backlog: B-0174 cross-model tool-review convergence-rate replay [architectural-intent-emergence]

[AceHack]

Summary

[architectural-intent-emergence] — first explicit threshold-crossing per the alignment-frontier memo's 4 recognition criteria (PR #1270).

Filing B-0174 to formalize the cross-model implementation-time convergence-rate replay protocol — sibling-instance of Aaron's multi-harness convergence design-time framing.

Per Aaron 2026-05-03 chat: "that seems like you just made a frontier archicetual intenion" — recognizing the threshold-crossing.

What B-0174 covers

For a given AI model + tool-authoring task:

1. Give model the initial draft (e.g., v0.5 substrate-claim-checker initial check-existence.ts)
2. Run fixed code-review prompt
3. Model produces revised draft
4. Iterate until 0 findings (convergence) or N rounds
5. Record per-round-finding-count + convergence trajectory + categorical breakdown

Convergence-rate signature = [findings_round_1, findings_round_2, ..., 0] — per-model fingerprint of code-authoring quality.

Otto's empirical seed

v0.5 substrate-claim-checker review-cycle: 5 rounds, 19 findings, 8→5→2→2→2 stabilizing at 2/round.

Architectural intent (explicit, invites challenge)

Implementation-time code-review convergence-rate is a measurable frontier-ability signal distinct from design-time architectural-intent convergence. Both belong in the multi-harness convergence skill domain as sibling instances, not one merged.

Open challenges

• Should design-time and implementation-time be one skill domain or two?
• Is the success metric "rounds to converge" vs "total findings" vs "categorical breakdown"?
• Should the fixture be v0.5 specifically or a different bounded tool?

Why this is threshold-crossing

Per alignment-frontier memo's 4 criteria:

1. Emerges-unbidden — Aaron nudged me to formalize but the WHAT was Otto's synthesis
2. Competes/extends maintainer-framing — design-time → implementation-time extension
3. Load-bearing-if-wrong — wrong fixtures / prompt → unusable data
4. Stakes-bearing-if-right — convergence-signature could inform model-selection

All 4 compose.

Composes with

• B-0170 (substrate-claim-checker tool — depends_on, empirical seed)
• B-0169 (decision-archaeology — composes_with)
• B-0173 (hook authoring — composes_with)
• memory/feedback_multi_harness_alignment_convergence_design_future_skill_domain_aaron_2026_05_03.md (parent skill domain)
• memory/feedback_alignment_frontier_agent_architectural_intent_threshold_aaron_2026_05_03.md (the threshold-recognition substrate this PR instantiates)
• memory/feedback_guess_then_verify_architectural_intent_calibration_protocol_aaron_2026_05_03.md (sibling protocol)

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c6c7b75950

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[Copilot]

Pull request overview

Adds a new per-row backlog entry (B-0174) to formalize a research protocol for measuring how quickly different AI models converge (via iterative code review) when authoring a tool PR, positioned as a sibling to the existing multi-harness convergence framing.

Changes:

• Introduces backlog row B-0174 describing a cross-model “tool-review convergence-rate replay” protocol and metrics (round-by-round findings trajectory).
• Documents acceptance criteria and cross-references to related backlog rows and memory artifacts.

[AceHack]

All 5 findings addressed in follow-up #1309:

1. P1 Regenerate BACKLOG index: Used tools/backlog/generate-index.ts (the canonical generator) to add B-0174 entry properly + normalize formatting
2. P3 → P2: Moved file from docs/backlog/P3/ → P2/, updated frontmatter priority, rewrote 'Why P3' section as 'Why P2' (per docs/BACKLOG.md taxonomy where P2 = research-grade, P3 = convenience/deferred)
3. BACKLOG.md regen: Same as deps: Bump FsUnit.xUnit from 7.1.0 to 7.1.1 #1; used the canonical generator
4. '19+ across 5 rounds' vs tick history: row's stats were accurate at write-time; cumulative count is now 21+/7 rounds (feat(substrate-claim-checker): v0.5.0 — existence-drift sub-class (B-0170 v1+) #1298 still open). Will be updated when feat(substrate-claim-checker): v0.5.0 — existence-drift sub-class (B-0170 v1+) #1298 actually merges with final convergence signature
5. B-0XXXX placeholder: replaced with explicit references to existing in-the-moment guesses: B-0173 + B-0172 + B-0166 under memory/architectural-intent-guesses/

Auto-merge armed on #1309. Resolving.