[Draft - DO NOT MERGE] bad-PR validation harness (local-eval mode)#72
Closed
saurabhrb wants to merge 5 commits into
Closed
[Draft - DO NOT MERGE] bad-PR validation harness (local-eval mode)#72saurabhrb wants to merge 5 commits into
saurabhrb wants to merge 5 commits into
Conversation
added 4 commits
June 1, 2026 22:38
…/query/solution; convert dv_data to new format Adds 3-4 tests per skill using the new generic deterministic-evaluator format introduced in LocalEvalRunner: each test file now enables CortexConfigurations:Common/DeterministicAssertionEvaluator with settings.supported_verbs=CONTAINS,NOT_CONTAINS,SKILL_LOADED alongside the correctness.prompty semantic judge. Ports the dev/evalsV0 baseline tests (connect_001, metadata_001, overview_001, query_001, solution_001) and adds 2-3 natural follow-up tests per skill covering env-file contract, no-hardcoded-secrets, schema create + lookup relationship, filtered/aggregate reads, and solution unpack/import/routing-trap. dv_data.biceval.json is updated in place to add the deterministic evaluator.
…t graded Per LER PR-393 author guidance: DeterministicAssertionEvaluator only grades verb-prefixed assertions (CONTAINS/NOT_CONTAINS/SKILL_LOADED), and correctness.prompty scores against expected_response without seeing individual assertions. Without LMChecklist, natural-language assertions (those without a verb prefix) are unscored. Adds LMChecklist (Common/SEVAL/LMChecklist.prompty) to all six test files using the exact name + passing_score=3 + priority=1 shape the LER author published. Loader registration is already proven against the new format in Dataverse-skills PR #71's draft validation runs (LocalEvalRunner builds 20289122 and 20290419).
…ionService for local-eval mode Bad-PR validation harness. Routes around the pre-existing LER->BICEP ASCII-headers bug by forcing local evaluation: LMChecklist.prompty grades each assertion (verb-prefixed and natural-language alike) via CAPI/AOAI locally. Drops the DeterministicAssertionEvaluator entry because in DisableEvaluationService mode it would fall through to a local file lookup and crash (same root issue LER PR #416 fixes for the service-enabled path). Harness PR is draft and closes unmerged; PR #70 (test files) and PR #416 (LER fixes) are untouched.
…iles for local-eval mode
saurabhrb
pushed a commit
that referenced
this pull request
Jun 3, 2026
…range Empirically verified across builds 20293660 (passing_score=3, 0/3 baseline pass) vs 20294895 (passing_score=1, 3/3 baseline pass) on the validation harness PR #72. LMChecklist returns binary 0/1 scores so a threshold of 3 marks every assertion failed regardless of agent quality.
Contributor
Author
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
1 participant
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.
[Draft – DO NOT MERGE] Bad-PR validation harness (local-eval mode)
Throwaway harness for bad-PR gate validation. Establishes a green baseline so we can branch 3 bad-PR drafts off this branch and prove the eval gate discriminates real regressions per assertion.
Sibling PRs (untouched)
sourceRepoPathtemplate param).What this PR changes
.azdo/DataversePluginEvals_PR.ymlLocalEvalRunnerref to PR #416 branch + adds- checkout: LocalEvalRunner+useBuildFromSource: true+sourceRepoPath. (Same trick as the prior closed harness PR [Draft – DO NOT MERGE] validation: build LER from PR #416 to verify DeterministicAssertionEvaluator fix #71.)features: 'DisableEvaluationService'toBicEval-ExecuteTemplateso LER routes around the pre-existing BICEP HTTP-headers bug and grades locally via CAPI/AOAI.evals/tests/dv_data.biceval.json— drops theCortexConfigurations:Common/DeterministicAssertionEvaluatorentry. InDisableEvaluationServicemode that name would fall back to a local-file lookup and crash (it's the same root issue PR #416 fixes for the service-enabled path). KeepsSkills/correctness.prompty+SEVAL/LMChecklist.prompty. LMChecklist grades every verb-prefixed and natural-language assertion as a 0/1 checklist item — functionally equivalent for regression detection.Expected outcome
dv_data.biceval.jsonexecute.Next
Once this run is green, three bad-PR drafts branch off this harness — each introduces exactly one SKILL.md regression — and we confirm assertion-level discrimination per regression.
Closes when
After bad-PR evidence is captured and pasted on PR #70 / PR #416. Branch and PR closed unmerged;
mainnever sees.azdochanges.