gradlab 0.1.0__tar.gz

gradlab-0.1.0/.gitignore

@@ -0,0 +1,70 @@
+# Python bytecode/caches
+__pycache__/
+*.py[cod]
+*$py.class
+
+# Virtual environments and local package managers
+.venv/
+.venv*/
+venv/
+env/
+.uv-cache/
+node_modules/
+
+# Python build artifacts
+build/
+dist/
+*.egg-info/
+.eggs/
+
+# Test, type-checker, and lint caches
+.pytest_cache/
+.ruff_cache/
+.mypy_cache/
+.pyre/
+.coverage
+coverage.xml
+htmlcov/
+
+# Local config/secrets
+.env
+.env.*
+!.env.example
+
+# Matplotlib and notebook caches
+.matplotlib/
+.ipynb_checkpoints/
+
+# Training outputs
+runs/
+logs/
+models/
+videos/
+wandb/
+experiments/goals/*/runs/
+experiments/goals/*/logs/
+experiments/goals/*/models/
+experiments/goals/*/videos/
+experiments/goals/*/wandb/
+experiments/goals/*/artifacts/
+
+# RL/model artifacts generated outside ignored run dirs
+*.ckpt
+*.onnx
+*.pt
+*.pth
+*.npz
+*.safetensors
+*.tfevents.*
+events.out.tfevents.*
+
+# OS/editor noise
+.DS_Store
+.idea/
+.vscode/
+*.swp
+*.swo
+
+.secret/
+
+artifacts/

gradlab-0.1.0/METRICS.md

@@ -0,0 +1,344 @@
+# Metrics schema v10
+
+This file is the human contract for gradlab telemetry. The Python registry in
+`src/gradlab/metric_names.py` is the executable source of truth. Every emitted metric must match an
+exact registry entry or a bounded template.
+
+## Surfaces and dimensions
+
+- W&B is the authoritative scientific metric surface. One supervisor process inside the training
+  container is the only process allowed to open and write the logical W&B run.
+- The learner writes structured events only to its embedded SQLite WAL outbox. It performs no
+  network I/O for metrics, checkpoint publication, or evaluation dispatch.
+- Modal never receives W&B credentials. The supervisor validates Modal results and appends their
+  metrics to the same W&B run.
+- SQLite and private-R2 JSONL metric segments are delivery and recovery transports, not competing
+  scientific metric stores. Verified terminal journals expire after seven days.
+- Public model R2 contains immutable checkpoint closures and a mutable no-cache run index. Private
+  eval R2 contains intents, results, and episode evidence. Private control R2 contains leases,
+  journals, promotions, and terminal receipts.
+- W&B config contains run-defining dimensions: `metrics_schema_version: 10`, `training_backend_id`,
+  `training_backend_config_hash`, `algorithm_id`, goal,
+  environment, starts, seed, frame skip, environment count, hyperparameters, eval protocol, and
+  runtime versions.
+- `goal_contract_sha256` is the semantic SHA-256 of the fully composed, rendered, validated goal
+  contract. Generated goal reports use it with `goal_slug` to keep current-contract leaderboards
+  comparable; runs without the current fingerprint remain historical evidence only.
+- Catalog-backed runs also record `effective_goal_contract_sha256`, `reward_program_kind`,
+  `reward_program_revision`, `reward_shape`, `reward_shape_sha256`, and
+  `reward_shape_is_default`. Reward-derived returns are comparable only when the selected reward
+  semantic identity and effective goal contract match; the readable key alone is not sufficient.
+- `leader/checkpoint/*` contains diagnostic projections of the selected checkpoint. The
+  create-only private-R2 `PromotionReceipt` is the authoritative selection.
+- Heavy model bytes, videos, replays, episode rows, diagnostics, and recovery payloads never go to
+  W&B.
+- Interactive playback uses local descriptor keys such as `reward/shaped`, `policy/value`, and
+  `action/executed` to configure live panels. They are typed projections of one streamed transition
+  or its bounded in-browser history, are not emitted metrics, and must not be interpreted as aliases
+  for similarly named W&B registry entries. `reward/provider` is the provider output before the
+  task program and gradlab-owned reward transform; `reward/shaped` is the final policy-facing reward
+  after task shaping, scaling, and clipping. Playback computes realized return-to-go and `V(s) -
+  G(s)` only for a terminal, stochastic, policy-driven trajectory whose policy environment,
+  reward stream, discount, action sampling, and boundary/bootstrap semantics match training.
