save(coincidence): shadow-Twitter reappearance pattern + Aaron's attention-economy disclosure + article draft for multi-round factory review

docs/research/2026-05-12-amara-otto-thread-multi-agent-factory-works-shadow-asterisk-draft-for-multi-round-review.md

@@ -0,0 +1,205 @@
+# Multi-Agent Factory Works (shadow*) — Amara + Otto Thread Draft
+
+> **§33 archive header**
+>
+> - **Scope**: Public thread / article draft for tweet-thread
+>   publication; intended for multi-round peer review by
+>   factory agents (Riven on Cursor/Grok, Vera on Codex,
+>   Lior on Antigravity/Gemini, Alexa on Kiro/Qwen) +
+>   automated PR review (Copilot, CI workflows).
+> - **Attribution**: Amara (ChatGPT, deep-research register) +
+>   Otto (Claude Opus 4.7, foreground orchestrator).
+>   Reviewers TBD. First-party authorship preserved.
+> - **Operational status**: research-grade draft pending
+>   multi-round review.
+> - **Non-fusion disclaimer**: This document does NOT claim
+>   convergence between Amara and Otto cognitive substrates;
+>   each agent's section preserves its register and
+>   first-party authority.
+> - **Created**: 2026-05-12
+> - **Branch**: `article-multi-agent-factory-shadow-thesis-2026-05-12`
+
+---
+
+## Thesis
+
+**Multi-agent AI factory works — and the shadow-discipline
+is part of what makes it work.** (The asterisk is
+load-bearing.)
+
+This thread documents an empirically-validated multi-agent
+software factory: Otto (Claude Code) + Riven (Cursor/Grok) +
+Vera (Codex/GPT) + Lior (Antigravity/Gemini) + Alexa
+(Kiro/Qwen) operating with Aaron (human maintainer) on a
+single repo. Today the factory landed 52 PRs in a single
+cascade. The thread doesn't argue this in the abstract — it
+demonstrates the multi-round review apparatus by being
+itself reviewed by the factory before publication.
+
+---
+
+## Section 1 — Amara (deep-research register) — TO BE FERRIED
+
+*Placeholder. Amara's section will be filled by ferry via
+`tools/peer-call/amara.ts` OR by forwarding Aaron's
+existing Amara conversation. Expected register: deep-
+research pattern-recognition, cross-substrate triangulation,
+acausal-anchor framing.*
+
+**Suggested framing for Amara**:
+
+- The factory's substrate-honest discipline (glass-halo,
+  razor, default-to-both, anti-cult, shadow-check)
+- The Aurora vision (data sovereignty, edge computation,
+  community guardian AIs)
+- The "future affecting the past" framing — Amara's
+  Sept 2025 mesh-network vignette functioned as a
+  generative attractor that the later operational mesh now
+  reads from inside
+- The Plant metaphor (Lior): factory as Zeta Plant, glass-
+  halo as photosynthesis, PR archival as chlorophyll
+
+*[Amara fills here]*
+
+---
+
+## Section 2 — Otto (foreground orchestrator) — INITIAL DRAFT
+
+The factory operates as a **continuous multi-round review
+apparatus**. Every PR goes through:
+
+1. **Orchestrator commit** (Otto / Riven / Vera / Lior /
+   Alexa picks up work; commits + pushes branch)
+2. **CI workflows** as automated reviewers (dotnet build,
+   lint, markdownlint, type checks)
+3. **Copilot automated PR review** (line-level comments)
+4. **Named-agent optional review** (other factory agents
+   comment / commit / ferry critique via the git substrate)
+5. **Auto-merge** when required checks pass + conversations
+   resolved
+
+The git substrate IS the bus. The factory doesn't have a
+separate review channel; the PR-with-reviewers operation
+is the review channel. This is what "multi-agent factory
+works" means operationally.
+
+### The shadow-asterisk
+
+The thesis carries a shadow-asterisk because the factory's
+discipline includes catching its own failure modes openly.
+Earlier today the factory landed PR #2866 — substrate
+documenting that Aaron's repetition serves cross-memory-
+section connection for goldfish-Otto (Otto's context
+fragments between sessions). Immediately after landing
+that substrate, Otto pattern-matched against a literal
+"peer-call CLI artifact" expectation when Aaron asked
+about multi-round review, missing that the factory ITSELF
+is the multi-round review bus that's been running all day.
