# Implementation Plan: Edge "Component Purity & Cannibalization Index" (Updated)

This feature delivers real-time, on-device detection of PCB cannibalization (stripping/depopulation of valuable components like CPU sockets, gold edge fingers, and tantalum capacitors) to calculate a **Board Integrity Score (BIS)**, generate vernacular voice warnings for scrap pricing adjustments, and propagate integrity data through the Dealer → Recycler chain to catch mid-chain stripping fraud.

Model training data collection is still in progress, so this plan is split into two tracks joined by a single frozen interface (`model_contract.md`): **Track A** builds and trains the detection model; **Track B** builds the full app pipeline against a mock model conforming to that same contract, so integration is not blocked on model readiness. Swapping the mock for the trained model later is a one-line config change, not a rewrite.

---

## User Review Required

> [!IMPORTANT]
> - **Labeling taxonomy refinement**: Split desoldered footprints into `forcibly_desoldered_pad` (jagged, torn traces, burnt flux) and `factory_unpopulated_pad` (pristine silkscreen, smooth solder dots) to eliminate false cannibalization alerts on unpopulated board regions. Added `occluded_region` to prevent penalizing components covered by heatsinks/dust.
> - **Form-factor baselines**: Replaced static class-wide component counts with form-factor sub-profiles (`desktop_atx`, `desktop_micro_atx`, `server_1u`, `gpu_pcb`, `ram_dimm`) in `board_profiles.json` to prevent small intact boards from receiving low BIS scores.
> - **Dynamic price floor**: Replaced the rigid 35% hardcoded price floor with category-based scrap grade floors ($\text{Floor}_{\text{category}}$ = 15% for high-grade server/GPU boards up to 60% for low-grade power boards).
> - **Sequential edge execution**: Adopted sequential model loading (`dispose()` calls between 21-class category classifier and YOLOv8 micro-model) to prevent memory crashes on 2GB–3GB RAM Android phones.
> - **BIS-delta tolerance bands**: Updated recycler intake fraud threshold to $>12\%$ for advisory flags and $>25\%$ for critical alerts (up from 5%) to accommodate field-vs-warehouse camera & lighting variances.
> - **Offline voice fallback**: Standardized pre-recorded `.opus` audio snippets via Service Worker `CacheStorage` as a fallback when Web Speech API offline regional voice packs are missing.
> - **Data sourcing strategy**: Confirmed that existing open datasets can bootstrap the generic (`*_intact`) component classes, removing most of the annotation burden for those; the fraud-specific classes (`forcibly_desoldered_pad`, `occluded_region`) still require primarily field-staged photography, supplemented — not replaced — by mined web sources. See Phase 1, Component 2a. **Requires a decision on whether forum/marketplace-sourced supplemental images are approved for use** (see licensing note in Component 2a) before any scraping work begins.

---

## Track A: Micro-Component Detection Model

### Phase 1: Data Collection & Labeling (Updated Guidelines)

#### Component 1: Refined Labeling Taxonomy (`taxonomy/component_classes.yaml`)
- [NEW] Defines fine-grained class states to eliminate false positives:
  - `bga_cpu_socket_intact`, `bga_cpu_socket_missing`
  - `gold_edge_finger_intact`, `gold_edge_finger_stripped`
  - `tantalum_cap_intact`, `tantalum_cap_missing`
  - `ram_dimm_slot_intact`, `ram_dimm_slot_stripped`
  - `forcibly_desoldered_pad` (jagged solder, torn traces, blowtorch/burnt flux residue — primary fraud signal)
  - `factory_unpopulated_pad` (clean silkscreen outline, smooth solder dots — unpopulated from factory, NOT fraud)
  - `occluded_region` (heatsinks, metal shields, heavy dirt/dust covering component pads)

#### Component 2: Data Collection & Sourcing Protocol (`data_collection/`)
- [NEW] `capture_protocol.md` — multi-angle capture protocol (Front side top-down, Reverse side top-down, and macro gold edge connector shot) to handle 2-sided PCB inspection.
- [NEW] `board_sourcing_log.csv` — logs board type, source, form factor (`desktop_atx`, `micro_atx`, `server_4u`, etc.), and intentional stripping state (pristine, blowtorched, organically degraded, factory unpopulated).
- Sourcing requirements: Must harvest dirty, corroded, dusty, and factory-unpopulated boards alongside pristine and stripped examples to match field scrap yard conditions.

