3.6 KiB
Release conformance evidence
check-conformance.py is a release gate for all 64 same-vendor cells. It
currently fails because the simulator and bench evidence producer is not yet
implemented. A green Rust build, host HAL replay probe, or rootfs archive does
not fill any cell.
research-evidence.json records scoped historical observations with source
hashes. evidence_status.py reports those separately from declarations in
conformance.json. Neither research claims nor manifest declarations count
as verified release cells. The current-run checker below is the only release
validator. experiment-queue.json ranks the next protocol and controller
experiments by the number of affected cells; blocked work stays visible.
Each CI run must put a JSON report at
ci/evidence/<GITHUB_RUN_ID>/<board>/<chip>.json. The report binds its
run_id, source_commit, os_binary_sha256, board, and chip to the
current workflow. It records stock_firmware_modified: false,
simulator_booted_stock: true, and a stock_result of mines or rejects.
For a rejection, stock_first_rejection names the first genuine firmware
gate. Reports for stock mining also reference stock_pool_accept_capture.
The report references files under ci/evidence/ with objects containing
path and sha256: sdk_image, stock_image, simulator_trace,
simulator_model, control_capture, chip_capture,
alkane_share_accept_capture, and bench_validation.
The checker opens and hashes each file. It loads the trace and capture JSON,
checks the current run/image/cell identities, verifies SHA-256d or Litecoin
Scrypt nonces against
a pool target at difficulty 1 or higher, and requires a matching
mining.submit and pool ACCEPT response ID. A SHA-256 capture must include
the matching subscription (extranonce1, extranonce2_size), full
notify message, exact decimal difficulty, and a configure_response
with an accepted BIP310 mask when version bits were submitted. A later
version_mask_event may supply the effective mining.set_version_mask.
The checker reconstructs
the coinbase, merkle root, and complete submitted header; its 80 bytes and
calculated difficulty target must match the capture. The trace binds
the capture digests.
The Scrypt verifier uses Litecoin's 80-byte, N=1024, r=1, p=1 proof hash.
Equihash proof verification remains pending, so its cells fail closed.
Scrypt cells also remain pending until real silicon reports, stock behavior,
and pool acceptance are captured for those chips.
Bench calibration names the physical control-board and chip devices, their
capture digests, and the simulator model digest. It can be reused across OS
commits until the model or a physical capture changes.
The fixed digests and artifact paths also appear in conformance.json so
source review can track what each cell claims.
The simulator runner must generate these reports by booting pristine stock firmware and Alkane OS in a machine model, with nonce data replayed from real-silicon captures. A rejected stock pairing must show its first rejection in the simulator trace while Alkane OS reaches pool ACCEPT. Bench devices provide independent observations for model validation, not CI dependencies. The checker verifies its input artifacts and reported nonce. It cannot prove that a self-asserted pool response came from a real pool or turn a report into physical evidence. The capture producer must preserve raw request/response bytes and independent provenance. Capture provenance and faithful machine models remain release blockers until the runner and bench validation are built.