[Tarzelf]

The learner needs a faithful bridge from demonstrated actions to the servo references the simulator actually applies. This adds matched absolute and reference-relative endpoint models using the same causal inputs, normalization, initialization, loss scales and original actuator projection. Neither uses teacher commands, native object/contact truth or future observations as inputs.

Both models trained on the same 96,000 training rows in 5.231 s and 5.098 s. Save/load preserved all 1,000 recorded-context outputs. A separate eight-attempt native development test completed but yielded 0/4 strict placements for each model. Absolute control reached one lift; relative control reached none and exceeded the strict TCP envelope in all four cases. Both unchanged weights, model/data cards, independent audit and prior infrastructure refusals are included. They remain unqualified research candidates; the frozen v0.1 result is unchanged.

Validation: 49 focused regressions passed; two real fits, recorded-context parity checks and all eight native episodes independently inspected. Runtime at measurement: 78ea4fbb4b2733e7738eb425eeecd7cb67a52195; subsequent commit adds evidence/docs only. The public API is Python-only; the new schema is not yet supported by the skill installer. Hardware control and ordinary human-video action recovery remain out of scope.

Stacked on #35. See docs/camera-action-reference-model-integration-v01.md and benchmarks/camera-action-reference-model-integration-v01/ for measurements, bounds and reproduction details.

Published artifacts: Hugging Face model repository, research report, and all eight recorded attempts. All 15 HF files were anonymously downloaded and matched at commit 8b2e983c; all 414 initial research files plus the two subsequently updated link documents matched their reviewed bytes. Current source head passed all 36 CI checks.

A clean outside-checkout install from the exact runtime Git tree matched all 227 installed package files and loaded both verified models. This was load-only: no fitting, inference or hardware. Fresh model-download requests failed DNS, so that smoke used previously anonymously verified local copies; remote Git installation itself was not tested.