Skip to content

Eval catalog plan versus direct brief A/B scaffold - #58

Draft
myobie wants to merge 8 commits into
mainfrom
agent/catalog-plan-ab-scaffold
Draft

Eval catalog plan versus direct brief A/B scaffold#58
myobie wants to merge 8 commits into
mainfrom
agent/catalog-plan-ab-scaffold

Conversation

@myobie

@myobie myobie commented Jul 30, 2026

Copy link
Copy Markdown
Collaborator

Outcome

This draft preserves the realistic model-free catalog-plan versus direct-brief comparison while adopting the accepted Resource-linked PR115 discovery contract.

  • Merged current evals main normally at merge commit 17bc728; no force-push or history rewrite.
  • Arm A discovers plan truth only through childless Agent Spec Resources such as resource "receipt-report" _tag="plan" uri="file:plans/receipt-report/plan.kdl".
  • The referenced plan.kdl owns identity, owner, versions, parents, reasons, and either external Markdown content or inline intent; the Agent Resource owns only its local role/link.
  • The existing A/B scenario remains external Markdown: initial and steered intent bytes are unchanged and still match the direct-message arm exactly.
  • One focused second plan proves inline intent through the same Resource-link boundary without changing the comparison.
  • Arm B still delivers the byte-identical briefs through isolated durable st2 message send operations with inReplyTo lineage, then recovers them through message ls|read.
  • Both arms lose volatile state and recover from local durable folders while their remote sources are offline.
  • Legacy plan-ref and agent-owned inline plan truth are absent.

Model-free comparison result

Endpoint Catalog plan Direct brief Verdict
Cold resume pass pass tie
Exact intent recovery pass pass tie
Human steering pass pass tie
Worker acceptance evidence evaluator-only evaluator-only tie

The conclusion remains no-measured-advantage for cold resume, intent recovery, or acceptance evidence. The plan adds static schema/graph validation and resolved-content provenance. Neither read-only plan inspection nor direct-message delivery reports worker acceptance.

Plain copied folders remain sufficient and casRequired is false.

Exact pairing

The contract stores no content digest or historical record, so it cannot prove that an earlier declaration, parent link, or content file stayed unchanged. The repository-wide released runner remains 0fed14b; the experimental 8a76b6e artifact is isolated to this cell.

Validation

  • Exact experimental runner: st2 eval ./cells/catalog-plan-vs-direct-brief/ --keep24 PASS / 0 FAIL / 24 gating judges.
  • Exact target forms: external Markdown content and inline intent both pass validate|list|show|inspect, with catalog bytes unchanged.
  • Released 0fed14b runner: bin/check-corpus.shPASS, 60 maintained cells, 119 KDL files, 12 nested catalogs.
  • Shell syntax and git diff --check: pass.
  • Model/provider calls: 0.

Honest limits

The reference implementation proves the frozen task and arm-neutral judges are valid; it is not substituted for two agent executions. Correctness, coordination messages/bytes, model input/output tokens, cost, and wall duration remain unresolved provider-backed endpoints. A live A/B requires separate authorization naming the model, effort, per-arm budgets, order/counterbalancing, spend ceiling, cleanup, and rollback.

This PR remains draft and unmerged pending Nathan review of the exact pair.

@myobie
myobie marked this pull request as ready for review August 3, 2026 16:20
@myobie
myobie requested a review from schickling August 3, 2026 16:21

myobie commented Aug 3, 2026

Copy link
Copy Markdown
Collaborator Author

Moved this shared PR from draft to ready at the unchanged accepted head a8e223f99c8f4b4e67e3390dec507fcbe992edf6 after Nathan's review, and requested @schickling for the normal review path. No merge is authorized until that review returns.

@myobie
myobie marked this pull request as draft August 3, 2026 16:51

myobie commented Aug 3, 2026

Copy link
Copy Markdown
Collaborator Author

Returning this shared PR to draft at unchanged head a8e223f99c8f4b4e67e3390dec507fcbe992edf6 because the accepted PR115 discovery contract changed. The eval shape and model-free conclusion will be preserved; plan discovery will move to an Agent KDL Resource with _tag="plan" whose referenced plan.kdl owns plan truth. No branch rewrite or product-cell run will occur until the exact corrected PR115 head and interface arrive.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant