CLI Usage Guide¶

The CLI has data, tools and remote groups plus the top-level train command. Use --help on any command for its exact options. Operator examples below require the named generated artifacts.

Installation¶

uv sync
uv run naics-embedder --help

Data Commands¶

data preprocess¶

Download and preprocess all raw NAICS data files. Each code's examples channel holds only its examples-role index entries, per the committed role table (see data roles); the code's other entries are outcome-panel queries. It also writes the redirection table: every cross-reference row and harvested "Excluded" paragraph once, with the codes it names. Each code's exclusion channel is built from that table, and a row that a held-out query leaks into is withheld from the channel. A code without official description text inherits its only child's, and description_source records whose text it is. An absent channel is null. Preprocessing fails if any validation or test query matches training text or an activity phrase.

Requires: conf/data/index_roles.csv
Generates: data/naics_descriptions.parquet, data/naics_index_roles.parquet, data/naics_redirections.parquet

uv run naics-embedder data preprocess

Options: - --source-dir PATH - Read the four Census files from this directory, by file name, instead of downloading them - --force - Overwrite a descriptions file that a configured supervision bundle pins. Refused by default: training against that bundle fails closed once the file changes

data supervision¶

Build one immutable, validated Stage-3 supervision bundle. It holds the codebook, the pair facts (D*, the structural relation, directional explicit exclusions and the unary-pair flag), legacy-compatible distance and relation views and matrices, and training pairs. It also holds the curriculum difficulty thresholds, the index roles and the redirection table. The manifest records the backbone's input window and each channel's texts beyond it. The manifest is written last and the bundle directory is published atomically.

Requires: data/naics_descriptions.parquet, data/naics_index_roles.parquet, data/naics_redirections.parquet (the last two become required bundle members), and the backbone's tokenizer in the local Hugging Face cache
Generates: data/supervision/stage3-supervision-v2/<bundle-id>/ (prints Supervision manifest: <path>; before training, set it as config.supervision.manifest_path in the gitignored conf/local.yaml, which conf/local.example.yaml shows)

uv run naics-embedder data supervision

data relations, data distances, data triplets¶

Deprecated stage commands. Publishing one partial authority would let artifacts from different generations mix, so each prints a migration notice and exits with status 1 without building anything; data supervision builds the complete bundle.

data all¶

Run the full data generation pipeline: preprocess, then build the supervision bundle.

uv run naics-embedder data all

data roles¶

Draw the frozen index-entry role table, once. Every Census index entry gets exactly one role, per code and stratified: examples-channel text, or a training, validation or test query for the outcome panel. No validation or test query matches any training text, exactly or as a near-duplicate (character-trigram Jaccard of at least 0.9). The table is committed and data preprocess applies it. Redrawing it unseals the validation and test splits, so an existing table is replaced only with --force.

Generates: conf/data/index_roles.csv, conf/data/index_roles_provenance.json

uv run naics-embedder data roles --source-dir ~/Downloads/Data

data summaries¶

Build the window-fitting summaries of over-long channel texts, once per backbone (roadmap Stage 6b). Every description, examples or excluded text whose marked form ('description: …', special tokens included) is over the backbone's trained window is summarized by whole units of the text: its sentences, or clauses and segmenter pieces of an over-long sentence, or its examples entries. The frozen backbone (from the local Hugging Face cache) keeps, greedily, the units whose pooled vectors best approximate the whole text's, in source order, while the marked summary fits. Every unit boundary is a boundary of the leakage segmenter, so a summary's segments are a subset of its text's and the leakage checks need no sealed read. The artifact is checked as every reader checks it before it is moved into place; commit it with the pin the command prints, in WINDOW_SUMMARIES (panels/window_summaries.py). An existing artifact is replaced only with --force, and a new artifact needs a new pin.

Generates: conf/data/window_summaries.csv, conf/data/window_summaries_provenance.json

HF_HUB_OFFLINE=1 uv run naics-embedder data summaries

Options: - --descriptions PATH - The descriptions parquet (default: ./data/naics_descriptions.parquet) - --backbone NAME - The backbone (default: data_loader.tokenization.tokenizer_name in conf/config.yaml) - --output PATH - The artifact (default: conf/data/window_summaries.csv) - --force - Replace an existing artifact

data regressor-groups¶

Draw the regressor panel's held-out four-digit groups, once: a fifth of each sector's groups (largest remainder, seeded) from the 980 six-digit codes Stage 1's finding names. The QCEW national slices are read from qcew_dir under the sha256 values pinned in conf/data/regressor_panel.yaml, and the codebook under its codes' fingerprint. The table is committed, and the panel reads it only under the fingerprint pinned there as heldout_groups_sha256. A redraw moves both regressor outer sets, so an existing table is replaced only with --force, and the panel reads a new draw only after a reviewed change of the pin.

