Skip to content

README.md

Latest commit

 

History

History
65 lines (44 loc) · 3.08 KB

README.md

File metadata and controls

65 lines (44 loc) · 3.08 KB
Raw

Known camera height ground scale R0

This directory contains the frozen, dataset-independent geometry core for KNOWN_CAMERA_HEIGHT_GROUND_SCALE_R0.

Run its focused mechanics tests with the canonical project Python:

E:\codex-tools\bin\blindassist-python.cmd -m unittest discover `
  -s scripts/research/known_camera_height_ground_scale_r0 -p "test_*.py" -v

The tests are synthetic algebra and rejection checks only. They are not effect evidence. The fresh ARKit source route was locked separately and stopped at its pre-DA source gate. The later TartanGround run was explicitly authorized as consumed Development evidence and cannot support a fresh, held-out, production, or generalization claim.

Validate the machine-readable freeze against the implementation constants with:

E:\codex-tools\bin\blindassist-python.cmd `
  scripts/research/known_camera_height_ground_scale_r0/validate_protocol.py `
  --protocol docs/research/hftf/KNOWN_CAMERA_HEIGHT_GROUND_SCALE_R0_PROTOCOL_2026-08-04.json

The metadata-only fresh-roster planner excludes every parent in the predecessor Spatial Head roster and refuses outcome-based replacement. Its selected public-data height is explicitly an oracle mechanism input, not a wearable deployment receipt.

evaluate_consumed_tartanground.py is the immutable single-frame R0 evaluator. evaluate_consumed_temporal_r1.py reads the cached R0 predictions and applies the posthoc, nine-valid-scale causal median fixed in the R1 optimization receipt. R1 uses no future frame, but its window was selected on the same consumed outcomes and therefore remains diagnostic only.

evaluate_camera_conditioned_student_r0.py performs the frozen five-fold leave-one-parent-out ridge evaluation. evaluate_student_external_replication_r0.py fits that unchanged student once on the five Development parents and runs the separately frozen ten-parent student-unseen replication. Both are synthetic evidence runners; neither is an Android or safety implementation.

sealed_student.py is the no-fit loader for the final consumed-evidence ridge head. Its JSON receipt and cross-language golden vectors live under configs/hftf/. preflight_phone_session.py fail-closes P0/R2 phone sessions unless measured camera height, fixed mount/camera identity, monotonic hashed RGB frames, and an independently hashed reference manifest are all present.

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。

产物边界

运行产物必须位于 artifacts.local/，不提交数据集、模型、设备日志或大文件。