Model tiering: Fable/Opus judgment-execution split (quality-boost + economy directions, default untouched)

[Imbad0202]

Motivation

Lance Martin (Anthropic), "Cost effective harnesses with Fable" (X article, 2026-07-10): many tasks have intelligence asymmetry across their tokens; harnesses can exploit it via three patterns (frontier model as orchestrator / advisor-at-checkpoints / verifier). Measured results: advisor-checkpoint config reached ~90% of Fable-solo quality at ~34% of token cost (Parameter Golf); delegation only pays when workers absorb enough tokens to offset the fixed per-handoff coordination cost (BrowseComp: +60% cost for nothing on the small subset, 96% score at 46% cost on the full set).

ARS maps onto this cleanly: light slash commands are already sonnet-pinned (correct per the small-task finding), but all pipeline agents are model: inherit — a Fable session burns Fable tokens on mechanical formatting/compilation work, and an Opus session gets no frontier judgment at the gates.

Design decision (frozen 2026-07-11)

One tiering mechanism, two explicit directions, default untouched:

1. Default: model: inherit everywhere — session model runs everything. Quality ceiling preserved; no opt-out surprise.
2. Quality-boost direction (Opus session): judgment-type agents upgrade to the frontier tier at Stage 2.5/4.5 gates and final review. Quality strictly rises vs Opus-solo; extra cost concentrated where the article shows it matters (mid-task re-ranking, not upfront planning).
3. Economy direction (Fable session): execution-type agents downgrade ONE tier max — floor is Opus, never Sonnet (frozen: academic prose quality tolerance is untested; the article's 90%/34% numbers came from ML tuning, not scholarly writing). Documented explicitly as quality-for-cost with the article's numbers cited.

Agent classification (frozen; 39 agents)

Judgment-type (25) — stays on session model / upgrade candidates:

• deep-research (10): socratic_mentor, research_question, research_architect, synthesis, devils_advocate, editor_in_chief, ethics_review, risk_of_bias, meta_analysis, source_verification
• academic-paper (6): socratic_mentor, argument_builder, structure_architect, peer_reviewer, revision_coach, literature_strategist
• academic-paper-reviewer (6): eic, methodology_reviewer, domain_reviewer, perspective_reviewer, devils_advocate_reviewer, editorial_synthesizer (mechanical by v3.6.2 design, but emits the final decision letter — kept judgment-type conservatively until data says otherwise)
• academic-pipeline + shared (3+1): pipeline_orchestrator, claim_ref_alignment_audit, integrity_verification, compliance_agent (has tier-based block authority)

Execution-type (12) — economy-direction downgrade candidates (floor: Opus):

• deep-research (4): bibliography (citation existence already handled by the deterministic gate; note the sha256 F2 lock on this file), timeline_extraction, report_compiler, monitoring
• academic-paper (6): intake, draft_writer (highest-token, highest-savings, highest-risk downgrade point — one tier max, flagged in docs as the suite's most quality-sensitive downgrade), abstract_bilingual, citation_compliance, visualization, formatter (STAMP-ONLY by design)
• academic-paper-reviewer (1): field_analyst
• academic-pipeline (2): collaboration_depth (advisory-only, never blocks), state_tracker

Implementation notes

• Opt-in switch (naming TBD, e.g. ARS_MODEL_TIERING=economy|quality-boost); absence = byte-equivalent current behavior (same philosophy as terminal_policies).
• Prompt caching: route repeated same-stage calls (e.g. reviewer re-review loops) to the SAME worker so its cache accumulates — a fresh worker per call re-pays the context write and can erase the savings (article guidance item 4).
• Lint: classification drift guard (an agent added without a tier assignment fails CI).
• v3.7.0 lesson applies: tiers must be relative to the session model (never hard-pinned model ids), or a floor becomes a downgrade ceiling on the next model generation.