Each record's step names its page and its directory in `# paths:`, and the workflow's trigger lists the same four paths, so a pull request that changes a reading, a fixture, the corpus, the rule, a dictionary or either page starts the workflow, and its gate runs the two steps that hold them. Before, only a change to the step files themselves or the nightly run did. Changelog: none Co-Authored-By: Claude Opus 5.5 <[email protected]>
15 lines
1 KiB
Shell
15 lines
1 KiB
Shell
# step: dcvc drift · the record and the text quoting it
|
|
# order: 32
|
|
# paths: wai/research/dcvc-drift.md wai/research/dcvc-drift/**
|
|
#
|
|
# The dcvc-drift record (research/dcvc-drift.md), over its committed readings: a second
|
|
# implementation that shares no code with the tool that took them
|
|
# (research/dcvc-drift/verify.py, standard library only, run isolated) decodes each y stream
|
|
# with its own rANS decoder under the encoder's buckets and under each reading's, recounts
|
|
# every desync event, holds each latent to its symbols, each drift reading to its histogram
|
|
# and digests and every summary and verdict to the readings, and checks every figure the page
|
|
# quotes (each is followed by a `dcvc` marker naming its place). --selftest then changes copies
|
|
# of the record, a changed reading, a miscounted desync and an event erased with its bucket
|
|
# among them, and requires each to be refused. No ONNX runtime and no build: the tool that
|
|
# took the readings is not run here.
|
|
python3 -I research/dcvc-drift/verify.py --selftest research/dcvc-drift research/dcvc-drift.md
|