+  Otherwise those critic diagnostics are explicitly unavailable rather than numerically compared.
+
+The active checkpoint protocol is `acceptance`; complete accepted evidence additionally emits the
+`full` metric family. Dimension IDs must be unique and match `[A-Za-z0-9_.-]+`; unsafe IDs are
+rejected rather than silently rewritten. Starts use the same readable ID in training and evaluation.
+Provider `info` fields never become metrics automatically.
+
+Configuration-selected internal learner feedback, such as an archive curriculum's per-start
+priority statistic, is not telemetry merely because it has a readable name. Internal feedback
+identifiers do not use metric paths and are not published to W&B unless a separately registered
+metric explicitly projects them. If projected, the emitted name and semantics must appear in the
+registry below. For SB3 PPO and A2C, archive-curriculum `priority_metric: value_error` specifically
+means the arithmetic mean of `abs(A_t)` over one completed archive-origin trajectory, where `A_t`
+is raw GAE before PPO minibatch normalization. That scalar updates the archive's cell-level EMA and
+is intentionally not emitted; `train/curriculum/archive/feedback/trajectory/count` reports only how
+many such trajectory updates were committed.
+
+An episode metric is a **return**. `reward` is reserved for per-step shaping and component
+attribution. Frame skip remains run config. W&B uses three explicit axes:
+
+- `train/global_step`: policy environment transitions consumed by training.
+- `eval/checkpoint_step`: step of the checkpoint represented by an evaluation row.
+- `orchestration/event_seq`: durable supervisor delivery order.
+
+Each axis is configured with a W&B `max` summary reducer. W&B's public API may therefore expose
+its summary value as a reducer mapping such as `{"max": 5046272}` rather than as a bare number.
+Catalog and report consumers must unwrap the configured reducer value; a recipe's requested
+`timesteps` cap is not a substitute for the observed `train/global_step`.
+
+Asynchronous evaluations may arrive after later training rows without changing their scientific
+X-axis. Each producer writes only its applicable scientific axis; durable delivery order uses
+`orchestration/event_seq`.
+
+Purged legacy W&B and R2 state has no compatibility guarantee. Newly materialized runs declare
+schema v10.
+
+## Research interpretation
+
+- Mario ranks checkpoints only after acceptance: earliest `leader/checkpoint/step`, then highest
+  `eval/full/episode/return/mean`. Breakout is training-only and ranks current-contract seeded recipe
+  cohorts using `train/episode/return/shaped/from/target/mean`, which excludes archive-curriculum
+  origins and non-episode control boundaries; tied cohorts prefer fewer policy transitions.
+- Aggregate training `current/rate/*` is cumulative. Aggregate `window_100/rate/*` is the latest
+  100 attempts. Global
+  window-100 min/mean appear only after every configured start has 100 attempts. Always pair early
+  current aggregates with `start_coverage/rate`.
+- A bounded training-only search may use per-start success counts and the history peak and first
+  threshold crossing of `train/outcome/success/window_100/rate/min` to screen and rank recipes.
+  That evidence is not checkpoint evaluation and cannot establish checkpoint promotion, goal
+  acceptance, or release evidence.
+- Failure reasons may overlap, so reason counts and rates need not sum to the terminal count.
+  Successful episodes contribute to success metrics, not the failure-reason families.
+- Positive PPO policy entropy, dominant-action rate, and the action histogram diagnose discrete
+  policy collapse. Value prediction and advantage histograms are sampled every 64 rollouts.
+- For reward-transform ablations, first compare `train/reward/raw/*` with
+  `train/reward/shaped/*`. `task.reward.reward_scale` is a positive divisor, so a value below one
+  amplifies the policy-facing reward. If raw rewards match but shaped magnitudes diverge, inspect
+  value loss and explained variance before policy entropy, dominant-action rate, KL, and clip
+  fraction: squared-error value loss can grow roughly with the square of the target scale, while
+  advantage normalization does not protect the critic from poorly conditioned targets.
+- Do not compare shaped episode-return or value magnitudes as policy quality across different reward
+  transforms. Use task success and acceptance-evaluation metrics for the outcome comparison; use
+  reward, critic, and policy metrics to locate the causal chain.
