. With `augment_nodes=1` no shard file is written, but the scheduler
claim, attempt-private clip prefix, and conditional canonical commit still fence a
late process from a prior loop or recovery.

Cosmos Transfer 2.5 itself also supports `torchrun --nproc_per_node=N` context
parallelism for *one* clip; NPA does not use it, because one-variant-per-GPU gives a
better throughput for a multiply fan-out. A single-variant run therefore does not go
faster on more GPUs.

**Authoring from chat (agent).** The NPA chat agent can WRITE this blueprint. Ask
it e.g. *"write me a paidf workflow: augment my robot clips and fan out 4 scenarios
on at least 4 RTX 6000 PRO GPUs"* — the deterministic router classifies
`create_data_factory_workflow`, `agent_workflow.choose_workflow_template` selects the
`physical-ai-data-factory` template, and `extract_data_factory_params` parses the
fan-out count → `config.n_augmentations`, the GPU count → `resources.gpu.accelerators`
+ `config.variant_parallelism` (capped to the GPU count), and the free-form
augmentation subject → `config.augment_subject`. The generated YAML is validated +
planned before it is returned (chat only emits runnable specs). `generate_data_factory_yaml(user_text=...)`
is the direct entry point; `generate_workflow_draft(intent="create_data_factory_workflow", user_text=...)`
is the chat path.

Chat authoring fails closed above 64 augmented variants or 8 GPUs. These are
generation ceilings, not workflow/job-count budgets: an operator can still edit
and validate a larger hand-authored spec after confirming real cluster capacity.

**Input conditioning (real augmentation of the caller's input).** The managed
`workbench.cosmos2.transfer_execute` path always conditions on the PAIDF run's
input. Its `config.trigger_uri` must contain captionable PNG/JPEG frames and may
also contain a supported video (`.mp4`, `.mov`, `.webm`, `.mkv`, or `.avi`). The
augment uses the first video when present; for a PAIDF frame-only prefix it
assembles those frames into a temporary 1280x720, 93-frame clip inside the GPU
runner. An empty, inaccessible, or image/video-free input fails closed before
inference. Bundled upstream media was removed for redistribution reasons and is
not a fallback. The runner builds a controlnet spec with `video_path` = that clip
and the selected `config.augment_control` signal, and the sampled
appearance prompt drives the new look — so the output preserves the input's
structure/motion with a new appearance. Generic direct CLI callers remain strict:
they opt in with `--condition-on-input` or `--input-video <path|s3://>` and must
supply a video. Conditioned runs record `mode: cosmos_transfer2.5_gpu` +
`input_conditioned: true` + `conditioned_input` in the augment `metadata.json` /
`manifest.json`, which the agent's provenance panel surfaces.

**Segmentation conditioning and region masks (`--var augment_control=seg`).**
`edge` (Canny), `vis` (bilateral blur), and `seg` (GroundingDINO-base + SAM2) may
be derived from the staged input. `depth` is deliberately precomputed-only and
requires `augment_control_asset_uri` produced by an operator-owned permissive,
weight-free method. NPA does not download, execute, or validate Video Depth Anything
Large or Small weights. Each modality selects an exact pinned ControlNet checkpoint
from `nvidia/Cosmos-Transfer2.5-2B`; submit verifies the caller-owned HF token can
access that exact revision/file before provisioning or GPU work. Token presence is
not treated as license consent. NPA used to rewrite requests outside `edge`/`vis` silently;
an unsupported modality now fails closed instead.