Generates: conf/data/regressor_heldout_groups.csv, conf/data/regressor_heldout_groups_provenance.json

uv run naics-embedder data regressor-groups --codebook PATH/naics_codebook.parquet

Tools Commands¶

tools config¶

Display the configuration validated over defaults, including run inputs and settings. The command exits 1 on a missing or invalid file.

uv run naics-embedder tools config --config conf/config.yaml

tools visualize¶

Plot the durable epoch summary. One epoch_metrics.png has four panels: monitor MRR, term and total losses, both learned scales, and per-level radius means with SD bands. Structural metrics are separate diagnostics. The tool reads no console logs or evaluation panels.

uv run naics-embedder tools visualize   --summary checkpoints/reference/epoch_summary.jsonl   --output-dir outputs/visualizations/reference

• --summary PATH is required: the run's epoch_summary.jsonl.
• --output-dir PATH defaults to visualizations beside the summary.

tools outcome-baseline¶

Score the training-free lexical encoder (hashed character trigrams under cosine distance) on the outcome panel's validation split: top-1 accuracy, MRR, Hit@1/5/10 and the level of the lowest common ancestor of the top-1 code and the truth. The read is appended to the selection log; the test split stays sealed.

Requires: data/naics_descriptions.parquet, data/naics_index_roles.parquet

uv run naics-embedder tools outcome-baseline

Options: - --purpose TEXT - Why this read happens; recorded in the selection log - --index-roles PATH, --descriptions PATH - Preprocessing outputs (default: the paths in conf/data/download.yaml) - --log PATH - Selection log (default: logs/selection_log.jsonl, from conf/data/outcome_panel.yaml) - --output PATH - Also write the summary as JSON

tools text-only-table¶

Embed every code's text with the arm's backbone, frozen: each of the four channels is mean-pooled over its tokens, and a code's vector is the mean of its present channels (roadmap D9). The backbone comes from the local Hugging Face cache (default: text_only.backbone in conf/data/regressor_panel.yaml). A channel text over the backbone's trained input window is read as its window-fitting summary, as the arm reads it, so no channel text is truncated, and text_only.max_length may not exceed the window (Req 9). The provenance records the summaries' sha256 (summaries). The decision store records it, and check_text_only compares it with the arm's (D9). The regressor panel reduces the table to the arm's dimension by PCA.

Generates: the table and <stem>_provenance.json beside it

uv run naics-embedder tools text-only-table --descriptions data/naics_descriptions.parquet \
  --output /tmp/text_only.parquet

Options: - --descriptions PATH - The arm's descriptions parquet: the text it reads - --output PATH - Where to write the table - --backbone NAME - The arm's backbone (default: the regressor config's)

tools regressor-panel¶

Score an arm on the regressor panel (roadmap Stage 3). Every comparator (covariates alone, and the arm's coordinates, one-hot (seen regime only), ancestor indicators and the text-only table, each alone and with the covariates) is fitted by ridge on standardized features, with the penalty tuned by nested grouped folds inside the remainder. The outcome is log employment in year t + 1 from year-t features. The seen and held-out regimes are separate panels. The validation split reads only the remainder. The test split opens each regime's sealed outer set first, and both the opening and the read are logged. The output holds one prediction per row, comparator and repeat (the test split has one repeat), with the row's outcome and the chosen penalty.

