Model Spar

Model Spar - the family-diverse panel

A panel is not "ask three models and pick the best answer." A panel exists to make correlated models disagree - which is exactly what a single-provider or single-family answer cannot do.

Ground rules

1. Family-diverse or nothing. Same-family models correlate (same training data, same blind spots). A "5 model panel" of one family is a fancy single opinion. Ship families: frontier + contending + local.
2. Identical context pack. Every seat gets the exact same input, verbatim
   • the panel measures the MODELS, not the prompt quality differences. (Bonus: identical context = cache-write once, cache-read N times, which is why panels are cheap - see the cost-runway-tracker skill.)
3. Roles give dissent structure. Assign each seat an explicit stance (structural, contrarian, momentum, risk-first) so disagreements are legible, not random.
4. Never average to mush. A verdict is produced FROM the disagreement: surface what they split on, then decide with the evidence table. "Record who said what" is mandatory; consensus-by-averaging hides information.
5. Pay attention to the lone voice. The single contrarian is often carrying the information the herd missed - specifically ask for each seat's "one risk the others won't flag."

The pattern

1. BUILD the context pack   - the problem, constraints, data, nothing else
2. FILL the roster          - 1 family per seat, roles assigned
3. RUN the panel            - same input, independent, no chat between seats
4. DIFF the outputs         - agreement map + split map
5. DECIDE from the split    - evidence table wins, name the deciding factor
6. RECORD the dissent       - the minority report is an artifact, not a footnote

Known pitfalls (learned the hard way)

• Family-blind "diversity". claude/gpt/grok/qwen is diverse; claude + a claude clone is not. Check provider, not name.
• Reasoning-token starvation. Reasoning flagships burn their whole output budget on chain-of-thought when max_tokens is set near the reasoning floor - content returns null with finish_reason=length. Always set max_tokens well above the reasoning budget (live verified: opus-class needs ~16000+ for long-form panel answers; flash-class ~4096).
• Panel drift on budgets. The free fallback roster must be pre-defined, so a failed seat degrades to a known model, not whatever is cheapest today.
• Cost illusion. 5 flagships feels expensive; with cache-read discounts (6-50x for shared context) it is not. Measure the actual panel cost, don't assume it.

Output discipline

• Every panel run produces: the verdict, each seat's answer verbatim, the split map, and the deciding evidence. The minority report is retained.
• No verdict without a named decision driver; "they disagreed" is not a decision.

Skill pattern: dissent-by-design -> decision from the split. The value is the disagreement, which correlated models cannot give you.