# PiPER contact history v0.3 — complete confirmation receipt

The `receipt/` directory is a **complete, byte-identical 90-file benchmark bundle**,
including the original protocol, report, HTML report, all 84 compressed attempts,
integrity manifest and original MIT model notices. Seed **813407** is the frozen
confirmation cohort. The sibling execution and publication records sit **outside
the original bundle seal**.

- [Complete paired comparison](receipt/report.html)
- [All outcomes, history decisions, actual motion and timing](receipt/report.json)
- [Frozen twelve-scene, eighty-four-attempt protocol](receipt/protocol.json)
- [Stored and decoded byte integrity](receipt/integrity.json)
- [Original MIT simulation-model license](receipt/licenses/agilex_piper/LICENSE)
- [Installed-command execution boundary](execution.json)
- [Pinned source and offline inspection guide](https://github.com/Dvidia-Inference/dvidia-training/blob/2a5fceaea8c2c7eeae808ea6656233d2a7f433fe/docs/piper-contact-memory-v03.md)
- [Full hosted original bundle](https://huggingface.co/spaces/Dvidia/robot-simulation-lab/tree/a6bf57fd64e61486dafaa93fd96168f953ccff80/tasks/piper-contact-v03)

**History completes 6/12 placements, versus 4/12 for both original and probe-only
motion; all 48 matched open-jaw and hold controls fail.** Every one of 84 attempts
remains numerically valid and retains 20 seconds, 1,000 commands and 1,001 states.
Two added completions occur at nominal friction with 60 g blocks. Four history
tasks resume and complete after five closed holds; only four of five resume
windows meet the independent low absolute TCP peak-speed criterion. All six
supported acquisition retries still fail. There are 75 missed 20 ms computation
windows; real-time operation and physical qualification are unestablished.

Seal-listed files total 55,578,406 stored bytes and 300,141,159 bytes when attempt
JSON is decoded. These are storage measurements, not a minimum RAM requirement.

The original controller, probe-only controller and history controller each have
their own exact requested-arm/open-jaw replay; a common initial-arm hold supplies
the seventh variant. Native model settings and starting state are identical
within each scene. Three friction settings and two masses share two XY layouts;
these related synthetic fixtures do not estimate population reliability.

For complete local inspection, preserve all paths inside `receipt/` and pass
that folder to the pinned `dvidia-piper-contact inspect` CLI. It verifies stored
gzip bytes, bounded decoded JSON, protocol/source identities, complete denominators,
paired control tapes and causal history/hold/reassessment consistency without
running native physics. The [separate exploratory development appendix](https://github.com/Dvidia-Inference/dvidia-training/tree/c8868c873f075116b71d908d464890e09778726b/benchmarks/piper-contact-development-20261011)
retains every scripted development attempt and its exact source snapshot; these
are not confirmation attempts and do not increase its denominator.

Synthetic rigid blocks and privileged state establish no learned model, camera
perception, calibrated material/contact model, shoe competence or physical robot
qualification. Requested reduced movement and hold targets do not imply instant
speed reduction or physical arrest; actual motion and per-attempt timing summaries/maxima/miss counts remain
visible. Research writing follows CC BY 4.0; DVIDIA tooling and preserved model
notices retain MIT terms. Seals establish recorded consistency, not publisher
authentication or independent physics recomputation.

[Verified Hugging Face publication](huggingface-publication.json) · [Publication provenance](provenance.json). Both records are outside the original receipt seal.