The panel's outcome setting in conf/data/regressor_panel.yaml (Req 21) is level, the panel above, or growth: log E(t + 1) − log E(t) on the same rows, read as growth_seen and growth_heldout. One opening of a regime's outer set counts for both outcomes (Req 22). This command reads level mode only. Every read scores the arm and the text-only table, and Req 24 records the economic margin before any candidate's growth error exists, so a config set to growth is refused before anything is read. The growth margin read is tools growth-margins's.

uv run naics-embedder tools regressor-panel --coordinates arm.parquet \
  --text-only /tmp/text_only.parquet --codebook PATH/naics_codebook.parquet

Options: - --coordinates PATH - The arm's 2,125-code table in the export form (tools export-table: tangent coordinates at the origin for a hyperbolic arm). Lorentz points and constant columns, such as a log map's zero time coordinate, are refused - --text-only PATH - The text-only table (tools text-only-table) - --codebook PATH - A supervision bundle's naics_codebook.parquet - --regime seen|heldout - Regime to score (repeatable; default: both) - --level INT - NAICS level 2-6 (repeatable; default: 6) - --split validation|test - The test split needs --open-purpose, which is logged - --purpose TEXT - Why this read happens; recorded in the selection log (default: regressor panel <split> read) - --open-purpose TEXT, --reopen-reason TEXT - Why the outer sets are opened, and why again - --log PATH - Selection log (default: logs/selection_log.jsonl) - --output PATH - Write the per-row predictions as parquet

tools export-table¶

Export a checkpoint's 2,125-code table as code, index, level, then float64 coordinates e0 through e{d-1}, in codebook order. Each code goes through the checkpoint's model in eval mode. A hyperbolic arm writes bounded tangents, with unit curvature and r = R * tanh(norm(v) / R) rather than a cap at 2; a Euclidean arm writes v and a spherical arm u = v / norm(v) (geometry arms).

The supervision contract must match the configured bundle. The checkpoint's encoder record is its own, so a dimension-8 checkpoint can export under a dimension-16 config. Pre-req11-v1 checkpoints are refused before model loading. Different window summaries or tokenizer pins are refused. Export writes <stem>_provenance.json, with checkpoint and contract identities, backbone revision, tokenizer/window/summaries identities, descriptions hash, geometry and its coordinates, and table fingerprint.

Use the monitor's selected checkpoint; epoch 1 is an example here:

uv run naics-embedder tools export-table --checkpoint checkpoints/reference/epoch=001.ckpt   --output data/reference/table.parquet

• --checkpoint PATH, --output PATH are required.
• --config PATH defaults to conf/config.yaml.
• KEY=VALUE ... overrides bundle/token-cache configuration. An argument without = is refused.

tools outcome-panel¶

Score the outcome validation split through the checkpoint's query encoder and the table exported from that checkpoint. The read logs table fingerprint and checkpoint hash. Provenance must match the checkpoint and preprocessing pins. The test split stays sealed; there is no --split.

uv run naics-embedder tools outcome-panel --checkpoint checkpoints/reference/epoch=001.ckpt   --table data/reference/table.parquet --purpose 'selected checkpoint validation'

• --checkpoint PATH, --table PATH, --purpose TEXT are required.
• --config PATH defaults to conf/config.yaml.
• --log PATH defaults to the outcome-panel configuration's selection log.
• --output PATH also writes JSON with panel/table/checkpoint identities and scores.
• KEY=VALUE ... takes the same config overrides as export.

tools sweep¶

Keep the store and JSON records under ~/naics-artifacts, outside every worktree. Build one arm record from complete trained seed directories. Before the first export or decision read, the runner checks every seed's epoch coverage, panel fingerprint, training-run id, seed, monitor distance, 21 settings, earliest best checkpoint and exact saved best score. The arm's geometry comes from model.geometry; each checkpoint's encoder record must name it. Both last and selected checkpoints must also match the current config in their saved LoRA rank/alpha/dropout and active MoE expert count/top-k/hidden dimension/load-balancing coefficient. Missing required values are refused; the arm retains its 21-key settings identity. A selected epoch with a versioned sibling is ambiguous and refused. The arm's backbone revision is resolved independently from its cached backbone; it cannot be copied from the text-only comparator. The sweep reads the regressor panel in level mode only (D8): a regressor config set to outcome: growth is refused before any seed is checked or read.

mkdir -p ~/naics-artifacts/records/stage7
uv run naics-embedder tools sweep --runs 'checkpoints/reference-seed-{seed}' \
  --seed 1 --seed 2 --seed 3 --seed 4 --seed 5 --seed 6 --seed 7 --seed 8 --seed 9 --seed 10 \
  --text-only data/reference/text_only.parquet --store ~/naics-artifacts \
  --output ~/naics-artifacts/records/stage7/reference.json --purpose 'reference campaign validation' \
  --name reference --accelerator cuda

• --runs TEXT is required and must contain {seed}.
• --seed INT is required and repeatable; seed ids must be distinct. Decisions need at least five.
• --text-only PATH, --store PATH, --output PATH, --purpose TEXT are required.
• --name TEXT defaults to reference.
• --accelerator TEXT defaults to cuda: the training accelerator for run-settings comparison, even when exports and reads run on the Mac.
• --log PATH overrides the selection log; --config PATH defaults to conf/config.yaml.
• KEY=VALUE ... configures the bundle, encoder and training settings of the arm.