#### Component 2a: Data Sourcing Strategy — Open Datasets, Mined Web Sources, and Field Photography (`data_collection/sourcing_strategy.md`)

Research into publicly available data confirms a **three-tier sourcing strategy**, since scraping alone cannot cover the fraud-specific classes:

**Tier 1 — Open datasets, for bootstrapping `*_intact` classes (use directly, no scraping needed):**
- A Dataset Ninja "PCB Component Detection" set — 1,410 images, 11,119 bounding-box-labeled objects across 9 classes (capacitor, transformer, MOSFET, resistor, etc.), released CC0 and explicitly tagged for the waste-recycling use case.
- HuggingFace `PCB_COMPONENTS_LABELLED` — 17-class YOLO-format set (IC, capacitor, connector, crystal oscillator, diode, ferrite bead, fuse, inductor, LED, resistor, switch, transistor, pins, pads, test points).
- HuggingFace `pcb-object-detection-dataset` — 21-class YOLO set including heatsinks and connectors.
- Roboflow Universe community PCB datasets (component-level bounding boxes, CC BY 4.0).
- **Use**: pretrain the detection backbone and pad out the `*_intact` and generic-component classes; re-map class names to `taxonomy/component_classes.yaml`. This does not reduce the need for field data on fraud-specific classes below, but meaningfully cuts the annotation volume needed for the "board still has its parts" side of the taxonomy.

**Tier 2 — Mined web sources, as a supplemental seed set for `forcibly_desoldered_pad` only (not a primary source):**
- Gold-recovery/e-waste hobbyist forums (e.g. Gold Refining Forum, finishing.com) contain user-posted photos of boards before, during, and after chip/pin removal, and active community discussion distinguishing "unpopulated" from "depopulated" boards — the closest organic analog to the `forcibly_desoldered_pad` class available online.
- Hobbyist "gold recovery" / heat-gun chip-pulling YouTube videos are a higher-yield variant of the same source: frame-extraction from a single video can produce many more usable images than a forum thread.
- **Known limitations, by design not treated as a primary source**:
  - Low volume per source and inconsistent photo quality/angle — doesn't match the phone-camera, scrap-yard conditions defined in `capture_protocol.md`.
  - No existing bounding boxes — still requires full manual annotation, same as field photos.
  - Licensing: forum and video content is user-generated and not freely licensed for reuse; scraping is low-risk for an internal/hackathon prototype but is **not cleared for production training data** without explicit permission or a licensed dataset. Flagged in User Review Required above — treat as blocked until a decision is made.
- **Use**: a small supplemental set to diversify what "stripped" looks like beyond staged photos, mined and annotated in parallel with, not instead of, Tier 3.

**Tier 3 — Deliberately staged field photography, mandatory and primary for fraud classes (this is the existing Component 2 protocol above):**
- Only source that guarantees: domain match to real collector phone conditions, controlled multi-angle capture, and clean separation between `forcibly_desoldered_pad`, `factory_unpopulated_pad`, and `occluded_region` at capture time (know the ground truth because you staged it, rather than inferring it after the fact from a forum photo).
- Remains the dataset of record for Phase 4 evaluation thresholds; Tier 1/2 data is for pretraining and augmentation only, not held-out test data.

#### Component 3: Annotation & QA
- [NEW] Bounding-box annotation via CVAT or Label Studio, exported in COCO/YOLO format.
- [NEW] `qa_annotation_review.py` — inter-annotator agreement check (min. 2 annotators per image, flag disagreement >10% of boxes for adjudication).
- Target: minimum 300 images per class, with at least 30% deliberately-stripped/burnt negative examples, before first training run.

### Phase 2: Form-Factor Baseline Definitions

#### Component 4: Dynamic Board Profiles (`baselines/board_profiles.json`)
- [NEW] Defines expected component counts/area ratios per board class **and form factor**:
  - `desktop_atx`, `desktop_micro_atx`, `desktop_itx`
  - `server_1u_2u`, `server_blade_4u`
  - `gpu_pcb_high_end`, `gpu_pcb_budget`
  - `ram_dimm_desktop`, `ram_so_dimm_laptop`
- Sourced from recycling domain experts — consumed downstream by `purity_assay.py`.

### Phase 3: Model Training

#### Component 5: Training Pipeline (`training/train.py`)
- [NEW] YOLOv8n/s fine-tuning with multi-scale training (`imgsz=960` or `1280`) and custom anchor ratios for elongated `gold_finger` boxes ($>4:1$).
- Heavy augmentations for field realities: specular glare on gold fingers, lighting variance, dirt noise, and angle tilt.
- [NEW] `train_config.yaml` — Focal loss & elevated class weights for `forcibly_desoldered_pad` vs `factory_unpopulated_pad`.