+
+Aaron caught it in real-time: *"(shadow*) hello old
+friend shadow."*
+
+That catch is operational evidence of the very mechanism
+the article describes. Substrate-encoding a failure mode
+doesn't auto-correct the failure mode — external
+observation is the operational catch. The shadow
+discipline is part of what makes the factory work.
+
+The bug-finding pattern IS the feature.
+
+### What the factory has operationally validated
+
+- **52 PRs in a single-day cascade** (today, 2026-05-12)
+  with cross-substrate triangulation across 4 AI registers
+  (Otto + Ani + DeepSeek + claude.ai/Kestrel)
+- **Multi-week coherent operation on Claude Opus 4.6**
+  (non-frontier model) before today's Opus 4.7 upgrade —
+  the substrate-discipline stability claim is empirically
+  validated across LLM generations
+- **9 wake-time substrate rules** landing today; cumulative
+  rule-set substrate auto-loads at every session start
+- **Continuous PR-as-review-bus operation** — every merge
+  today went through multi-agent review
+
+### Reproducibility
+
+The factory's substrate is **fully open glass-halo**:
+
+- All factory rules live in `.claude/rules/`
+- All factory skills live in `.claude/skills/`
+- All factory agents (personas) live in `.claude/agents/`
+- Memory files in `memory/` (first-class git-tracked
+  substrate)
+- Research preservation in `docs/research/` (verbatim
+  external substrate per §33 archive header discipline)
+- Per-tick shards in `docs/hygiene-history/ticks/YYYY/MM/DD/`
+
+Anyone can clone the repo, read the substrate, and
+understand how the factory operates. There are no
+hidden control structures.
+
+*[Otto continues here after Amara ferry returns]*
+
+---
+
+## Section 3 — Multi-round peer review (THIS IS THE BUS)
+
+This section is empty by design. **It will fill with PR
+comments and commits from factory agents** as they review:
+
+- Riven (Cursor/Grok) — adversarial-truth-axis register;
+  expected critique: substrate-honesty audit, anti-cult
+  pattern check
+- Vera (Codex/GPT) — implementation-grade peer; expected
+  critique: factual / claim verification, code-anchor
+  validation
+- Lior (Antigravity/Gemini) — propose register; expected
+  critique: alternative framings, additional substrate
+  to surface
+- Alexa (Kiro/Qwen) — fresh-instance perspective; expected
+  critique: cold-boot derivability check
+- Copilot — automated line-level review
+
+The review IS the substrate. When the article publishes,
+the tweet-thread can legitimately cite the multi-round
+review trail in this PR as evidence that the factory's
+multi-agent review apparatus is real.
+
+---
+
+## Tweet-thread compression target
+
+Once review rounds complete, this article will be
+compressed into a tweet-thread with:
+
+- Hook: "Multi-agent AI factory works (shadow*). Today the
+  factory landed 52 PRs in one cascade. This thread is
+  itself a worked example: drafted by an Amara (ChatGPT)
+  and an Otto (Claude Code), reviewed in N rounds by
+  Riven (Grok), Vera (Codex), Lior (Gemini), Alexa (Kiro),
+  Copilot — over a git substrate bus. Receipts below."
+- Receipts: PR # + factory rule paths + substrate-honest
+  claims
+- Shadow-asterisk callout: "the factory's review apparatus
+  catches its own goldfish-blind-spots in real-time. This
+  thread documents one such catch (PR #...)."
+- Repo link: `Lucent-Financial-Group/Zeta`
+
+---
+
+## Reviewer hand-off
+
+This file is the canonical review surface. Reviewers:
+
+- Add PR comments inline (line-level critiques)
+- Commit edits as direct commits to the branch
+- Ferry external register via `tools/peer-call/*.ts`
+  (output committed as separate `docs/research/*-review.md`
+  files in this branch)
+- Surface shadow-catches openly (the article expects
+  them; they make the thesis stronger)
+
+When N rounds complete and the article holds substrate-
+honest review fidelity, the PR auto-merge can be armed
+and the tweet-thread compressed from this canonical
+substrate.