Each seed exports its selected checkpoint and reads outcome validation plus the seen and held-out regressor regimes at level 6. The record stores all monitor reads, selected epoch and training-run id, as well as decision reads and content-addressed artifacts. Test splits stay sealed. Fix reference margins before candidate monitor or decision reads; see the campaign workflow.

tools radius-report¶

Read a checkpoint and its exported table on the CPU, without reading an evaluation panel. The report checks live anchor-radius gradients, per-level SD > 1e-3, positive distinct sector radii and their least gap, manifold error at the largest radius, and chunked all-pairs float32/float64 distance agreement. It reports nonzero gradient norms for all three weighted terms and both scales on the saved seed's epoch-zero, step-zero batch. The radius checks are the hyperbolic arm's: a flat arm's report names its geometry, records radius as null and checks the task and code-to-code terms and both scales.

uv run naics-embedder tools radius-report --checkpoint checkpoints/reference/epoch=001.ckpt   --table data/reference/table.parquet --output data/reference/radius_report.json

• --checkpoint PATH, --table PATH are required and must have matching provenance.
• --output PATH also writes the JSON report.
• --config PATH defaults to conf/config.yaml; KEY=VALUE ... supplies config overrides.

A failed criterion writes the report and exits 1. Distance relative error must be at most 1e-3 on noncoincident pairs. Coincident tangents must have zero float32 training distance; their float64 residual is reported separately. Largest-radius manifold error must be at most 1e-9 * x0**2. This is a verification report, not a panel decision.

tools margins¶

Fix each panel's non-inferiority margin δ from a reference arm (Req 5): δ is --multiple times the reference's across-seed standard deviation of the panel's decision statistic (roadmap D10: per-query MRR on the outcome panel, and the covariates+embedding comparator's mean squared error on each regressor regime at level 6). Fix the margins before any other arm of the decision reads a panel, including its training monitor reads: a decision refuses every run that read before its margins were fixed.

Generates: the margin record (JSON)

uv run naics-embedder tools margins \
  --reference ~/naics-artifacts/records/stage7/reference.json --multiple 3 \
  --name reference-margins --store ~/naics-artifacts \
  --output ~/naics-artifacts/records/stage7/margins.json

Options: - --reference PATH - The reference configuration's arm record, written by the seed-sweep driver (naics_embedder.decision.sweep.run_seed_sweep) - --multiple FLOAT - Each δ as a multiple of the reference's across-seed standard deviation - --name TEXT - Names the margins in the decision records that use them - --store PATH - The artifact store the arm record references - --output PATH - Where to write the margin record; an existing file is refused before any work, never overwritten

tools growth-margins¶

Fix each growth panel's non-inferiority margin δ from a reference-only read (Req 24). The read scores each growth regime's level-6 validation split, growth_seen and growth_heldout, with only the comparators that read no candidate. The seen regime reads the covariates, one-hot and ancestor indicators, each alone and with the covariates; the held-out regime has no one-hot comparator and reads covariates, ancestors and covariates+ancestors. It computes neither embedding nor text_only, so no candidate's growth error exists before δ is recorded. δ is relative_margin in conf/data/decision.yaml (r = 0.05) times the validation MSE of the reference covariates+ancestors. The panel reads growth whatever the regressor config's outcome says, which stays level for tools regressor-panel and tools sweep.

Run it once, before any candidate is read on growth: a decision refuses a candidate read before its margins were fixed. The record names the platform, Python and uv.lock sha256 the read ran under. A decision pairs only bit-identical reference predictions, so every candidate read must run on the same platform and lock.

