260707 Fugu Orchestration Adoption

Source: devlog/_plan/260707_fugu_orchestration_adoption/000_fugu_adoption.md

260707 Fugu Orchestration Adoption

Source: Sakana Fugu Technical Report, arXiv:2606.21228v1 (2026-06-19). Learned

orchestrator models (Fugu = hidden-state single-worker router; Fugu-Ultra =

Conductor-style GRPO-trained workflow generator) over a frontier worker pool

(GPT-5.5, Opus 4.8, Gemini 3.1). Analysis method: two parallel gpt-5.5 subagent

reads (technical deep-dive + PABCD transfer mapping, artifacts at

/tmp/arxiv2606/analysis_technical.md and

/tmp/arxiv2606/analysis_pabcd_transfer.md) cross-checked by a direct main-agent

read of the full paper text.

Why these mechanisms transfer

Fugu's value is trained routing, which a static skill cannot replicate. What DOES

transfer is what its trained orchestrators converged on behaviorally — the paper

documents the converged policies with concrete incidents:

• Intra-workflow isolation via access lists (paper §3.2.2): without isolation,

the first agent to touch the environment "sets the trajectory for all future

agents" — they name it orchestration collapse. Access lists make peer-output

visibility an explicit, per-dispatch decision; cross-workflow shared memory stays

on so lanes do not re-discover artifacts.

• Crux-matched aggregation (paper §4.4): tree topologies picked the aggregator

by the disputed domain (Gemini for trivia synthesis, GPT for a spectral-number

derivation) and the paper argues fixed-aggregator systems are capped at the

aggregator's own ceiling.

• Cross-model builder/verifier alternation (paper §4.4): GPT-as-builder with

Opus dispatched at critical debugging moments; a clean-slate GPT pass broke an

Opus dead-end on a SWE Bench Pro OTP bug (root cause was client-side concurrency,

not the server path Opus tunneled into). Different model families decorrelate

blind spots.

• Specialist first-principles re-derivation (paper §4.4 FEAL differential

cryptanalysis): after Opus built the attack, GPT was tasked as a math specialist

to re-derive the differential constant bit-by-bit before adoption.

• Late-stage fan-out gains (paper §4.3.1 AutoResearch, 123 experiments x 3

seeds): orchestration was "competitive early... pulls ahead after mid-training,"

i.e. once coarse configuration search shifted to fine optimizer/schedule tuning;

it also reduced cross-seed variance.

Adopted rules

Rule	Tag	Placement
DISPATCH-ISOLATION-01	DEFAULT	dev-pabcd §7.1 (canonical, cli-jaw); codexclaw Delegation Model
SPECIALIST-CRUX-01	HEURISTIC	same
REVIEW-DECORRELATE-01	HEURISTIC	same
LOOP-FANOUT-TIMING-01	HEURISTIC	§10 Optimization-Loop Meta-Rules (all three)
COLLAPSE-AGGREGATOR-01	DEFAULT	§11.9 (canonical bullet, cli-jaw section); codexclaw Meta-Rules

Single-source induction caveat: all five rules are grounded in one vendor report

whose qualitative examples are curated. Tags were chosen accordingly (three

HEURISTIC, two DEFAULT where the paper gives a mechanism-level failure mode:

orchestration collapse, fixed-aggregator ceiling). Revise when a second domain's

evidence contradicts them.

Explicitly NOT adopted

• Hidden-state routing head, soft-label SFT routing, sep-CMA-ES, GRPO Conductor:

require trained models and reward infrastructure; no static analogue.

• Model-brand role bindings ("GPT builds, Opus debugs, Gemini aggregates"): roles

come from current observed capability evidence, never hard-coded brands.

• Single-interface abstraction hiding orchestration: PABCD's value is visible

phase discipline and attested evidence; hiding the process would invert it.

• Workflow depth cap (<=5 steps) and quality/latency mode split: already covered

by DIVERGE-TIER-01 and C0-C5 class scaling.

Patch map (2026-07-07)

• pabcd_initiative/skills/dev-pabcd/SKILL.md — new §7.1, §10 bullet, §11.9

bullet; README hardening line.

• cli-jaw/skills_ref/dev-pabcd/SKILL.md — new §7.1 (generalizes §11.8 star

topology to all parallel dispatch), §10 bullet, new §11.9 section.

• codexclaw/plugins/codexclaw/skills/pabcd/SKILL.md — three Delegation Model

bullets, two Optimization-Loop Meta-Rules bullets; record at codexclaw devlog

260707_fugu_orchestration_adoption/000_record.md.

Follow-ups

• jawcode port: no dev-pabcd doc there; fold the five rules into

packages/coding-agent/src/defaults/jwc/skills/{team,plan,goal} phrasing.

• references/loop-engineering.md (canonical + codexclaw): sync §11.9 and the

fan-out timing rule into the full-rules reference on its next touch.

• cxc-loop skill (codexclaw): mirror COLLAPSE-AGGREGATOR-01 into its

divergence/collapse section on next touch.