#### Component 6: Auxiliary Desoldered Pad Classifier Head
- [NEW] Secondary classification head specifically trained to distinguish desoldered blowtorch pads from factory empty pads.

#### Component 7: Export & Quantization (`export/export_tflite.py`)
- [NEW] Export to TFLite INT8 quantized, target size <15MB, target inference <1s on low-end Android hardware.
- [NEW] `model_contract.md` — freezes input/output tensor schema.

### Phase 4: Evaluation

#### Component 8: Model Evaluation (`eval/evaluate.py`)
- [NEW] Per-class mAP, plus dedicated fraud recall on `forcibly_desoldered_pad` and false-positive rate on `factory_unpopulated_pad`.
- Thresholds: $\ge 90\%$ recall on intact classes, $\ge 85\%$ recall on desoldered pad classes, and $\le 5\%$ false positive rate on factory unpopulated pads.

---

## Track B: Integration into Kabadiwala Connect

### Component 9: Purity Engine & Memory-Conscious Orchestrator (`purity_engine/`)

- [NEW] `purity_engine/model_client.py` — wrapper supporting `MockModelClient` and `TFLiteModelClient` with explicit tensor memory management (`model.dispose()` between Stage 1 and Stage 2).
- [NEW] `purity_engine/purity_assay.py` — implements `calculate_board_integrity_score(detected_components, form_factor)`:
  - Dynamic price formula:
    $$\text{Adjusted Price} = \text{Base Price} \times \left( \text{Floor}_{\text{category}} + (1 - \text{Floor}_{\text{category}}) \times \frac{\text{BIS}}{100} \right)$$
  - Scrap Grade Floors: Server/GPU = 15%, Desktop = 30%, Low-grade Power = 60%.
- [NEW] `purity_engine/pcb_analyzer.py` — single entry point for Collector Lite & Recycler Portal.

### Component 10: Vernacular Voice Alerts with Audio Fallback (`purity_engine/vernacular_voice.py`)

- [NEW] Extends Collector Lite's 8-language Web Speech API audio system with Service Worker pre-recorded `.opus` audio fallbacks for offline devices lacking regional voice packs.
- Message templates parameterized by BIS band (Pristine $\ge 85\%$, Partial $50\text{--}84\%$, Severe $<50\%$).

### Component 11: Collector Lite Multi-Angle UI Integration (`mobile/kabadiwala-lite.html`)

- Extend Step 1 of Collector Lite: Stage 1 21-class classifier detects `PCB` $\rightarrow$ prompts guided Front/Back photo capture $\rightarrow$ triggers `pcb_analyzer`.
- Displays BIS score, missing-component breakdown, and transparent price adjustment in Step 3.
- Persists assay JSON into IndexedDB with the scrap slip.

### Component 12: QR Handover Payload Enrichment (`mobile/dealer.html`, backend)

- QR payload includes collector BIS score, form factor, and assay report hash.

### Component 13: Recycler Portal BIS-Delta View (`recycler-dashboard-frontend/`)

- Re-scan intake lot via `pcb_analyzer`.
- Tolerance Bands:
  - $< 12\%$ BIS Drop $\rightarrow$ Normal capture variance (Pass).
  - $12\text{--}25\%$ BIS Drop $\rightarrow$ Discrepancy Flagged.
  - $> 25\%$ BIS Drop $\rightarrow$ Critical Fraud Alert & CPCB Form 6 PDF Audit Log Entry.

### Component 14: Backend API Changes (`backend/`)

- [NEW] `POST /api/lots/{lot_id}/assay` & `GET /api/lots/{lot_id}/bis-delta`.

---

## Sequencing Summary

| | Track A (Model & Dataset) | Track B (App Integration) |
|---|---|---|
| Now | Tier 1: bootstrap from open datasets. Tier 3: harvest dirty/stripped/unpopulated boards; annotate fine-grained taxonomy. Tier 2 (mined web sources) blocked pending licensing decision | Build multi-angle UI & pricing engine against `MockModelClient` |
| Parallel | Multi-scale YOLOv8 training with focal loss & TFLite INT8 quantization | Collector Lite, audio fallbacks, Recycler Portal BIS-delta tolerance UI |
| Hand-off | Ship model clearing Phase 4 thresholds | Swap `TFLiteModelClient` in with zero code changes |