Generates: the relative margin record (JSON), with the read's selection-log records, and its predictions and scores in the store

uv run naics-embedder tools growth-margins --codebook PATH/naics_codebook.parquet \
  --name growth-margins --store ~/naics-artifacts \
  --output ~/naics-artifacts/records/economic/growth_margins.json

Options: - --codebook PATH - A supervision bundle's naics_codebook.parquet - --name TEXT - Names the margins in the decision records that use them - --store PATH - The artifact store for the read's predictions and scores - --output PATH - Where to write the margin record; an existing file is refused before anything is read, never overwritten - --purpose TEXT - Why the read happens, recorded in the selection log (default: economic margin read) - --log PATH - Selection log (default: the regressor config's) - --lock PATH - The uv.lock the read runs under, which the record names (default: uv.lock)

tools decide¶

Decide among arms under Req 5's rule. With a margin record from tools margins, it decides among two or more arms over the three panels of roadmap D8: the outcome panel and the regressor panel's seen and held-out regimes at level 6. With one from tools growth-margins, it makes the economic decision over the two growth panels (Req 24), below. Each comparison reads the difference Δ on paired resamples of each panel's units (codes with their queries; four-digit groups), with each arm's seeds resampled inside every replicate. A is adopted over B when it is non-inferior on every panel (the 95 % interval's lower bound above −δ) and superior on at least one (the 1 − 0.05/n interval above zero, for n panels: 98⅓ % over D8's three, 97.5 % over the two growth panels). The survivors are the arms no other arm is adopted over, and the tie order picks among them: fewer components, then lower dimension, then non-hyperbolic geometry, then the higher held-out gain over covariates+ancestors (roadmap D11): on regressor_heldout for D8's decision, on growth_heldout for the economic decision, where it is the reference. The number of replicates and the bootstrap seed come from conf/data/decision.yaml, as does the seed floor, min_seeds, which applies to D8's decision only: the economic decision takes none, because each candidate is a single replicate.

The economic decision. Its arms are economic arm records: a deterministic candidate's table, with the provenance its builder wrote, read once on each growth panel by naics_embedder.decision.economic.read_deterministic_arm, which has no command yet. Each candidate is a single replicate. A trained candidate's seeds have no economic record yet. Each candidate is compared with the reference covariates+ancestors, whose predictions in the margin read its read must reproduce bit for bit. Only a candidate adopted over the reference is eligible, and the rule above picks among the eligible; its last tie goes to the higher held-out gain over the reference. With none eligible, the reference stands, and there is no economic artifact.

Generates: the decision record (JSON): the arms with their selection-log records and artifact references, the margins, every comparison with both intervals, the non-dominated set, the tie order, the chosen arm, and each arm's reported gains over the sparse comparators. The economic decision record names its reference, lists the candidates adopted over it and each candidate's MSE for every comparator, and chooses none when the reference stands

uv run naics-embedder tools decide \
  --arm ~/naics-artifacts/records/stage7/candidate.json \
  --arm ~/naics-artifacts/records/stage7/reference.json \
  --margins ~/naics-artifacts/records/stage7/margins.json \
  --name dimension-8 --question "Is dimension 8 enough?" \
  --store ~/naics-artifacts --output ~/naics-artifacts/records/stage7/decision.json

The economic decision, over one candidate:

uv run naics-embedder tools decide \
  --arm ~/naics-artifacts/records/economic/tree.json \
  --margins ~/naics-artifacts/records/economic/growth_margins.json \
  --name economic --question "Which candidate, if any, is the economic artifact?" \
  --store ~/naics-artifacts --output ~/naics-artifacts/records/economic/decision.json

Options: - --arm PATH - An arm record, or for the economic decision an economic arm record; repeat for each arm - --margins PATH - The margin record (tools margins, or tools growth-margins for the economic decision) - --name TEXT, --question TEXT - Name the decision and say what it settles - --store PATH - The artifact store the arm records reference - --output PATH - Where to write the decision record; an existing file is refused before any work, never overwritten

tools diagnostics¶

Report Req 6's structural diagnostics over every codebook code: sector separation (an AUC), within-sector rank correlation (over queries and over sectors), MAP over ancestors, NDCG@5/10/20 with integer lowest-common-ancestor grades, the Pearson correlation of distance with the tree metric D*, and parent retrieval@1/5 without the 522 unary pairs. The report describes an arm: nothing selects on it, and no statistic in it has a threshold. To compare two tables, such as one before and one after graph refinement, report on each.

