RISKSEG-R0 training data preparation

Status

Development-only infrastructure for the frozen RISKSEG_R0 task and data contract. Full sequential execution authority is already recorded in that contract; this utility does not bypass its event-data, technical-preflight, or promotion gates.

Stable interface

materialize_reencoded_view.py reads the frozen data-role ledger and only the three source manifests declared by that ledger. It selects exactly the declared train/dev source sessions and frame counts, verifies source hashes, and writes a new four-class mask view. dev_200 and the old blind 120 use canonical_passthrough; the fresh 200 first use the SHA-bound frozen R2-P0 source_native canonicalizer. Only then is the old canonical order remapped:

legacy 0 -> RISKSEG-R0 0
legacy 1 -> RISKSEG-R0 2
legacy 2 -> RISKSEG-R0 1
legacy 3 -> RISKSEG-R0 3

Example:

python -m scripts.research.riskseg_r0_training.materialize_reencoded_view `
  --repo-root E:\linnan\linnan `
  --output-root artifacts.local\evidence\riskseg-r0\training-reencoded-view

The output directory is atomically published below artifacts.local/ and contains:

• manifest.jsonl: one row per selected frame with role, explicit dataset root and decoder, native source hash, in-memory old-canonical PNG hash, new mask path/hash, both mapping hashes, dimensions, and four-class pixel counts.
• receipt.json: total, role, session, and class coverage; source/ledger identity; class/session coverage gates.
• masks/{train,dev}/: re-encoded single-channel PNG masks.

Source RGB images are hash-verified and referenced, not copied.

Safety boundary and stop conditions

The materializer never modifies the ledger or source data. It fails closed and publishes no output for manifest/source hash drift, path traversal, unknown class IDs, duplicate selected IDs, exact session/frame-count drift, image/mask dimension mismatch, or insufficient obstacle/boundary coverage. In particular, both blocking_obstacle and boundary_level_change must have nonzero pixels in at least two train sessions and two dev sessions; failure is DATA_SPLIT_NOT_READY.

This module does not inspect event-eval review records, candidate outputs, or model outputs. The 90-frame fixed-regression role is not an input.

Research question and evidence ceiling

This utility answers only whether the already-consumed 520-frame train/dev view can be reproducibly remapped to the frozen four-class contract with closed source ancestry. Pixel coverage is a data-readiness check, not event-level model evidence, deployment evidence, or safety evidence. Previously consumed assets remain Development-only.

PIDNet-S formal training

train.py is the implementation-locked, single-architecture training path for the three frozen seeds. It consumes only the 320-frame train and 200-frame dev roles, uses the official PIDNet auxiliary/final/boundary training structure, and never reads event-eval model outcomes.

The recipe is defined as constants in train.py; formal runs accept only the three predeclared seeds and write checkpoints plus complete dev/session metrics under ignored artifacts.local/ evidence directories.

状态：development

稳定 Interface

公开入口、输入不变量和失败模式以本目录脚本帮助和专项协议为准；跨域调用不得依赖私有 Implementation。

输出

只写入 artifacts.local/ 下的明确证据目录；不写仓库根目录或正式 App 资产。

安全边界

本模块不产生默认 App、生产、安全或 unseen confirmation authority；结果按当前协议声明的 Development/diagnostic 角色使用。

停止条件

最小判别实验完成、输入权威缺失、预算耗尽或重复失败时停止当前 evidence version，并保持最小 failure scope。