- **What seg buys you.** `edge` preserves every texture edge, so a prompt that
  restyles a surface fights the old material's edge detail. `seg` preserves class
  boundaries only, which lets the prompt change what a region is *made of* while
  keeping the region's shape and motion.
- **`config.augment_control_prompt`** names what to segment (`"robot arm, conveyor,
  bin"`). Upstream defaults it to the first 128 words of the appearance prompt.
- **Region masks** restrict any modality to part of the frame: white pixels are
  where the control applies, black pixels follow the prompt freely.
  `config.augment_mask_prompt` has SAM2 segment the region from text;
  `config.augment_mask_asset_uri` supplies a precomputed binary spatiotemporal mask
  video. They are mutually exclusive — upstream accepts one or the other.
- **`config.augment_control_weight`** is finite and bounded `0.0`–`1.0`. The
  shared semantic contract rejects bad weights, mask mutual exclusion, non-seg
  control prompts, missing depth assets, and nodes exceeding variants during
  validate/plan/submit, before image, cluster, or GPU work.
- **`config.augment_control_asset_uri`** substitutes a precomputed control video
  (e.g. a segmentation map from an earlier pipeline) for the on-the-fly one. A named
  asset that does not exist fails rather than quietly reverting to on-the-fly.
- **Published conditioning.** The control map and mask land under
  `config.augment_control_uri` as `cosmos_control/<clip>/control_<modality>.mp4`,
  `mask_<modality>.mp4`, and extracted frames beneath each. That prefix is a
  **sibling** of `cosmos_augmented/`, never a child: `cosmos_evaluator` treats every
  child directory of the augment prefix as a variant and falls back to the
  alphabetically first PNG inside one, so a nested `control/` would hand the
  attribute-verify VLM a segmentation map instead of the frame it must grade. Rerun
  logs them as `control/<clip>/control_<modality>` next to `augmented/<clip>`, and
  the augment `manifest.json` records `control`, `control_weight`,
  `control_prompt`, `mask_prompt`, and `control_uris`.

Example:

```bash
npa workbench workflow submit physical-ai-data-factory.yaml --run-id <id> \
  --var augment_control=seg \
  --var augment_control_prompt="robot arm, conveyor, bin" \
  --var augment_mask_prompt="robot arm"
```

`NPA_COSMOS_CONTROL`, `NPA_COSMOS_CONTROL_PROMPT`, `NPA_COSMOS_CONTROL_ASSET`,
`NPA_COSMOS_MASK_PROMPT`, and `NPA_COSMOS_MASK_ASSET` override the same knobs for a
submit that cannot change the toolRef argv.

**Cosmos Evaluator grading (`evaluate` stage).** `npa workbench cosmos-evaluator
evaluate` runs two of upstream's checks per augmented variant and writes
`grade/cosmos_evaluator.json` (schema `npa.cosmos_evaluator.report.v1`):

- *attribute verification* — upstream's protocol: an LLM writes one
  multiple-choice question per sampled appearance attribute (guided JSON schema,
  with upstream's tolerant text fallback), then a VLM answers it from a frame of
  the variant. Both hops run on Token Factory, because upstream drives them
  through a configurable OpenAI-compatible endpoint. The sampled combo is
  upstream's `selected_variables` and `APPEARANCE_VARIABLES` is its
  `variable_options`, so a variant that ignored its prompt fails.
- *hallucination* — per-frame dynamic-mask comparison of the source clip against
  the variant. CPU only. It delegates to upstream's own `HallucinationProcessor`
  when a checkout is importable (`NPA_COSMOS_EVALUATOR_SRC`, else
  `/opt/cosmos-evaluator`) and otherwise runs the in-repo port of the same
  algorithm; the result's `engine` field says which ran, and the two agree to
  ~1e-3. Managed variants are input-conditioned, so this comparison contributes
  to their score. For a generic unconditioned transfer the source and output are
  different scenes, so it remains informational and the score is the attribute
  pass rate.

`grade_gate` thresholds on that report's `score`. It also still accepts the older
`vlm_eval` report (`vlm_eval_stub.json`, a LEGACY filename of the vlm_eval tool's
`RESULT_FILENAME`, never a stubbed stage), so runs started before the `evaluate`
stage existed keep grading. Both filenames come from the producing tool's own
constant, so the gate cannot drift from its producer.

**Cosmos Curator curation (`cosmos-curate` stage).** `npa workbench cosmos-curate
curate-augmented` drives upstream's real stage classes in-process — no Ray
scheduler, no GPU: `VideoDownloader` → `FixedStrideExtractorStage` →
`ClipTranscodingStage` → `MotionVectorDecodeStage` + `MotionFilterStage` →
`ClipWriterStage`. It writes upstream's canonical tree under `curated_clips_uri`
(`clips/<clip-uuid>.mp4`, `metas/v0/<clip-uuid>.json` with real per-clip motion
scores, `processed_videos/`) plus a summary at `curator_report_uri`, which the
`curate` review stage merges into its report under `cosmos_curator`.