+- Snapshot-curriculum `sampling/probability/max` and `sampling/effective_cell/count` summarize the
+  current cell-probability distribution. They do not report realized per-cell selection frequency
+  or identify which resident cells were selected.
+- Derived throughput phase timing satisfies `loop wall time = env_step_seconds +
+  rollout_overhead_seconds + between_rollouts_seconds`. Compare those three phase durations on
+  matching workloads to identify a training-loop bottleneck. `rollout_overhead_seconds` includes
+  policy inference plus wrapper, buffer, reset, task, and callback work outside the native provider.
+  `between_rollouts_seconds` includes optimizer updates, callbacks, and logging, so it is deliberately
+  not named optimization time.
+- Reward components are emitted only when active. Each component has mean, nonzero rate, and share;
+  raw reward appears only when it differs from shaped reward. Mario's `progress` component includes
+  both its base new-progress reward and any configured additional new-progress reward above
+  `progress_reward_boost_start_x`.
+- Episode-return means are neither a best-episode metric nor the score of a currently visible lane:
+  they reduce the latest 100 completed episodes across all applicable vector lanes. W&B chart
+  smoothing, when enabled, is applied on top of that already-rolling value. Under the root Breakout
+  contract (`reward_mode: native`, unclipped), shaped episode return is the sum of Atari row-score
+  deltas, so individual 400-plus games can coexist with a much lower mean when other lanes finish
+  with lower scores.
+- Episode returns, success rates, failure reasons, policy entropy, and optimizer diagnostics
+  describe performance or mechanism; no one of them should be treated as a generic stall-stop
+  signal. A configured plateau condition may watch any registered numeric training metric, with
+  direction, minimum meaningful improvement, warmup, and patience owned by the goal or recipe.
+  `train/early_stop/{condition}/*` projects that condition's local state for diagnosis and shadow
+  calibration. It means only that the selected metric has not improved under the declared
+  condition, not that the task is impossible or that a checkpoint is accepted. Private control-R2
+  receipts, never W&B diagnostics, are authoritative for an active early-stop outcome.
+
+## Full-evaluation table
+
+`eval/full/by_start` has one row per start and observed failure reason with these columns:
+
+`checkpoint_step`, `start_id`, `episodes`, `success_count`, `success_rate`, `return_mean`,
+`return_std`, `return_median`, `reason`, `reason_count`, `reason_rate`.
+The reason value is empty, with zero count and rate, when a start has no recorded failure reason.
+
+Episode-level evidence stays in R2. Confidence intervals and start-by-reason scalar products
+are intentionally computed offline rather than added to W&B history.
+
+An acceptance rejection is complete evidence of failure, but not a complete 100-episode
+evaluation. W&B history always receives `eval/checkpoint_step`, pass, planned/completed episodes, and
+acceptance duration. It receives no partial `eval/full/*` result. Accepted projections additionally
+include variable return, length, progress, episode count, artifact, source, and `eval/full/by_start`.
+Constant acceptance success rates, per-start success scalars, failure-reason scalars, duplicate full
+duration, and constant leader-success fields remain in private-R2 evidence but are suppressed from
+acceptance W&B history.
+
+`eval/acceptance/pass` is per-checkpoint history. W&B summarizes that history with `max`, so the
+summary means that some checkpoint passed; it is not the run verdict. The authoritative verdict is
+the create-only private-R2 `PromotionReceipt`, whose selected result is hash-bound to the complete
+acceptance evidence. At terminal publication, that receipt restamps `gradlab/goal/outcome`, the diagnostic
+`leader/checkpoint/*` fields, and the accepted W&B projection. Later rejected checkpoint projections
+remain in history and never modify the active projection. Raw acceptance aggregates and episode
+evidence remain authoritative in private eval R2.
+
+## Delivery, backpressure, and recovery
+
+Every event has a stable content-derived event ID. Delivery to W&B is at least once; the durable
+`orchestration/event_seq` is also W&B's internal step, so replay after an interrupted local
+acknowledgement cannot append a second scientific point. Reports and summary projection retain the
+event ID for explicit deduplication. Promotion, terminal state, and early-stop authority are exactly
+once through conditional private-R2 receipts.
