06. Cost-aware panel re-architecture

This example walks one real panel design from cost-unconscious (every lens on planner-class) to cost-aware (gradient workflow with disagreement-gated synth promotion). It is the canonical reference for any review-shaped workload (PR review, security audit, multi-lens critique).

Starting design (cost-unconscious)

A 6-lens review panel: 5 specialist lenses (security / UX / cost / architecture / correctness) plus 1 synthesizer. Each lens reads the same PR diff, applies its rubric, surfaces findings. The synthesizer reads all 5 sets of findings and produces one verdict.

flowchart LR
  PR[PR diff] --> L1[lens 1: security<br/>planner-class]
  PR --> L2[lens 2: UX<br/>planner-class]
  PR --> L3[lens 3: cost<br/>planner-class]
  PR --> L4[lens 4: architecture<br/>planner-class]
  PR --> L5[lens 5: correctness<br/>planner-class]
  L1 --> S[synthesizer<br/>planner-class]
  L2 --> S
  L3 --> S
  L4 --> S
  L5 --> S
  S --> V[verdict]
  style L1 fill:#fcc,stroke:#c00
  style L2 fill:#fcc,stroke:#c00
  style L3 fill:#fcc,stroke:#c00
  style L4 fill:#fcc,stroke:#c00
  style L5 fill:#fcc,stroke:#c00
  style S fill:#fcc,stroke:#c00

R5 trigger fires. 6 planner-class calls per PR. CLASS-UNIFORM GRAPH.

Cost-aware re-architecture

Apply A12 GRADIENT WORKFLOW + B12 MODEL ROUTER + B13 CACHE-AWARE PREFIX + S4 VALIDATION DECORATOR.

flowchart LR
  PR[PR diff] --> R[trivial router]
  R --> L1[lens 1<br/>reviewer-class]
  R --> L2[lens 2<br/>reviewer-class]
  R --> L3[lens 3<br/>reviewer-class]
  R --> L4[lens 4<br/>reviewer-class]
  R --> L5[lens 5<br/>reviewer-class]
  L1 --> DD[trivial disagreement<br/>detector]
  L2 --> DD
  L3 --> DD
  L4 --> DD
  L5 --> DD
  DD -->|agree ~80%| LS[light synth<br/>reviewer-class]
  DD -->|disagree ~20%| HS[heavy synth<br/>planner-class]
  LS --> V[verdict]
  HS --> V
  style R fill:#cfc,stroke:#0c0
  style L1 fill:#ffc,stroke:#cc0
  style L2 fill:#ffc,stroke:#cc0
  style L3 fill:#ffc,stroke:#cc0
  style L4 fill:#ffc,stroke:#cc0
  style L5 fill:#ffc,stroke:#cc0
  style DD fill:#cfc,stroke:#0c0
  style LS fill:#ffc,stroke:#cc0
  style HS fill:#fcc,stroke:#c00

Patterns applied:

• A12 GRADIENT WORKFLOW — heavy class only on the disagree path.
• B12 MODEL ROUTER — trivial-class router at entry.
• B13 CACHE-AWARE PREFIX — rubric + persona stable (cached); PR diff variable (suffix).
• S4 VALIDATION DECORATOR — the disagreement detector. Trivial- class classifier reading 5 lens outputs, deciding agree/disagree.
• R5 COST PRUNE — the refactor trigger.

Cost projection

Anthropic billing, verified 2025-11-14 (see claude-code adapter Section 9):

Version	Calls	Approx input	Approx output	$/run
Cost-unconscious	6 x planner (Opus)	30K	6K	~$0.90
Cost-aware (agree, ~80%)	1 trivial router + 5 reviewer (Sonnet) + 1 detector + 1 reviewer synth	29K	~3.9K	~$0.10-0.15
Cost-aware (disagree, ~20%)	adds 1 planner arbiter	+~5K	+~1K	~$0.20-0.30

Arithmetic check. Opus: $15/Mtok input + $75/Mtok output = 30K x $15/M + 6K x $75/M = $0.45 + $0.45 = ~$0.90. Sonnet: $3/Mtok input + $15/Mtok output = 28K x $3/M + 3.5K x $15/M ~ $0.08 + $0.05 + tiny router/detector = ~$0.13.

Blended: ~$0.12-0.18 per run across 80/20 agree/disagree mix. Reduction: ~5-7x on blended workload, ~6-9x on the dominant agree path.

What stays the same

• Quality envelope on agree case. Same 5 lenses fire. Same rubric. Reviewer-class meets the capability profile of “match against rubric, surface findings.”
• Quality envelope on disagree case. The planner-class arbiter still adjudicates when genuinely needed.
• Structural correctness. Still fan-out + synthesizer. Still no SHARED MUTABLE STATE. Still no CONTEXT THRASH.

When this is the WRONG call

A12 GRADIENT WORKFLOW pays off when N >= ~4 mid workers per planner call. For N < 4, or for:

• Low cadence + high disagreement rate (e.g. quarterly strategic reviews where every lens always disagrees) — the disagree path fires too often; route everything to planner-class directly.
• Quality stance explicitly declared — the operator paid for premium and wants premium throughout.
• N < 4 — flat single-class is structurally simpler with comparable cost.

Then the cost-unconscious version IS the right call.

Anti-patterns this re-architecture would create if done wrong

• ROUTER-AS-PLANNER — if the trivial router actually evaluates the PR instead of dispatching, it cannot be trivial class.
• INVERTED GRADIENT — if the lens fan-out is planner-class and the synthesizer is reviewer-class, you have paid heavy for the bulk and cheap for the decision.
• BUDGET-DRIVEN PROMOTION — promoting lenses to planner-class because the budget allows it, not because capability requires it.
• INVALIDATOR LEAK — a timestamp in the rubric, or mid-session effort change between detector and synth, would re-bill the entire prefix at input rate.
• TIMESTAMP IN RUBRIC — “as of 2025-11-14” in the stable rubric breaks cache on the rubric-update day.

Operator tuning

This walkthrough assumes balanced stance. With other stances:

• frugal — the heavy synth on the disagree path becomes reviewer-class too; arbiter capability ceiling lowered.
• quality — the gradient flattens upward; light synth becomes planner-class; reviewer lenses become implementer-class.
• unbounded — no gradient at all; everything planner-class (back to the cost-unconscious version, recorded in projection).

See stance and cap.