feat(calibration): counterfactual prompt replay — offline harness + budget-gated advisory CI check#8271
Merged
Merged
Conversation
…udget-gated advisory CI check (#8221, #8222) Closes the #8218 sub-epic's phases 3+4 against the #8219 contract verbatim: - scripts/counterfactual-replay-core.ts (pure, unit-tested): abstention-safe variant-output parsing (never coerced), scoreBacktest mapping over the #8220 fixture plan, budget arithmetic + resume cursor, run-accounting invariant, artifact cache keys, Pareto baseline comparison - scripts/counterfactual-replay.ts (thin IO): ollama smoke path only (BYOK refused loudly), raw outputs cached under a local artifacts dir so re-scoring is free, --baseline prints the shared renderer's comparison; exit reflects operational success only, never verdict - src/services/ai-review.ts: REVIEW_PROMPT_VERSION + buildCanonicalJudgePrompt (the pure, input-free judge prompt the dual-checkout replay diffs) - scripts/print-review-prompt.ts: the #8139 dual-checkout extractor - .github/workflows/counterfactual-replay-check.yml: path-filtered on the judge-prompt surface, hard-gated on an explicitly configured replay budget (visible skipped-notice otherwise), byte-identical prompts short-circuit, fail-open on every step, update-in-place marker comment, persisted calibration.counterfactual_backtest_run events feeding the SAME REGRESSED-verdict track record as the threshold/logic backtests - docs: worked ollama example + the prompt-change workflow
Contributor
Logic backtestReplayed 0 historical case(s) for Backtest comparison:
|
Deploying with
|
| Status | Name | Latest Commit | Preview URL | Updated (UTC) |
|---|---|---|---|---|
| ✅ Deployment successful! View logs |
loopover-ui | 7d81d1a | Commit Preview URL Branch Preview URL |
Jul 23 2026, 09:47 PM |
Contributor
|
Superagent didn't find any vulnerabilities or security issues in this PR. |
Bundle ReportChanges will increase total bundle size by 8.42kB (0.11%) ⬆️. This is within the configured threshold ✅ Detailed changes
Affected Assets, Files, and Routes:view changes for bundle: loopover-uiAssets Changed:
|
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #8271 +/- ##
=======================================
Coverage 92.16% 92.16%
=======================================
Files 786 786
Lines 78941 78943 +2
Branches 23840 23840
=======================================
+ Hits 72757 72759 +2
Misses 5062 5062
Partials 1122 1122
Flags with carried forward coverage won't be shown. Click here to find out more.
|
This was referenced Jul 23, 2026
JSONbored
added a commit
that referenced
this pull request
Jul 23, 2026
orb-v3.3.0 shipped stable, so the beta cutter is idle until the manifest's declared target moves forward. The calibration-expansion epic's merges since (per-repo loosening #8264, reviewer votes #8265, tightening #8267, AMS loop #8270, counterfactual replay #8271) are feats — the inferred bump is minor, so the next automated snapshot targets orb-v3.4.0-beta.1.
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.
Summary
Closes #8221. Closes #8222. (Phases 3+4 of sub-epic #8218, one PR per the fewer-larger-PRs direction.)
Implements the #8219 contract verbatim — every shape cited, none re-invented:
scripts/counterfactual-replay.tswith the pure core incounterfactual-replay-core.ts(the backtest-corpus-export split). Abstention-safe verdict parsing (string-aware balanced-object extraction; anything unparseable is counted, never coerced),scoreBacktest/compareBacktestScoresapplied unchanged over the counterfactual: pure fixture assembler — replayable cases from the raw-context corpus #8220 fixture plan, neuron-budget arithmetic with a first-gap resume cursor, a run-accounting invariant asserted before any artifact persists, and a content-keyed raw-output cache under a local artifacts dir so re-scoring is free. Ollama is the only wired provider (BYOK refused loudly, per the contract); exit codes reflect operational success only — never a verdict.REVIEW_PROMPT_VERSION+buildCanonicalJudgePrompt()insrc/services/ai-review.ts(pure, input-free — the exact default-config system prompt), extracted per checkout byscripts/print-review-prompt.ts(the Extend the backtest check to logic/regex changes via a dedicated CI job, using captured raw context #8139 dual-checkout mechanism, sentinel-tolerant for pre-counterfactual: advisory check when reviewer prompt versions change #8222 bases)..github/workflows/counterfactual-replay-check.yml, path-filtered on the judge-prompt surface. Hard spend guard first: without the explicitly configuredCOUNTERFACTUAL_REPLAY_BUDGET/COUNTERFACTUAL_REPLAY_MODELvariables + provider secret it posts a visible "skipped: no replay budget configured" notice and does nothing else. Byte-identical base/head prompts short-circuit; the sample seed is the head SHA (deterministic per push); every step fails open; the comparison lands as an update-in-place marker comment; each run persists acalibration.counterfactual_backtest_runevent feeding the SAME REGRESSED-verdict track record as the threshold/logic backtests (backtest-track-record.tsnow aggregates all three siblings).qwen3:8bis finishing — the mechanism already ran end-to-end: 40 scored / 0 abstained / 0 accounting drift on the seeded production sample).Validation
npm run test:ci: green except one unrelated suite (test/unit/miner-package-skeleton.test.ts) that hit the known subprocess-spawn timeout while two gates ran concurrently on this machine — 6/6 standalone in this worktree; no miner packaging file is touched here.npm audit --audit-level=moderateclean;actionlintclean on the new workflow.src/**diff (the two new exports); the core has 12 tests covering every parsing/scoring/accounting arm.