Runs in the `npa-cosmos-curate` image, which bakes a pinned upstream checkout and
a conda-forge ffmpeg carrying `libopenh264` — upstream's transcoding stage accepts
only `libopenh264` or `h264_nvenc`, and Debian/Ubuntu ffmpeg builds have neither.
Outside that image the stage records `engine: unavailable` with the exact reason
(`npa workbench cosmos-curate engine` prints the same diagnosis) rather than
pretending to curate, and the FiftyOne review stage still runs. Operators with the
full curator container can instead run upstream's GPU pipeline — `npa workbench
cosmos-curate plan-pipeline` prints the documented `video-pipeline split` command,
which adds TransNetV2 shot detection, aesthetic filtering, embeddings, and VLM
captioning; both paths write the same layout, so the ingest reads either.

## Containers And Model Weights

Both NVIDIA tools are containerized as workbench images, and **neither image
carries model weights**. That is a licensing boundary, not an optimization:
upstream's *code* is Apache-2.0 and redistributable, its *weights* are not ours to
ship. A build-time check in each Dockerfile fails if a weight file is present, a
guardrail test (`npa/tests/docker/test_cosmos_oss_images.py`) fails if a Dockerfile
grows a build-time download, and both checkouts are fetched with
`GIT_LFS_SKIP_SMUDGE=1` so upstream's Git-LFS payloads never enter a layer.

| Image | Tier | Weights | Credential at run time |
| --- | --- | --- | --- |
| `npa-cosmos-evaluator` | job, CPU | none needed | `NEBIUS_TOKEN_FACTORY_KEY`, for attribute verification only |
| `npa-cosmos-curate` | job, CPU | none baked; GPU stages fetch on demand | `HF_TOKEN`, for `fetch-models` |

**The evaluator needs no weights at all.** The hallucination check is classical
computer vision, and attribute verification calls a hosted VLM instead of loading
one — so its golden eval runs with `--network none`. Upstream's *objects/obstacle*
check would need an EULA-gated SegFormer ONNX and a CWIP checkpoint; those stay as
LFS pointers and that check is not wired. Anyone who needs it must accept
upstream's EULA and fetch the weights themselves.

**The curator's GPU stages do need weights**, so the image fetches them at run time
with the operator's own Hugging Face token, into a `/config/models` volume that
survives across runs:

```bash
docker run --rm -e HF_TOKEN=... -v curator-weights:/config/models \
  <registry>/npa-cosmos-curate:0.1.0 fetch-models --models split-annotate
docker run --rm -v curator-weights:/config/models \
  <registry>/npa-cosmos-curate:0.1.0 models --output text
```

Model sets name a capability, and each set's membership mirrors the
`model_id_names` of the upstream model class that stage instantiates:
`split-transnetv2`, `embed-internvideo2`, `embed-cosmos-embed1`,
`filter-aesthetic`, `caption-qwen`, `dataset-t5`, and `split-annotate` (what
upstream's `video-pipeline split` needs with default flags). The model ids **and
their pinned revisions come from upstream's own registry**
(`cosmos_curator/configs/all_models.json`), and the download is upstream's own
`huggingface_hub` call, so a pin moves only when the checkout does — never because
NPA hardcoded one. `fetch-models` skips anything already complete, records
per-model failures instead of aborting the batch (a gated repo shows up as its own
403), and refuses outright with an actionable message when no token is set.

NGC versus HF: every curator model is a Hugging Face repo, so `HF_TOKEN` is the
credential that fetches weights. `NGC_API_KEY` is what pulls NVIDIA *containers*
(and what upstream's own NVCF path uses); `models` reports whether each is visible
so an operator can tell which one is missing.

Both images are mode-based (`engine`, `smoke`, plus each tool's commands), so one
image serves the workflow stage, the golden eval, and interactive debugging.