+
+The supervisor seals immutable metric-journal segments to private R2 every five seconds or 1,000
+events and batches pending frames to W&B. A retry reconstructs its local SQLite state from those
+segments before producing new events. It resumes the same W&B run with `resume="must"`.
+
+Backpressure is sampled every 15 seconds. The supervisor reports queue depth, oldest unpublished
+age, ingress and publish rates, observed publication-capacity ratio, local/R2/W&B high-water marks,
+remote-visible W&B lag, checkpoint backlog, pending evals, scratch utilization, accepted-result to
+stop latency, and post-learner idle-GPU tail. Publication capacity is healthy only when measured
+publish capacity is at least twice peak ingress.
+
+Unpublished W&B age warns at 45 seconds and is unhealthy at 60 seconds. Terminal W&B drain has a
+300-second deadline. If neither W&B nor private R2 can preserve pending metrics, or task scratch
+usage reaches 80%, the supervisor requests a safe learner stop and emits a resumable failure rather
+than discarding evidence.
+
+A logical run succeeds only when its private-R2 `TerminalReceipt` proves complete checkpoint and
+evaluation inventories, a promotion, the W&B high-water mark, and a complete drain. dstack process
+exit alone is never scientific success.
+
+## Registry
+
+<!-- METRIC_REGISTRY_START -->
+| Metric or template | Meaning | Unit | Cadence | Surface |
+|---|---|---|---|---|
+| `train/episode/return/shaped/mean` | Rolling mean shaped return over the latest 100 genuine completed training episodes across target and archive origins; a archive-origin return starts at restoration, and control boundaries are excluded. | scalar | rollout | history |
+| `train/episode/length/mean` | Rolling mean length over the latest 100 completed training episodes. | steps | rollout | history |
+| `train/outcome/terminal/count` | Cumulative terminal episode records. | episodes | rollout | history |
+| `train/outcome/reason/{reason}/count` | Cumulative failed episodes containing a reason. | episodes | rollout | history |
+| `train/outcome/reason/{reason}/rate/window_100` | Failure-reason incidence over the latest 100 terminal episodes. | fraction | rollout | history |
+| `train/outcome/success/from/{start}/count` | Cumulative successful genuine target-origin episodes from a start; archive-origin episodes are excluded. | episodes | rollout | history |
+| `train/outcome/success/from/{start}/attempts` | Cumulative genuine target-origin episode attempts from a start; archive-origin episodes are excluded. | episodes | rollout | history |
+| `train/outcome/success/from/{start}/rate/window_100` | Success rate over the latest 100 genuine target-origin attempts from a start. | fraction | rollout | history |
+| `train/outcome/success/current/rate/min` | Minimum cumulative target-origin success rate across observed starts. | fraction | rollout | history |
+| `train/outcome/success/current/rate/mean` | Mean cumulative target-origin success rate across observed starts. | fraction | rollout | history |
+| `train/outcome/success/window_100/rate/min` | Minimum target-origin window-100 success rate after every start has 100 genuine target-origin attempts. | fraction | rollout | history |
+| `train/outcome/success/window_100/rate/mean` | Mean target-origin window-100 success rate after every start has 100 genuine target-origin attempts. | fraction | rollout | history |
+| `train/outcome/success/start_coverage/rate` | Configured starts with a genuine target-origin attempt divided by configured starts. | fraction | rollout | history |
+| `train/early_stop/{condition}/value` | Current finite value consumed by a configured metric early-stop condition. | scalar | watched metric sample | history |
+| `train/early_stop/{condition}/best` | Best value retained by a configured metric early-stop condition. | scalar | watched metric sample | history |
+| `train/early_stop/{condition}/patience/elapsed_steps` | Policy steps elapsed in the condition's current patience interval. | steps | watched metric sample | history |
+| `train/early_stop/{condition}/patience/progress` | Condition patience progress capped at one; one means the condition would trigger. | fraction | watched metric sample | history |
+| `train/early_stop/{condition}/would_trigger` | Whether the configured condition would trigger at this sample, regardless of observe or stop action. | boolean | watched metric sample | history |
+| `train/reward/shaped/mean` | Distribution of learner-facing per-step reward after gradlab applies the task reward scale and then clipping. | scalar | rollout | history |
