TASK_STATUS.md

HoopSense Task Status

This file is a human-readable execution snapshot derived from docs/plan/PLAN_TREE.yaml.

The old chapter-style status model is retired. This file now tracks the current plan frontier and active L-level work only.

Current Frontier

The current highest-priority frontier is:

• L3.72 extend attributed MVP event emission to shot and rebound evidence once ball-result signals are available
• L3.66 add bounded multi-hypothesis identity infrastructure for ambiguous short-gap track continuity
• L3.64 add a minimal ball artifact to Layer 1 review outputs and use it to refine live-play gating
• L3.63 add explicit playback transport controls and conservative jersey-OCR display gating in the labeller
• L3.62 add a rule-based live-play/dead-ball gate with per-frame scores, stitched segments, and labeller review support
• L3.53 add bidirectional short-gap repair for clustered or briefly missed player tracks in Layer 1 artifacts
• L3.61 add a coarse torso-color appearance cue to further suppress seated spectators and sideline bystanders
• L3.46 check in canonical MVP stats header and formula templates with provenance rules
• L3.51 define a HoopSense-native declarative rules layer for stat attribution without reusing Shush implementation
• L3.47 map each MVP stat to required event primitives and possession fields
• L3.9 define PossessionContext fields and ledger serialization contract
• L3.10 track ballhandler_id, dribble_count, and pass_count for one possession slice
• L3.12 define the first stat-ready event set: pass, catch, dribble, shot_attempt, rebound, turnover
• L3.48 materialize auditable box-score rows that match the core scorebook columns and formulas
• L3.49 add scorebook-backed validation fixtures that compare generated stats to trusted totals
• L3.50 defer shot-profile and advanced report outputs until the core box score is reliable
• L3.37 decide the first migration target, if any, based on performance, determinism, and implementation risk
• L3.24 make Docker docs explicitly cloud-oriented and not the native Orin story
• L3.26 publish separate runbooks for cloud/x86 training and Jetson/ARM64 runtime validation
• L3.29 define functional-core CI ownership and path triggers
• L3.30 add a lightweight integration pipeline that depends on selected core workflows
• L3.31 gate training smoke/evaluation workflows behind ML-specific changes or manual triggers
• L3.32 define which GCP resources stay on direct gcloud versus move into Terraform
• L3.33 add initial Terraform layout for shared buckets, registry, and service accounts
• L3.20 add first slice-based evaluation outputs
• L3.27 define dataset version lineage and promotion rules across Oracle/synthetic training sets
• L3.28 define multi-arch or explicitly split cloud-vs-Orin container strategy

L0 Product Goal

• [>] Trustworthy basketball understanding from video

L1 Workstreams

• [>] Perception and geometry foundation
• [>] Action Brain and synthetic Oracle training loop
• [>] Possession context and game-state reasoning
• [>] Event attribution and stats generation
• [>] Deployment, runtime, and operator workflow
• [>] Collaboration, review, and project control
• [>] MLOps governance and model lifecycle control
• [>] DevOps reproducibility and delivery control

L2 Status By Workstream

Perception and Geometry Foundation

• Multi-source ingestion and deterministic video contracts
• Court geometry, homography, and camera state
• [>] Identity fusion, tracking, and temporal continuity
• [>] Pose estimation and kinematic lifting
• [>] Perception-and-geometry readiness gate
• [>] Python-versus-Rust boundary decision for perception and geometry

Action Brain and Synthetic Oracle

• Stable Action Brain feature contract (features_v2)
• [>] Oracle MoCap ingestion, FK, and dataset generation
• [>] Action Brain training, evaluation, and checkpoint lifecycle
• Sim-to-real fine-tuning and validation loop

Possession Context and Game-State Reasoning

• [>] Possession context contract in the ledger
• [>] Ball control, dribble count, and pass-chain tracking
• [>] Offense zone, transition, and drive semantics
• Referee-assisted rewind and correction logic

Event Attribution and Stats Generation

• [>] Event attribution rules combining motion, ball, and possession state
• [>] Stat ledger and box-score generation
• Shot chart, defensive metrics, and report outputs