**Custom images must satisfy SkyPilot's Kubernetes setup.** Its provisioner runs, in
the image and as the image's own user, `$(prefix_cmd) apt install openssh-server
rsync -y` then `service ssh restart`, where `prefix_cmd` is `sudo` for a non-root
user. An image missing `sudo` or those packages fails that script, the container
exits, and SkyPilot reports `container not found ("ray-node")` — which reads like a
scheduling fault and is the reason operators historically reached for unpinned
submits. All three Cosmos images install
`openssh-server`, `rsync`, and `sudo` and grant their user passwordless sudo;
`npa/tests/docker/test_cosmos_oss_images.py` fails if that regresses. The same test
covers the entrypoint contract: a bare `ENTRYPOINT ["/bin/bash"]` swallows the args
Kubernetes passes, so an entrypoint must exec its arguments.

Verified Token Factory model roles: `MiniMaxAI/MiniMax-M3` (VLM),
`nvidia/Nemotron-3_5-Lightning` (LLM), `MiniMaxAI/MiniMax-M3`
(hosted reasoning critic). Cosmos Transfer 2.5 is the GPU augment engine, not a
Token Factory model.

## Commands

```bash
SPEC=npa/workflows/workbench/npa-workflows/physical-ai-data-factory.yaml
npa workbench workflow validate-spec "$SPEC" --json
# --var bucket= is required for a meaningful plan; without it the spec's
# `example-bucket` placeholder is planned (plan-spec warns). The shipped
# plan_assume_decision previews the accepted path; runtime still reads S3.
npa workbench workflow plan-spec "$SPEC" --run-id demo \
  --var bucket=<bucket> --json

# Prerequisites, in order, on a fresh machine/account:
npa workbench health preflight
npa workbench health access --capability paidf   # Cosmos Transfer gate must PASS
npa skypilot bootstrap                          # persists skypilot.sky_bin
npa provision-if-absent --project <alias> --cluster-name <context> \
  --cpu-nodes 1 --cpu-preset 8vcpu-32gb \
  --accelerator RTXPRO6000:1                    # CPU hosts controller + CPU stages
# Submit stages missing/outdated content-addressed NPA source automatically.
# `stage-src` or submit `--stage-src` remains the explicit force/restage path.

# Render/submit on GPUs:
npa workbench workflow submit "$SPEC" --run-id "$(date -u +paidf-%Y%m%dt%H%M%sz)" \
  --assume-decision promote_checkpoint --var bucket=<bucket> \
  --var n_augmentations=1 \
  --infra k8s/<context> \
  --secret-env NEBIUS_TOKEN_FACTORY_KEY --secret-env AWS_ACCESS_KEY_ID \
  --secret-env AWS_SECRET_ACCESS_KEY --secret-env HF_TOKEN
```

The secret names above are resolved from the environment first and then the
selected project's configured NPA credentials; operators do not re-export values
already stored by `npa configure`. With no input selector, submit fetches,
checksum-verifies, caches, normalizes, and stages the pinned real RoboPro starter.
Use `--input-video`, `--input-uri`, or `--lerobot-uri` to replace it; LeRobot
selection may add an exact `--lerobot-camera` and non-negative
`--lerobot-episode`, and materializes only the chosen observation video. For a
subject-sensitive run, privately review representative beginning, middle, and
end frames and pass `--require-explicit-lerobot-selection`; it fails before
object-store access unless both selectors were explicitly supplied. The flag is
data-agnostic and records no selectors or subject labels in Git. Omitting it
retains the compatibility defaults and is not semantic-subject proof. Use
`--seed-fixture` only for explicitly synthetic developer/test input.
The one-variant override keeps the first real run decisive; omit it for the
spec's default two-variant multiply or raise it with the requested GPU count.

`submit` reports missing prerequisites together: SkyPilot, source staging,
bucket/S3, the four forwarded runtime secrets, gated Cosmos Transfer access,
and a Ready, schedulable, appropriately untainted node with 6 CPU/24 GiB
allocatable for one PAIDF CPU stage plus the SkyPilot controller. A qualifying
GPU node is valid. Image preflight proves each selected manifest;
private Nebius pulls refresh the Kubernetes secret before launch. `--plan-only`
skips runtime-only checks; `--skip-preflight` bypasses them.

The immutable infrastructure plan also checks boot-disk count and
`compute.disk.size.network-ssd` byte capacity before any mutation. The shipped
one-CPU/one-GPU cluster requires 1,151 GiB (128 + 1,023); the README path with a
new 100 GiB agent requires 1,251 GiB. JSON and human output include exact bytes
and GiB for required, available, and shortfall. Preemptible GPU selection does
not change these disk requirements.

