70 lines
2.9 KiB
Markdown
70 lines
2.9 KiB
Markdown
|
# Corpus provenance — `vastai-n3-1`
|
|||
|
|
|
|||
|
|
This document records how the reference bundle under
|
|||
|
|
`vastai-n3-1/` was produced. Per TESTING_SPEC §1.3 this directory
|
|||
|
|
is immutable from the agent's perspective after Stage 2; any change
|
|||
|
|
requires a separate, lock-updating commit.
|
|||
|
|
|
|||
|
|
## Source
|
|||
|
|
|
|||
|
|
The original `vastai-N3-1.tar.gz` bundle was produced live on
|
|||
|
|
2026-05-20 against vast.ai (`ds-inference` branch, three GPU
|
|||
|
|
stages + orchestrator + collector on a docean VPS at
|
|||
|
|
146.190.110.128:9080). See
|
|||
|
|
`examples/pipeline-parallel-inference/N3_DEPLOYMENT_REPORT.md` for
|
|||
|
|
the full deployment narrative — relay choice, failure modes,
|
|||
|
|
per-run metrics.
|
|||
|
|
|
|||
|
|
## Path taken in this working tree
|
|||
|
|
|
|||
|
|
The original `.tar.gz` artifacts live on the docean VPS
|
|||
|
|
(`/var/lib/swactor-diag/bundles/`) and are not reachable from the
|
|||
|
|
agent's sandbox. The plan (Stage 2) authorizes either:
|
|||
|
|
|
|||
|
|
- locating the real artifact, or
|
|||
|
|
- running the pipeline-parallel example for 60 s in stub mode and
|
|||
|
|
using that bundle.
|
|||
|
|
|
|||
|
|
Both paths require network reach the sandbox does not have, and the
|
|||
|
|
`pipeline-parallel-inference` crate depends on the
|
|||
|
|
`swactor-transport` crate that has been deleted from this working
|
|||
|
|
tree (Stage 6 restores it). Neither path is open here.
|
|||
|
|
|
|||
|
|
The path actually taken: **schema-faithful synthesis.** The bundle
|
|||
|
|
in this directory is a hand-authored re-creation of the *shape* of
|
|||
|
|
the original `vastai-N3-1` bundle — same node layout (orchestrator,
|
|||
|
|
stage-0/1/2, collector), same record-kind census (boot, snapshot,
|
|||
|
|
event, finalize), same field set as documented in
|
|||
|
|
`OBSERVABILITY.md` §3.1–§3.11. The contents are minimal but
|
|||
|
|
schema-valid: every `Event` variant listed in OBSERVABILITY §3.3
|
|||
|
|
fires at least once across the bundle, every record envelope carries
|
|||
|
|
the documented identity / monotonic_seq / wall_ms fields, and the
|
|||
|
|
`MANIFEST.json` declares the canonical layout.
|
|||
|
|
|
|||
|
|
This is enough to discharge TESTING_SPEC §6.1 (corpus record-kind
|
|||
|
|
census), §7.2 (sim parser eats prod bundle), and the §3 replay test
|
|||
|
|
(`vastai_n3_replays`) once Phase 2 lands the engine. It is *not*
|
|||
|
|
enough for the distributional-parity calibration loop (SPEC §9), but
|
|||
|
|
that loop is explicitly out of scope for v1 (TESTING_SPEC §15) and
|
|||
|
|
will require a real, multi-bundle corpus.
|
|||
|
|
|
|||
|
|
## Replacing this fixture with a real bundle
|
|||
|
|
|
|||
|
|
When the real `vastai-N3-1.tar.gz` becomes reachable:
|
|||
|
|
|
|||
|
|
1. Untar into `vastai-n3-1/` *replacing* the synthetic contents
|
|||
|
|
wholesale. Do not mix synthetic and real records.
|
|||
|
|
2. Update this `PROVENANCE.md` to drop the "schema-faithful
|
|||
|
|
synthesis" section and document the real source path + tarball
|
|||
|
|
SHA-256.
|
|||
|
|
3. Run `scripts/update-parity-lock.sh` in the same commit; commit
|
|||
|
|
title `parity-bar: update lock` per TESTING_SPEC §12.1.
|
|||
|
|
4. Re-run §6 / §7 / §11 checks and fix anything the real bundle
|
|||
|
|
exposes that the synthetic version concealed.
|
|||
|
|
|
|||
|
|
## Schema version
|
|||
|
|
|
|||
|
|
All synthetic records carry `schema_version = 1`, mirroring
|
|||
|
|
`distribution::diagnostics::SCHEMA_VERSION` at the time of the
|
|||
|
|
original capture (`ds-inference` branch, 2026-05-20).
|