Deployment, Runtime, and Operator Workflow

• [>] Native Orin environment and ARM64 runtime validation
• [>] Cloud/x86 training environment and artifact parity
• Real-time inference pipeline and model optimization

Collaboration, Review, and Project Control

• Gemini/Codex collaboration bridge and review workflow
• Plan-driven execution and task synchronization
• [>] Documentation and task-status alignment

MLOps Governance and Model Lifecycle Control

• [>] Dataset manifests, lineage, and validation policy
• [>] Training run lineage and checkpoint lifecycle
• Slice-based evaluation and drift monitoring
• Deployment compatibility reporting across cloud and edge targets

DevOps Reproducibility and Delivery Control

• [>] Guix-first development and environment reproducibility
• [>] Docker fallback packaging and cloud image discipline
• [>] CI quality gates for code, contracts, and docs
• Cloud and edge delivery guidance with explicit target boundaries
• Multi-pipeline CI architecture by functional core
• Terraform adoption for stable shared GCP infrastructure

Key L3 Execution Slices

Completed

• Preserve the Oracle MVP parser/FK path for one fixture motion
• Maintain a reproducible GPU training smoke test and checkpoint write path
• Require plan-tree updates before substantial architecture or implementation changes
• Define measurable readiness checks for ingestion, tracking, pose, geometry, and lifting
• Add a perception-quality report artifact for representative clips
• Materialize one real Layer 1 annotation artifact from GPU-backed Ultralytics inference on a representative 5-second clip
• Render real detection, track, and pose overlays in the labeller from a persisted annotation artifact
• Verify the Orin GPU container command used for representative Layer 1 artifact generation
• Add coarse light-versus-dark uniform bucket estimation to representative Layer 1 perception artifacts
• Add active-player scoring to suppress spectators and bench-side false positives in Layer 1 artifacts
• Add a rule-based live-play/dead-ball gate with per-frame scores, stitched segments, and labeller review support
• Refactor extract_game_dna into focused inference pipeline classes without changing output contract
• Add video discontinuity detection and continuity segments to Layer 1 artifacts before stronger identity repair
• Check in an explicit Layer 1 identity-policy contract and machine-readable rules file
• Check in machine-readable MVP event-evidence rules and a loader for deterministic stat attribution
• Make the first deterministic event-attribution path consume the checked-in MVP event rules
• Emit first attributed MVP event payloads in the shape required by the deterministic event-rule engine
• Add a minimal ball artifact to Layer 1 review outputs and use it to refine live-play gating
• Add bounded multi-hypothesis identity infrastructure for ambiguous short-gap track continuity
• Add explicit playback transport controls and conservative jersey-OCR display gating in the labeller
• Add bidirectional short-gap repair for clustered or briefly missed player tracks in Layer 1 artifacts
• Attach jersey-number evidence and consensus fields to persistent player identity without rewriting raw tracker IDs
• Add a coarse torso-color appearance cue to further suppress seated spectators and sideline bystanders
• Add Kalman-smoothed track state and motion-aware scoring to Layer 1 artifacts
• Support typed partial-court calibration landmarks and solver flow in the labeller without depending on visible corners
• Add an explicit general-note save path in the labeller so freeform perception review notes are not silently lost
• Replace named-point calibration clicks with primitive-based line and arc sampling in the labeller
• Normalize labeller feedback into a categorized regression fixture for Layer 1 tracking and postprocess work
• Add a regression-audit report over reviewed Layer 1 failure cases to guide tracking changes without clip-specific overfitting
• Add a structured perception-feedback workflow for false positives, misses, merges, and track errors in the labeller
• Verify actual CUDA acceleration in the Orin validation path and publish a repeatable probe artifact
• Publish a mobile-friendly external HoopSense demo page with one representative clip, overlay, and feedback prompt
• Publish a normalized MVP stats contract for a sellable scorebook-style output
• Scale Oracle ingestion to Subject 124 while preserving features_v2
• Define PossessionContext fields and ledger serialization contract
• Track ballhandler_id, dribble_count, and pass_count for one possession slice
• Define the first stat-ready event set: pass, catch, dribble, shot_attempt, rebound, turnover
• Generate MVP box-score rows from attributed events
• Check in canonical MVP stats header and formula templates with provenance rules
• Define a HoopSense-native declarative rules layer for stat attribution without reusing Shush implementation
• Map each MVP stat to required event primitives and possession fields
• Materialize auditable box-score rows that match the core scorebook columns and formulas
• Add scorebook-backed validation fixtures that compare generated stats to trusted totals
• Defer shot-profile and advanced report outputs until the core box score is reliable
• Implement offense zone and transition flag derivation in the possession context
• Document candidate Python-versus-Rust ownership split for the perception layer
• Implement first-pass dataset manifests with SHA-256 hashing
• Integrate training lineage recording (git commit + data hash) into training loop
• Establish initial CI plumbing (.github/workflows/ci.yml)
• Implement dynamic perception audit script (scripts/run_perception_audit.sh)