## Key Operational Notes

- **Prepare a verified video before GPU work.** The submit path selects the
  pinned RoboPro physical capture by default, or an explicit `--input-video` /
  `--input-uri` / `--lerobot-uri`; LeRobot inputs validate `meta/info.json`, a
  declared video feature, and the selected episode before media staging. It
  validates H.264 MP4 media, verifies the default digest,
  caches/reuses safely, stages `source.mp4`, and derives the exact
  `conditioning.mp4` plus caption frames. `--seed-fixture` is the only synthetic
  geometry path. Conflicts, offline cache misses, invalid media, or checksum
  failures stop before automatic provisioning and never fall back. The catalog's
  mandatory `--condition-on-input` makes the staged conditioning clip the real
  Cosmos control. Consequently the Dataset & provenance tab and the full
  `reports/sim2real.rrd` Rerun recording is built after grading on both outcomes:
  accepted runs include downstream captions/curation/finalization when available;
  rejected runs get a partial evidence recording before the explicit fail-closed
  state. Runs that fail before usable input/augmented frames exist cannot produce
  an RRD.
- **Schema is not semantic proof.** A valid LeRobot `meta/info.json`, checksum,
  or automatic camera choice cannot establish the depicted subject. For a
  subject-sensitive run, inspect representative frames outside the repository,
  supply exact camera and episode selectors, and enable
  `--require-explicit-lerobot-selection` so compatibility defaults cannot choose
  a different stream. Ambiguous visual candidates fail closed operationally.
- **SAM2 is explicit and optional.** `segmentation_mode: off` is the default and
  keeps the non-SAM path unchanged. `sam2-auto` invokes upstream Meta SAM2 in the
  Cosmos GPU task once per source clip, publishes an exact
  frame-aligned mask contract, and reuses it across augmentation variants for
  protected source-chroma/luma-bounded compositing. Raw prompt coordinates are
  not a workflow config surface. Keep the official source and model revisions
  immutable and fail closed on an incomplete mask sequence. The SAM2 masks are
  distinct from Cosmos Evaluator comparison masks.
- **Control A/B prompt sampling.** Set the same non-sensitive
  `augmentation_seed` on two fresh runs when measuring an optional component.
  Empty preserves deterministic run-ID-derived sampling.
- **Refinement policy and precedence.** The first pass retains Cosmos Transfer's
  established `control_weight=1.0` with guidance `3`. A failed gate consumes the
  durable decision/report contract and must select a different in-bounds pair
  (the default retry lowers guidance). Attempts are immutable JSON objects with
  commit markers; only the current pointer is mutable. Explicit CLI values are
  overridden by `NPA_COSMOS_*` conditioning variables, and a validated refinement
  artifact overrides both so ambient worker settings cannot change a committed
  retry. Prior live evidence showed that stronger control is non-monotonic with
  attribute score, so this is compatibility policy rather than an improvement
  claim.
Run either NVIDIA component on its own, against a run prefix or local files:

```bash
# Which engine will each resolve to here, and why?
npa workbench cosmos-evaluator engine --output text
npa workbench cosmos-curate    engine --output text

# Grade one run's variants (writes grade/cosmos_evaluator.json).
npa workbench cosmos-evaluator evaluate \
  --augment-uri s3://<bucket>/physical-ai-data-factory/<run>/cosmos_augmented/ \
  --output-uri  s3://<bucket>/physical-ai-data-factory/<run>/grade/ \
  --input-uri   s3://<bucket>/physical-ai-data-factory/<run>/input/ \
  --configs-uri s3://<bucket>/physical-ai-data-factory/<run>/configs/

# One check at a time (both take local paths).
npa workbench cosmos-evaluator hallucination --original-video a.mp4 --augmented-video b.mp4
npa workbench cosmos-evaluator attribute-verify --video b.mp4 \
  --variables '{"color_grade": "warm"}' --options '{"color_grade": ["warm","cool","neutral"]}'

# Curate one run's variants (writes the curator tree + cosmos_curator.json).
npa workbench cosmos-curate curate-augmented \
  --augment-uri s3://<bucket>/physical-ai-data-factory/<run>/cosmos_augmented/ \
  --curated-uri s3://<bucket>/physical-ai-data-factory/<run>/curation/cosmos_curator/

# Curate a local directory of clips.
npa workbench cosmos-curate curate-videos --input-dir ./clips --output-dir ./curated
```

