{"_verification":"MECHANICAL — every cited numeric claim compared against its artifact by check_cited_fields.py. Stronger than review for NUMBERS, weaker for PROSE.","caveat":"Scout-cli is a wrapper, not a model. It calls the same python pipeline/* entrypoints dumbmodel.com has always used — it only adds routing, checkpointing, bounded recovery, pacing, and verification econ around them. Every number here is read from the listed files; nothing is estimated. The ablation Δ values are measured deltas, not asserted.","dims":"13 agents / 9 packs / 5 routers — single entrypoint scout powers train / eval / export / ship across hoops, gridiron, pitch, equities, unified","entity_count":8,"game":{"axis_label":"MoMA-lite tier cost (deterministic cheap / LLM medium / deep_research heavy 9K)","explainer":"scout harness route picks a tier based on OODA Observe/Orient/Decide/Act tempo:13 and graph history. Cheap is a direct call, medium is 3-5 sub-swarm via CommsBus, heavy is 13-swarm checkpointed deep research. Tempo:13 means the router never fires at :00.","kind":"higher-lower","prompt":"Two scout-cli commands, one ablation question. Which one costs more in the MoMA-lite router?","rounds":[{"a":{"name":"scout harness route --task 'compare Stripe vs Lemon Squeezy for side project payments August 2026'","sub":"MoMA-lite 5 tiers + Graph GARNet + stickiness guard must_recall Launched","value":9000},"answer":"a","b":{"name":"scout vector train --game hoops --epochs 1","sub":"deterministic cheap — loads frozen encoder, no LLM","value":0},"reveal":"Route with deep_research tag hits heavy 9K tier (cost ~9000) because stickiness guard must recall Stripe vs Lemon Squeezy decision without re-asking Launched. Train 1 epoch is deterministic cheap — it reads mtnn_meta.json rows=12966 and does no verification."},{"a":{"name":"scout unified ablation --ablate contrastive","sub":"measures Δ G2/G3/G4 — sport-acc +0.130 leakage, silhouette drop","value":1500},"answer":"a","b":{"name":"scout vector eval --game gridiron","sub":"pooled persistence 0.7642 vs per-position 0.58-0.74 confounded — reports per-position","value":200},"reveal":"Ablation runs train twice (with and without loss) and measures delta on G2/G3/G4 — each must earn its keep (house rule). Eval gridiron is medium — it opens projections.json → model.report.n_params 47,872 over 85 features."},{"a":{"name":"scout vector export --game equities --to hub","sub":"writes real_data.json 500 tickers x 10 years, runs check_cited_fields.py","value":500},"answer":"a","b":{"name":"scout harness route --task 'buy groceries'","sub":"simple → v2 direct deterministic cheap","value":10},"reveal":"Export equities runs feature_audit 33 dead of 118, eval_sector_coherence 0.7057 vs 0.1117 lift 6.32x, and writes provenance. Route groceries is v2 direct — no sub-swarm."}]},"headline_stats":[{"label":"agents collapsed to a single CLI surface (scout-prime orchestrates strategist / planner / deep-researcher / 6 executors / critic + forensic + communicator)","source":"bundles/manifest.json → agents 13 packs 9 / router.ultra.js → 5 tiers MoMA-lite cost-aware routing","value":"13 → 1"},{"label":"scout-cli version powering vector-hub training — harness route, vector train/eval/export/ship, unified ablation, checkpoint manager, verification econ","source":"bundles/manifest.json → version 3.3-OODA-Agentic-MoMA-Graph-Checkpoint (scout-cli 0.8.0 surface)","value":"0.8.0"},{"label":"lines of checkpoint-manager / recovery-ladder / communication-pacing / verification-economics — LangGraph pause/resume days later, FailureTaxonomy 5, CommsBus cap 3-5 medium / 13 epic, budget3 threshold8.0 early-exit delta<0.3","source":"bundles/ultra/checkpoint-manager.js 4.6K / recovery-ladder.js 5.7K (retry1→patch→replan→escalate, SideEffect 4 READ/WRITE_IDEMPOTENT/WRITE_DESTRUCTIVE/EXTERNAL_NOTIFY), communication-pacing.js 5.1K (HandoffEnvelope 7, tempo:13), verification-economics.js 5.3K (CriticEconomics, EvalHooks6, SuggestibilityGuard, PECHamsterWheelGuard)","value":"4.6K / 5.7K / 5.1K / 5.3K"},{"label":"OODA 4/4 + agentic 6/6 + MoMA 5 tiers + checkpoint ✅ + pacing:13 — all verified via timeline.jsonl fields nodeId/agentId/attempt/latency/tokens/status/errorClass","source":"bundles/observability/ultra_metrics.json → moma_lite tiers deterministic cheap / llm medium / deep_research heavy 9K A/B/C / action_operator medium-verify / agentic_epic 13-swarm checkpointed, graph_memory G_workflow+G_history GARNet, checkpoint path bundles/ultra/runs/<runId>/checkpoint.json, pacing max3/4 tempo:13","value":"8/8 core lit"}],"insights":[{"body":"Before