uv run naics-embedder tools diagnostics --table arm.parquet --geometry hyperbolic \
  --codebook PATH/naics_codebook.parquet

Options: - --table PATH - The arm's code table in the export form (tangent coordinates at the origin for a hyperbolic arm; Lorentz points are refused) - --geometry euclidean|spherical|hyperbolic - The arm's distance: Euclidean, cosine, or the geodesic distance after the exponential map at the origin; a table whose export provenance names another geometry, or cannot be read as JSON, is refused (Req 12) - --codebook PATH - A supervision bundle's naics_codebook.parquet; the table must hold exactly its codes - --curvature FLOAT - A hyperbolic arm's curvature magnitude (default: 1.0) - --output PATH - Also write the report as JSON

Training Commands¶

train¶

Train the shared reference encoder with task, code-to-code and radial terms. The epoch has two streams and a per-epoch code cache; it has no text curriculum or validation loader. The outcome validation monitor's val/outcome_mrr selects the earliest best checkpoint and drives plateau scheduling and early stopping.

uv run naics-embedder train --config conf/config.yaml

• --config PATH defaults to conf/config.yaml.
• --ckpt-path PATH supplies an exact checkpoint; last resolves the current experiment's latest checkpoint. Continue only last, rather than rewinding to an older kept epoch.
• --checkpoint-load-mode TEXT accepts only exact; pre-objective checkpoints and weights-only migration are refused.
• --skip-validation skips advisory data/cache checks. The immutable supervision gate always runs.
• KEY=VALUE ... overrides configuration, such as data_loader.queries_per_step=64 or training.learning_rate=1e-5. A fresh run needs a distinct experiment name.

The configuration is the YAML file, then the gitignored conf/local.yaml in the working directory, then the overrides. conf/local.yaml sets only this machine's supervision manifests (config.supervision.manifest_path, and graph.supervision_manifest_path for HGCN), each a repo-relative path under data/; any other key is refused. NAICS_EMBEDDER_LOCAL_CONFIG=off ignores it.

Defaults are experiment reference, dimension 16, masked mean, radius bound 8 and 128 queries per step. CUDA uses bf16-mixed for the backbone while fusion, projection, head, distances and losses stay float32. CPU and MPS use 32-true. Text training uses one device; devices > 1 is refused. training.trainer.val_check_interval is retained but unread.

Fresh starts refuse used checkpoint directories. Exact resume requires the same experiment directory, bundle, encoder, preprocessing, seed and all 21 run settings, including the epoch budget. A supplementary saved-constructor check compares LoRA rank/alpha/dropout always, and MoE expert count/top-k/hidden dimension/load-balancing coefficient under active MoE fusion, before model/data construction. Missing required values fail closed; inactive MoE controls are ignored. The code cache is rebuilt; monitor_reads.jsonl and epoch_summary.jsonl retain lines through the resumed epoch and then continue. A run that early stopping ended exits 1 with early stopping ended the run at epoch k. A spent epoch budget is a harmless no-op. Launchers must treat the early-stopping message as finished rather than retrying it.

uv run naics-embedder train --ckpt-path last

See exact resume for artifacts and guards. The former sequential stage-chain workflow is removed. Run HGCN separately with uv run python -m naics_embedder.graph_model.hgcn; see HGCN training.

Getting Help¶

uv run naics-embedder --help
uv run naics-embedder data --help
uv run naics-embedder tools --help
uv run naics-embedder train --help
uv run naics-embedder tools sweep --help
uv run naics-embedder tools radius-report --help
uv run naics-embedder tools visualize --help

Configuration Files¶

conf/config.yaml defines the text encoder, objective, two-stream loader and training settings. Panel pins and the selection log are configured under conf/data/; HGCN has conf/graph.yaml. See the configuration API and text training.

Remote Commands¶

The remote workflow guide covers remote up, train, sync, finish and status, including transport configuration, canonical input precedence, exact continuation, coherent checkpoint/history pulls and termination verification. Run these adapters on the Mac; Lambda launch and termination remain the user's actions. Help is nonmutating:

uv run --locked naics-embedder remote --help
uv run --locked naics-embedder remote train --help

A finished resume skips successfully; refusals exit 1. Only a verified safe finish prints Safe to terminate. Local fixture tests do not qualify an image: real-instance qualification is a new-experiment run, which Plan 10's Phase 2 did for the recorded private source and a public release or later campaign repeats.