## Key Operational Notes

- **Cancellation distinguishes planning from launch.** A durable
  planned/reserved/staged ledger with no workflow, stage, controller, or job ID
  returns `NOT_SUBMITTED` without S3 or SkyPilot. Once any durable evidence says
  submission began, missing S3/SkyPilot remains `VERIFICATION_UNAVAILABLE` until
  exact provider or terminal receipt evidence resolves it. Receipt/exact flags
  follow exact > receipt > live-config precedence and conflicts fail closed.
- **Managed-job identity is per stage/wave attempt.** Runtime state records each
  wave key, encoded stage members, attempt number, job ID/name, timestamps, and
  outcome. Status and cancellation reconcile those records independently; they
  never copy one discovered job ID (for example `8`) onto every stage. Parallel
  JobGroup members share an ID only because the durable wave proves that
  membership. Retries retain all attempts and select the final attempt
  deterministically; missing/conflicting evidence is `unknown`/ambiguous rather
  than an invented mapping. Legacy root-ID fan-out remains only for the proven
  single-managed-job manifest contract.
- **Controller launch is a transaction, including wave 1 and resume.** NPA
  requires stable exact-context Kubernetes API readiness, serializes the durable
  logical wave/attempt identity locally, and reconciles structured SkyPilot queue
  evidence before retry or cancellation. A client-side refusal after acceptance
  adopts the immutable job ID; authoritative absence plus transient transport
  may recover within the same submit; indeterminate existence blocks and prints
  the exact `--resume-run <same-id>` remedy. No ID means cancellation is
  `not_applicable`, never a name-based teardown or a fabricated `CANCELLED` state.
- **Image and completion evidence are digest/run bound.** Selected PAIDF images
  satisfy the SkyPilot 0.12.2 bootstrap contract at immutable digest; FiftyOne
  uses its declared non-root user and UID-0 overrides are forbidden. A terminal
  current-schema ten-wave ledger with exact durable artifacts remains terminal
  after managed jobs disappear. Stale planning cannot produce `NOT_SUBMITTED`;
  resume skips succeeded waves without launch.
  Supply distinct validation artifacts with repeatable
  `--image-override TOOL_REF=IMAGE`, using the same exact mapping for preflight,
  initial submit, and resume; each verified digest is rendered only for its tool.
- **Neither NVIDIA component fetches code at run time.** The evaluator looks for
  a checkout at `NPA_COSMOS_EVALUATOR_SRC` / `/opt/cosmos-evaluator` and falls
  back to the in-repo port; the curator requires `NPA_COSMOS_CURATE_SRC` /
  `/opt/cosmos-curate` (baked into `npa-cosmos-curate`) and reports
  `engine: unavailable` without one. If a report says `unavailable`, read its
  `reason` — it names the missing piece (checkout, importability, or ffmpeg
  encoder) instead of guessing.
- **Curator encoder gotcha.** `ffmpeg -encoders | grep -E 'libopenh264|h264_nvenc'`
  must match something, or upstream's `ClipTranscodingStage` cannot write clips.
  Debian/Ubuntu ffmpeg has neither; conda-forge's build carries `libopenh264`, and
  any GPU node's ffmpeg carries `h264_nvenc`.
- **Curator needs Python >= 3.12** (upstream declares
  `requires-python >=3.12,<3.13`). On an older interpreter the failure otherwise
  surfaces as `cannot import name 'Self' from 'typing'` from deep inside an
  upstream import, so the availability probe checks the version first and says so.
  The `npa-cosmos-curate` image is 3.12; a dev box on 3.10/3.11 needs its own
  3.12 environment (`uv venv --python 3.12`) to run the curator locally.
- **A VLM can answer outside the options it was given.** Token Factory's VLM does
  return `D` for a three-option question. That reads as `UNKNOWN` (a failed check
  with no verdict about the pixels), not as a concrete wrong answer, so check
  `vlm_answer` before drawing conclusions from a failed attribute.