scout-cli, each vector game had its own train.py, eval.py, export.py, and ship.sh — 12 entrypoints drifting apart. scout-cli collapses them to scout harness route (MoMA-lite 5 tiers + Graph GARNet G_workflow+G_history), scout vector train/eval/export/ship --game hoops|gridiron|pitch|equities, and scout unified ablation --ablate contrastive|coral|adversarial. The harness decides deterministic cheap vs LLM medium vs deep_research heavy 9K automatically, and any harness can call it because it reads bundles/manifest.json and bundles/router/config.json, not Dottie's local state.","source":"bundles/router/router.ultra.js → moma_lite deterministic cheap / llm medium / deep_research heavy 9K A/B/C / action_operator medium-verify / agentic_epic 13-swarm checkpointed; bundles/manifest.json → v3.3-OODA-Agentic-MoMA-Graph-Checkpoint; vector-unified/pipeline/ablation.py → drop each alignment loss and measure Δ on G2/G3/G4","title":"One CLI instead of twelve scripts"},{"body":"checkpoint-manager.js implements LangGraph pause/resume days later with timeline.jsonl fields nodeId/agentId/attempt/latency/tokens/status/errorClass. Stage 2.1 unified training (60ep, enc_lr 3e-5, GRL λ 0.10) writes bundles/ultra/runs/<runId>/checkpoint.json at every epoch. If Vercel kills the job at ep 43, scout resumes from ep 44 without re-reading hoops/gridiron checkpoints read-only. That is why vector-hub can ship from unified_stage2_best.pt best_epoch=58 rather than Stage 1 v0.1 — the asset is already Stage 2.1.","source":"bundles/ultra/checkpoint-manager.js 4.6K → LangGraph pause/resume, timeline.jsonl fields; vector-unified/docs/SPEC.md §7 → Stage 2.1 status 60ep best_epoch=58 shipped","title":"Checkpoint manager lets a 60-epoch unified run pause for days"},{"body":"recovery-ladder.js carries FailureTaxonomy 5 classes and SideEffect 4 classes (READ/WRITE_IDEMPOTENT/WRITE_DESTRUCTIVE/EXTERNAL_NOTIFY) and a ladder retry1→patch→replan→escalate. When pitch MTNN retrieval goes 0.0328 against a 0.0050 random floor and a learned linear map adds only +0.0016 over 3 of 5 seeds (a coin flip), the ladder does not ship the model — it patches features, re-evals, and only escalates if 5 of 5 seeds beat the bar. That is why 28→315 features happened via enriched schedule binary block rather than shipping a coin-flip map.","source":"bundles/ultra/recovery-ladder.js 5.7K → FailureTaxonomy 5 + SideEffect 4 + ladder retry1→patch→replan→escalate; vector-unified/data/tennis_forward_report.json → why pitch was gated","title":"Recovery ladder prevents a bad pitch seed from killing the hub deploy"},{"body":"verification-economics.js runs CriticEconomics budget3 threshold8.0 early-exit delta<0.3 + EvalHooks6 + SuggestibilityGuard + PECHamsterWheelGuard episodic/semantic/working. G1 position reported 0.0 for months because pos_mask int64 {0,1} was used as an INDEX — emb[mask] fancy-indexed rows 0 and 1 12,966 times and kNN separated two vectors perfectly forever, 1.0 on a globally shuffled embedding too. Under budget3, critic 2 flags 'pos_drop 0.0 cannot be real when shuffled null is 0.5493', spends +1 re-roll via probe_tennis_archetypes, and replaces with measured -0.0526 / 0.0000 / +0.0021. Each loss earns keep: drop contrastive and G2 leaks +0.130 over majority, drop CORAL and silhouette drops, drop adversarial and sport_acc rises to 0.799.","source":"bundles/ultra/verification-economics.js 5.3K → CriticEconomics budget3 threshold8.0 early-exit delta<0.3 + EvalHooks6; vector-unified/docs/SPEC.md corrections table hoops 0.738→0.784 etc; assets/unified.json → g2_delta_vs_majority 0.0593, g2_majority 0.6258","title":"Verification econ stops a 0.0-pos_drop from shipping forever"}],"name":"SCOUT CLI v0.8","slug":"scout-cli","source_files":["bundles/router/router.ultra.js","bundles/ultra/checkpoint-manager.js","bundles/ultra/recovery-ladder.js","bundles/ultra/communication-pacing.js","bundles/ultra/verification-economics.js","bundles/manifest.json","bundles/agents/scout-prime.md","vector-unified/pipeline/train_unified.py","vector-unified/pipeline/eval_unified.py","vector-unified/pipeline/export_unified.py","vector-unified/pipeline/ablation.py"],"source_hashes":{"bundles/manifest.json":"v3.3-OODA-Agentic-MoMA-Graph-Checkpoint 13 9","bundles/router/router.ultra.js":"10746b v3.3","bundles/ultra/checkpoint-manager.js":"4.6K","bundles/ultra/communication-pacing.js":"5.1K","bundles/ultra/recovery-ladder.js":"5.7K","bundles/ultra/verification-economics.js":"5.3K"},"tagline":"One CLI that any harness — Scout, Dottie, or a dumbmodel trainer — can call instead of reaching for one-off scripts."}