+| `train/reward/shaped/std` | Distribution of learner-facing per-step reward after gradlab applies the task reward scale and then clipping. | scalar | rollout | history |
+| `train/reward/shaped/min` | Distribution of learner-facing per-step reward after gradlab applies the task reward scale and then clipping. | scalar | rollout | history |
+| `train/reward/shaped/max` | Distribution of learner-facing per-step reward after gradlab applies the task reward scale and then clipping. | scalar | rollout | history |
+| `train/reward/shaped/nonzero_rate` | Distribution of learner-facing per-step reward after gradlab applies the task reward scale and then clipping. | scalar | rollout | history |
+| `train/reward/raw/mean` | Distribution of completed task reward immediately before gradlab-owned scaling and clipping, emitted when distinct from shaped reward. For an identity task this is the untransformed provider reward. | scalar | rollout | history |
+| `train/reward/raw/std` | Distribution of completed task reward immediately before gradlab-owned scaling and clipping, emitted when distinct from shaped reward. For an identity task this is the untransformed provider reward. | scalar | rollout | history |
+| `train/reward/component/{component}/mean` | Active reward-component attribution in pre-transform task-reward units. | scalar | rollout | history |
+| `train/reward/component/{component}/nonzero_rate` | Active reward-component attribution in pre-transform task-reward units. | scalar | rollout | history |
+| `train/reward/component/{component}/share` | Absolute contribution share computed from components in pre-transform task-reward units. | scalar | rollout | history |
+| `train/reward/signal/{signal}/mean` | Configured reward-source signal. | scalar | rollout | history |
+| `train/reward/signal/{signal}/max` | Configured reward-source signal. | scalar | rollout | history |
+| `train/reward/signal/{signal}/nonzero_rate` | Configured reward-source signal. | scalar | rollout | history |
+| `train/algorithm/ppo/update/approx_kl` | Approximate KL divergence for the PPO update. | scalar | rollout | history |
+| `train/algorithm/ppo/update/clip_fraction` | Fraction of policy ratios clipped by PPO. | scalar | rollout | history |
+| `train/algorithm/jerk/retained/count` | Distinct action sequences retained by JERK search. | sequences | rollout | history |
+| `train/algorithm/jerk/best/return_mean` | Mean observed return of JERK's highest-ranked retained sequence. | return | rollout | history |
+| `train/algorithm/jerk/best/sequence_length` | Action length of JERK's highest-ranked retained sequence. | steps | rollout | history |
+| `train/algorithm/jerk/archive/selected_prefix_return_mean` | Cumulative mean retained-prefix return selected for JERK archive replay. | return | rollout | history |
+| `train/algorithm/jerk/exploit/probability` | Probability that JERK starts an episode by sampling a retained archive sequence. | fraction | rollout | history |
+| `train/algorithm/go_explore/archive/cell_count` | Semantic cells currently retained by Go-Explore. | cells | interval | history |
+| `train/algorithm/go_explore/archive/entry_count` | Immutable entries in the ephemeral working archive, including replacements not yet removed by periodic compaction. | entries | interval | history |
+| `train/algorithm/go_explore/archive/blob_count` | Distinct content-addressed provider-state blobs retained by the search. | blobs | interval | history |
+| `train/algorithm/go_explore/archive/blob_bytes` | Uncompressed bytes in distinct retained provider-state blobs. | bytes | interval | history |
+| `train/algorithm/go_explore/archive/selection_count` | Cumulative archived-cell selections for restoration. | selections | interval | history |
+| `train/algorithm/go_explore/archive/visit_count` | Cumulative semantic cell visits. | visits | interval | history |
+| `train/algorithm/go_explore/archive/update_count` | Cumulative replacements of an existing cell by a better trajectory. | updates | interval | history |
+| `train/algorithm/go_explore/archive/recent_new_cell_rate` | New semantic cells divided by visits in the bounded recent visit window. | fraction | interval | history |
+| `train/algorithm/go_explore/archive/recent_visit_window` | Visit count represented by the recent new-cell-rate window. | visits | interval | history |
+| `train/algorithm/go_explore/archive/visits_per_cell` | Cumulative visits divided by current cell count. | ratio | interval | history |
+| `train/algorithm/go_explore/success_guided/cell_count` | Cells on the current best-success lineage. | cells | interval | history |
+| `train/algorithm/go_explore/success_guided/selection_count` | Cumulative selections made by success-lineage guidance. | selections | interval | history |