- **GPU accelerator name is cluster-specific.** The spec uses canonical
  `RTXPRO6000:1`; some clusters advertise `RTXPRO-6000-BLACKWELL-SERVER-EDITION`.
  If `sky` reports `FAILED_PRECHECKS` / no matching resources, check
  `sky gpus list` and resubmit with the cluster's accelerator name.
- **Reproducible deploy/submit:** unset a stale `NEBIUS_IAM_TOKEN` before
  `sky`/`terraform` (the Nebius provider prefers the ambient token over the
  fresh one). `provisioner._run` scrubs it for agent deploys.
- **Augment output contract (per-clip layout):** `cosmos2.transfer` writes a
  contract manifest by default; with `--execute` on S3 output,
  `publish_transfer_to_s3` uploads the real Cosmos Transfer 2.5 result in the
  **per-clip** layout the consumers require:
  `cosmos_augmented/<clip>/{augmented_video.mp4, frame-*.png, metadata.json}`
  plus a run-level `cosmos_augmented/manifest.json`. Every scheduler-managed
  augment puts its clip dirs below `_attempts/<attempt-id>/`; a multi-node augment
  also puts `manifest-rank-<k>.json` there. Consumers resolve only variants named
  by the executed canonical manifest. `build_run_rrd` reads each selected clip's
  `metadata.json` for its Rerun label. Producer and consumers share this shape;
- **Real FiftyOne curation:** the `curate` stage invokes
  `workbench.fiftyone.curate_augmented` in the `npa-fiftyone` image with
  `--require-fiftyone`. It builds a `fiftyone.Dataset`,
  computes a GPU-free per-variant embedding (downsampled RGB + color histogram),
  and runs `compute_uniqueness` + `compute_similarity().find_duplicates()` +
  `compute_visualization(method="pca")`. The report gains `curation_engine:
  fiftyone-brain`, per-variant `uniqueness`, near-duplicate clusters, and a
  kept/dropped `selection` (schema stays `npa.fiftyone.curation.v1` — new fields
  are additive). A completed real Cosmos Curator report is required, and if
  Curator, Brain, or its database is unavailable, every caller fails this stage;
  no report-only success mode remains. `--require-fiftyone` is retained as an
  explicit workflow contract marker.
  The agent's Dataset & provenance tab surfaces uniqueness + kept/dropped per
  card, curation stats, original-versus-synthetic grouping, and source metadata
  (`build_fiftyone_dataset`).
  `test_publish_transfer_layout_interoperates_with_curate_and_viz` guards it. The
  augment stage "multiplies": it runs one inference per sampled combo and emits
  one clip dir per variant (`publish_transfer_clip` per clip + a single
  `write_run_manifest`), so N sampled combos → N clip dirs.
- **Full or rejected-evidence pipeline in the Rerun panel:** `build_run_rrd` logs the run's input
  frames and each augmented clip's frames/label PLUS a static text doc per stage
  — sampled scenarios (config), the attribute-verify / hallucination grade + gate
  decision, final quality disposition, any curation/finalize reports, and a stage
  log — under the `pipeline/*` entities. Rejected runs take `visualize-rejected`
  before `reject-quality`, so the evidence survives while promotion remains
  fail-closed.
- **Viewing in the NPA agent:** every stage lands under one S3 run prefix
  (`input/ configs/ labeled_original/ cosmos_augmented/ grade/ labeled_augmented/
  curation/ reports/`). The `visualize` stage writes `reports/sim2real.rrd`,
  which the agent's embedded Rerun viewer renders. Browse via
  `/api/artifacts/runs?prefix=physical-ai-data-factory`. See
  `docs/workbench/guides/physical-ai-data-factory.md`.
- **Restricted-egress visualization:** the `visualize` state uses
  `workbench.nurec.visualize`, which the renderer pins to the prebuilt
  `npa-rerun-viewer` image. It never performs `pip install` at task runtime.

## Testing (live-infra is a priority)

Follow `skills/atomic/testing-conventions/SKILL.md` ("Live-Infra Testing Is A
Priority"). The blueprint is registered in `SUBMIT_LIVE_MATRIX` and
`DYNAMIC_SPECS`; smoke via `test_all_workflow_yamls.py`; the Rerun builder via
`npa/tests/workflows/test_data_factory_viz.py`.