Repository navigation
feat(metric-benchmark): collect verified Helix results in one command - #174
Draft
mstaats-nvidia wants to merge 1 commit into
Draft
mstaats-nvidia wants to merge 1 commit into
mstaats-nvidia wants to merge 1 commit into
Conversation
This was referenced Nov 4, 2026
mstaats-nvidia
force-pushed
the
explore/metric-benchmark-helix-import
branch
from
November 6, 2026 09:46
d18bf25 to
86dbf41
Compare
mstaats-nvidia
force-pushed
the
explore/metric-benchmark-helix-import
branch
from
November 6, 2026 10:27
86dbf41 to
525d238
Compare
mstaats-nvidia
changed the title
mstaats-nvidia
force-pushed
the
explore/metric-benchmark-helix-import
branch
from
November 9, 2026 11:51
f1cb98d to
0ff8340
Compare
mstaats-nvidia
changed the title
Signed-off-by: mstaats <mstaats@nvidia.com>
mstaats-nvidia
force-pushed
the
explore/metric-benchmark-helix-import
branch
from
November 9, 2026 12:15
0ff8340 to
efd76bb
Compare
asutermo
reviewed
Nov 9, 2026
| with native execution errors preserve those errors and their available artifacts. | ||
| One failed task does not discard the other trials. | ||
|
|
||
| The hidden sibling `.OUTPUT.helix` directory retains collection state and native |
Collaborator
There was a problem hiding this comment.
should it be .benchmark.helix? or soemthing along those lines?
asutermo
reviewed
Nov 9, 2026
| ) | ||
| if output.exists(): | ||
| receipt = _object(resolve_evidence(output, "import.json")) | ||
| if receipt.get("importer_versions", {}).get("harbor") not in ("0.20.0", "0.24.0"): |
Collaborator
There was a problem hiding this comment.
might be useful to move these to a constants list? and allow > 0.20.0?
asutermo
approved these changes
Nov 9, 2026
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Completes the four-tool workflow in #185 with
helix_collect.py: download selected runs, recover missing signed sources, verify what ran, and write the existing benchmark results contract in one command. There is no separate public import step.Collection requires exactly one saved run per prepared agent and preserves original task/agent identities, native checksums, job specifications, traces, statuses and unavailable observations. Source recovery uses the shared operation from #173, so the upload directory is unnecessary. Pending runs return exit 2 without publishing final output; completed evidence is validated and published atomically. Repeat collection verifies and reuses identical evidence.
Stack: #185 → #173 → #162 → #174. Standard validation and scoring consume the resulting
execution.jsonand per-case results.Validation: full CI passed on Python 3.12/3.13 (4,519 tests per version, 112 optional/environment skips); 149 focused tests passed with 3 skips, and 85 checks passed in the qualified client environment with 1 skip. Lint, formatting, type checks, copyright and license checks passed. An earlier external license-download checksum failure passed on retry with its integrity check unchanged.
Live qualification: the corrected six-trial registered-agent matrix collected, passed standard validation/scoring, and recollected with all 212 evidence files byte-identical. All six scores were correctly unavailable due to native recursion-limit, agent-timeout or sandbox-start failures. The first Oracle/NOP control matrix also validated and replayed unchanged, preserving its sandbox-start failures. These checks establish the evidence/format/failure-handling path, not successful model performance. A serial Oracle retry reproduced startup failure before agent execution. The final-head one-task alternative-model probe also collected, validated and replayed unchanged (86 files); both trials had native Fabric/API HTTP 424 errors. No clean three-task live performance run has been established.