+| `train/algorithm/go_explore/best/progress` | Greatest task progress reached by the best retained trajectory. | progress | interval | history |
+| `train/algorithm/go_explore/best/return` | Shaped return of the best retained trajectory. | return | interval | history |
+| `train/algorithm/go_explore/best/program_steps` | Environment steps in the best retained JERK program. | steps | interval | history |
+| `train/algorithm/go_explore/best/program_runs` | Canonical action runs in the best retained JERK program. | runs | interval | history |
+| `train/algorithm/go_explore/best/completed` | Whether the best retained trajectory completed the task. | boolean | interval | history |
+| `train/algorithm/go_explore/best/improvement_count` | Successful-trajectory improvements after the first completion. | improvements | interval | history |
+| `train/algorithm/{algorithm}/value/explained_variance` | Actor-critic value-function explained variance. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/update/policy_gradient_loss` | Actor-critic policy-gradient loss. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/update/value_loss` | Actor-critic value loss. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/update/learning_rate` | Current actor-critic learning rate. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/policy/entropy` | Positive actor-critic policy entropy. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/policy/distribution_std` | Continuous-action distribution standard deviation. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/policy/dominant_action_rate` | Fraction assigned to the most frequent sampled discrete action. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/policy/action_hist` | Sampled discrete-action histogram. | histogram | every 64 rollouts | history |
+| `train/algorithm/{algorithm}/rollout/value_prediction/mean` | Rollout value-prediction distribution diagnostic. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/rollout/value_prediction/std` | Rollout value-prediction distribution diagnostic. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/rollout/value_prediction/min` | Rollout value-prediction distribution diagnostic. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/rollout/value_prediction/max` | Rollout value-prediction distribution diagnostic. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/rollout/value_prediction/hist` | Rollout value-prediction histogram. | histogram | every 64 rollouts | history |
+| `train/algorithm/{algorithm}/rollout/advantage/mean` | Rollout advantage distribution diagnostic. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/rollout/advantage/std` | Rollout advantage distribution diagnostic. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/rollout/advantage/min` | Rollout advantage distribution diagnostic. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/rollout/advantage/max` | Rollout advantage distribution diagnostic. | scalar | rollout | history |
+| `train/algorithm/{algorithm}/rollout/advantage/hist` | Rollout advantage histogram. | histogram | every 64 rollouts | history |
+| `train/algorithm/{algorithm}/hyperparameter/entropy_coefficient` | Current scheduled entropy coefficient. | scalar | rollout | history |
+| `train/throughput/loop_fps` | Policy transitions divided by rollout-start-to-next-rollout-start wall time. | steps/second | rollout | history |
+| `train/throughput/rollout_fps` | Policy transitions divided by rollout-collection wall time. | steps/second | rollout | history |
+| `train/throughput/env_step_fps` | Policy transitions divided by native-provider step wall time accumulated during the rollout. | steps/second | rollout | history |
+| `train/throughput/rollout_seconds` | Wall time spent collecting one rollout. | seconds | rollout | history |
+| `train/throughput/env_step_seconds` | Native-provider step wall time accumulated while collecting one rollout. | seconds | rollout | history |
+| `train/throughput/rollout_overhead_seconds` | Rollout wall time outside native-provider step calls, including policy inference and wrapper, buffer, reset, task, and callback work. | seconds | rollout | history |
+| `train/throughput/between_rollouts_seconds` | Wall time after rollout collection and before the next rollout, including optimizer updates, callbacks, and logging. | seconds | rollout | history |
+| `train/artifact/save/seconds` | Local model save duration. | seconds | artifact | history |
+| `train/artifact/upload/seconds` | Public R2 checkpoint publication duration. | seconds | artifact | history |
+| `eval/{protocol}/episode/return/mean` | Evaluation episode-return distribution accumulated from learner-facing rewards after gradlab-owned scaling and clipping. | return | evaluation | history |
