scale_free_traversability_r0

状态：development / frozen operator / external replication evaluation

研究问题与版本

R0 tests whether one fixed scale-invariant, three-band relative-depth operator executes stably. R1 and R2 evaluate that unchanged operator against public RGB-D sensor truth. Authority is Development only and follows docs/RESEARCH_GOVERNANCE.md.

稳定 Interface

• core.py: frozen score and five-frame causal decision mechanics;
• evaluate_phone_consumed.py: consumed fixed-phone mechanics diagnostic;
• evaluate_bonn_rgbd_consumed.py: Bonn RGB-D R1 evaluator;
• evaluate_arkitscenes_rgbd_consumed.py: ARKitScenes RGB-D R2 evaluator;
• validate_bonn_rgbd_result.py: evaluator-independent ledger/result validator.

All evaluators require explicit input and new output roots and fail if an output root already exists. Checkpoint, protocol, roster, and/or source archive hashes are bound before inference. Run focused tests with:

E:\codex-tools\bin\blindassist-python.cmd -m unittest test_core.py test_bonn_rgbd_evaluation.py

输出

Machine evidence is written only to a caller-selected new directory under artifacts.local/evidence/hftf/. Durable protocol and result summaries live in docs/research/hftf/.

安全边界

Outputs mean only relative image-band agreement. They do not mean clear, safe, blocked, distance, collision risk, an alert, App integration, or production fitness. Consumed public data must keep its disclosed Development role.

停止条件

Do not change the R0 checkpoint, ROI, bands, valid-coverage rule, percentiles, causal window, margin, winner count, source cohort, truth reconstruction, or gates after reading the corresponding round's outputs. A failed source-support precondition remains NOT_EVALUABLE; a failed accuracy gate remains NOT_SUPPORTED.

失败资产复用

Failed or not-evaluable ledgers remain valid source-characterization, regression, validator, and counterexample fixtures. They may be reused with the consumed role disclosed, but cannot be repackaged as unseen Confirmation.