Active or Next

• Extend attributed MVP event emission to shot and rebound evidence once ball-result signals are available
• Add bounded multi-hypothesis identity infrastructure for ambiguous short-gap track continuity
• Add a minimal ball artifact to Layer 1 review outputs and use it to refine live-play gating
• Add explicit playback transport controls and conservative jersey-OCR display gating in the labeller
• Add a rule-based live-play/dead-ball gate with per-frame scores, stitched segments, and labeller review support
• Add bidirectional short-gap repair for clustered or briefly missed player tracks in Layer 1 artifacts
• Add a coarse torso-color appearance cue to further suppress seated spectators and sideline bystanders
• Check in canonical MVP stats header and formula templates with provenance rules
• Define a HoopSense-native declarative rules layer for stat attribution without reusing Shush implementation
• Map each MVP stat to required event primitives and possession fields
• Define PossessionContext fields and ledger serialization contract
• Track ballhandler_id, dribble_count, and pass_count for one possession slice
• Define the first stat-ready event set: pass, catch, dribble, shot_attempt, rebound, turnover
• Materialize auditable box-score rows that match the core scorebook columns and formulas
• Add scorebook-backed validation fixtures that compare generated stats to trusted totals
• Decide the first migration target based on performance and risk (TrackManager identified)
• Separate ARM64/Jetson runtime guidance from cloud/x86 Docker guidance
• Add first slice reports for action class, camera/view, and pose quality

Product Reality Check

What is currently true:

• the Action Brain is a narrow local-motion classifier
• features_v2 remains the current frozen neural contract
• the layered feature architecture is documented
• the geometry layer has a shared module and readiness report artifact
• an Orin container logic probe exists
• basic manifest hashing, training lineage scaffolding, and initial CI plumbing now exist locally
• possession and stat primitives exist locally in Rust and Python, but need tighter end-to-end verification
• the MVP product contract is now explicitly captured as a normalized scorebook-style output spec

What is not yet true:

• Oracle scale-out has not yet been re-materialized and verified as an on-disk dataset artifact
• the possession context contract is not yet proven end to end in inference output
• the scorebook-style box-score contract is not yet materialized as headers, formulas, and fixture-backed validation
• dataset promotion rules are not yet defined
• the CI architecture is not yet split by functional core
• no Terraform layer exists yet for shared GCP infrastructure

Quality Gates

• Rust core logic verified for PossessionContext and box-score changes
• Python behavior-engine/perception gate tests verified
• Synthetic Oracle tests passing
• Tooling review tests passing
• Documentation updated for layered feature architecture
• Planning invariant checked into AGENTS.md
• [>] Orin hardware validation (CUDA/PyTorch)
• Scorebook-style stat contract materialized as checked-in templates and validation fixtures
• Oracle dataset manifest validation implemented
• Checkpoint lineage recording implemented
• Slice-based evaluation report implemented
• Geometry readiness report artifact implemented
• Initial CI workflow implemented
• Check in an explicit Layer 1 identity-policy contract and machine-readable rules file