+| `eval/{protocol}/episode/return/std` | Evaluation episode-return distribution accumulated from learner-facing rewards after gradlab-owned scaling and clipping. | return | evaluation | history |
+| `eval/{protocol}/episode/return/median` | Evaluation episode-return distribution accumulated from learner-facing rewards after gradlab-owned scaling and clipping. | return | evaluation | history |
+| `eval/full/episode/return/best` | Best full-evaluation episode return accumulated after gradlab-owned scaling and clipping. | return | evaluation | history |
+| `eval/{protocol}/episode/length/mean` | Mean evaluation episode length. | steps | evaluation | history |
+| `eval/{protocol}/episode/count` | Evaluation episodes represented. | episodes | evaluation | history |
+| `eval/{protocol}/outcome/success/from/{start}/rate` | Evaluation success rate from a start. | fraction | evaluation | history |
+| `eval/{protocol}/outcome/success/rate/min` | Aggregate per-start evaluation success rate. | fraction | evaluation | history |
+| `eval/{protocol}/outcome/success/rate/mean` | Aggregate per-start evaluation success rate. | fraction | evaluation | history |
+| `eval/{protocol}/outcome/reason/{reason}/rate` | Evaluation failure-reason incidence. | fraction | evaluation | history |
+| `eval/full/progress/{progress}/mean` | Goal-configured full-evaluation progress summary. | value | evaluation | history |
+| `eval/full/progress/{progress}/max` | Goal-configured full-evaluation progress summary. | value | evaluation | history |
+| `eval/{protocol}/checkpoint/artifact` | Evaluated checkpoint artifact reference. | metadata | evaluation | history |
+| `eval/{protocol}/duration/seconds` | Evaluation wall duration. | seconds | evaluation | history |
+| `eval/{protocol}/source` | Evaluation execution source. | text | evaluation | history |
+| `eval/acceptance/pass` | Per-checkpoint acceptance result; W&B summarizes its history with max, not as the verdict. | boolean | acceptance evaluation | history |
+| `eval/acceptance/episodes/planned` | Exact episode identities required by the acceptance manifest. | episodes | acceptance evaluation | history |
+| `eval/acceptance/episodes/completed` | Valid planned episode rows completed before acceptance or fail-fast rejection. | episodes | acceptance evaluation | history |
+| `eval/acceptance/failure/count` | Failed planned episodes; zero for acceptance and one for fail-fast rejection. | episodes | acceptance evaluation | history |
+| `eval/acceptance/duration/seconds` | Acceptance-worker evaluation wall duration. | seconds | acceptance evaluation | history |
+| `eval/full/by_start` | Structured full-evaluation evidence by start and reason. | table | evaluation | history |
+| `leader/checkpoint/acceptance_pass` | Diagnostic projection of the promoted-checkpoint verdict; `eval_scope_exact` is authoritative. | boolean | selection | summary |
+| `leader/checkpoint/success_rate_min` | Diagnostic selected-checkpoint projection. | summary | selection | summary |
+| `leader/checkpoint/success_rate_mean` | Diagnostic selected-checkpoint projection. | summary | selection | summary |
+| `leader/checkpoint/objective` | Diagnostic selected-checkpoint projection; never a ranking input. | summary | selection | summary |
+| `leader/checkpoint/return_mean` | Diagnostic selected-checkpoint projection. | summary | selection | summary |
+| `leader/checkpoint/best_return` | Diagnostic selected-checkpoint projection. | summary | selection | summary |
+| `leader/checkpoint/rank_values` | Diagnostic rendering of authoritative rank values. | summary | selection | summary |
+| `leader/checkpoint/progress_max` | Diagnostic selected-checkpoint projection. | summary | selection | summary |
+| `leader/checkpoint/step` | Diagnostic selected-checkpoint projection. | summary | selection | summary |
+| `leader/checkpoint/artifact_ref` | Diagnostic immutable artifact reference projection. | summary | selection | summary |
+| `leader/checkpoint/eval_source` | Diagnostic evaluation-source projection. | summary | selection | summary |
+| `leader/checkpoint/updated_at` | Diagnostic projection update time. | summary | selection | summary |
+| `train/global_step` | Scientific training X-axis: policy environment transitions consumed. | steps | frame | history |
+| `eval/checkpoint_step` | Scientific evaluation X-axis: step of the evaluated checkpoint. | steps | evaluation | history |
+| `orchestration/event_seq` | Monotonic local outbox event sequence used as W&B delivery order. | events | frame | history |
+| `orchestration/event_id` | Stable content-derived identifier used to deduplicate at-least-once delivery. | metadata | frame | history |
+| `orchestration/outbox/queue_depth` | Metric outbox frames not yet acknowledged by the W&B SDK. | events | supervisor sample | history |
+| `orchestration/outbox/oldest_unpublished_seconds` | Age of the oldest metric frame not yet acknowledged by the W&B SDK. | seconds | supervisor sample | history |
+| `orchestration/outbox/ingress_rate` | Observed local metric-frame creation rate over the latest supervisor interval. | events/second | supervisor sample | history |
+| `orchestration/outbox/publish_rate` | Observed W&B SDK acknowledgment rate over the latest supervisor interval. | events/second | supervisor sample | history |
+| `orchestration/outbox/publication_capacity_ratio` | Observed W&B publication rate divided by observed metric ingress rate. | ratio | supervisor sample | history |
+| `orchestration/outbox/local_high_water` | Largest metric-frame sequence durably present in local SQLite. | events | supervisor sample | history |
+| `orchestration/outbox/r2_high_water` | Largest metric-frame sequence sealed in immutable private R2 journals. | events | supervisor sample | history |
+| `orchestration/outbox/wandb_high_water` | Largest metric-frame sequence acknowledged by the W&B SDK. | events | supervisor sample | history |
+| `orchestration/outbox/wandb_remote_high_water` | Largest orchestration event sequence observed through the W&B API. | events | remote visibility probe | history |
+| `orchestration/outbox/wandb_remote_visible_lag_seconds` | Age of the newest local metric event not yet observed through the W&B API. | seconds | remote visibility probe | history |
+| `orchestration/checkpoint/backlog` | Ready local checkpoints not yet verified in public model R2. | checkpoints | supervisor sample | history |
+| `orchestration/eval/pending` | Persisted evaluation intents without a terminal verified result. | evaluations | supervisor sample | history |
+| `orchestration/eval/result_to_stop_seconds` | Time from observing an accepted eval result to signaling the learner. | seconds | accepted evaluation | history |
+| `orchestration/drain/idle_gpu_tail_seconds` | Time the training container retained its GPU after the learner exited. | seconds | terminal drain | history |
+| `orchestration/scratch/used_fraction` | Fraction of the task scratch filesystem currently used. | fraction | supervisor sample | history |
+| `train/episode/return/shaped/from/target/mean` | Rolling mean shaped return over the latest 100 genuine target-origin training episodes. | return | rollout | history |
+| `train/curriculum/archive/cell/count` | Current archive-curriculum cell count. | cells | rollout | history |
+| `train/curriculum/archive/entry/count` | Current immutable entry count retained by the curriculum view. | entries | rollout | history |
+| `train/curriculum/archive/admission/candidate/count` | Non-terminal cell-crossing candidates observed during the rollout. | transitions | rollout | history |
+| `train/curriculum/archive/admission/accepted/count` | Candidate entries accepted into cell reservoirs during the rollout. | entries | rollout | history |
+| `train/curriculum/archive/evicted/count` | Curriculum cells evicted during the rollout. | cells | rollout | history |
+| `train/curriculum/archive/capture/call/count` | Batched portable provider-state capture calls during the rollout. | calls | rollout | history |
+| `train/curriculum/archive/restore/episode/count` | Archive-origin episodes started during the rollout. | episodes | rollout | history |
+| `train/curriculum/archive/restore/forced_boundary/count` | Non-episode control truncations used to activate archive lanes. | boundaries | rollout | history |
+| `train/curriculum/archive/feedback/trajectory/count` | Completed archive-origin trajectories committed to the priority sampler. | trajectories | rollout | history |
+| `train/curriculum/archive/transition/share` | Fraction of policy transitions whose origin is the archive curriculum. | fraction | rollout | history |
+| `train/curriculum/archive/sampling/probability/max` | Largest final cell probability in the archive sampler. | fraction | rollout | history |
+| `train/curriculum/archive/sampling/effective_cell/count` | Inverse-Simpson effective cell count of the archive sampling distribution. | cells | rollout | history |
+| `train/curriculum/archive/capture/seconds` | Portable state capture wall time accumulated during the rollout. | seconds | rollout | history |
+| `train/curriculum/archive/restore/seconds` | Provider restore wall time for reset calls containing archive lanes. | seconds | rollout | history |
+<!-- METRIC_REGISTRY_END -->