Skip to content

red-lines-and-anti-patterns.md

Latest commit

 

History

History
54 lines (46 loc) · 3.81 KB

red-lines-and-anti-patterns.md

File metadata and controls

54 lines (46 loc) · 3.81 KB
Raw

Red Lines And Anti-Patterns

This page records public-safe rationale for v0.6 control-surface regression checks.

Naming note: BriefLoop is the public project name, and MABW remains the current implementation lineage and compatibility surface. The red lines apply to both names.

Red Lines

  • Do not turn briefloop run into a Python brief generator.
  • Do not let Python auto-repair prose or claim the repair is semantically complete.
  • Do not let finalize stand in for quality gates.
  • Do not treat feedback text as claim evidence.
  • Do not treat FeedbackIssue as guidance: guidance must be Human-authored, separately approved, compatible with the successor direction, and explicitly opted into; there is no automatic path from issue or Gate finding to a successor snapshot.
  • Do not tell users "the system learned this." Even approved guidance has no later-run effect until a Human explicitly starts a compatible successor with reuse enabled; utility is NOT MEASURED.
  • Do not let future-stage artifacts or issues globally block a fresh workspace.
  • Do not let planned blocking issues be bypassed by continue.
  • Do not require reader-facing briefs to retain internal [src:<claim_id>] markers.
  • Do not fetch live market data inside deterministic gate checks.
  • Do not present provenance projection as semantic proof that a source supports a claim.
  • Do not present a precision gate as sufficient by itself when it can be passed by omitting important information.
  • Do not present coverage checks or Screener output as full-recall proof.
  • Do not execute arbitrary shell strings from evaluation fixtures.
  • Do not ship private prompts, internal paths, real URLs, tokens, or commercial benchmark cases as public fixtures.
  • Do not treat the BriefLoop public naming layer as a runtime rename; multi-agent-brief, briefloop, /briefloop, /mabw, package/module paths, artifact names, workspace formats, and MABW experiment IDs remain compatible until an explicit migration changes them.

Anti-Patterns

Anti-pattern Why it is unsafe
Full Python pipeline resurrection Breaks the subagent-first runtime boundary.
Global required-artifact blocking Makes fresh workspaces look blocked by downstream artifacts that are not due yet.
Repair plan equals repair completion Lets the Orchestrator skip unresolved feedback.
Gate findings auto-create feedback issues Hides the Orchestrator decision that should route repair or human review.
Precision-only quality gate Can reward omission: unsupported claims disappear, but important missing topics are not surfaced. Pair with coverage-side checks when the risk is omission.
Screener as recall proof Screener ranks and caps candidates; it does not prove the source universe contains no other material item.
Final brief as auditable source surface Reader-facing output should not need internal claim markers; use the delivery source appendix and output/source_appendix.md audit copy for source lists.
Provenance graph as truth graph The projection records citation/control relationships; semantic support still needs audit or human review.
LLM-as-judge default eval Turns deterministic regression checks into model-dependent scoring.
Shell-string eval commands Expands the fixture runner into an arbitrary command executor.

These red lines are enforced through focused tests, packaged public-safe evaluation cases, and the support matrix.

中文红线：FeedbackIssue 只是问题记录，不是读者偏好。Guidance 必须由 Human 撰写、独立批准、与 successor direction 兼容，并在启动 successor 时明确 opt-in； 从 issue 或 Gate finding 到 successor snapshot 不存在自动路径。不要告诉用户 “系统学会了这一点”；即使已批准 guidance，效用仍是 NOT MEASURED，也没有隐式 后续 run effect。