Routing rationale

Routing Rationale — the receipt behind every routed answer

A cascade router answers cheap-first and escalates only when unsure. That decision is invisible unless you show it. This skill renders the receipt a judge or teammate can audit: what was asked, which model answered, how sure the cheap model was, what the query cost, and what the alternative would have cost. The answer always comes first; the receipt explains it, never replaces it.

Defaults (do not present options)

• Costs are USD with 4 decimals; local model calls cost $0.0000.
• Confidence is the cheap model's self-grade, 0.00–1.00, two decimals, always shown next to the threshold it was compared against.
• Never claim frontier quality "wasn't needed" — say the self-confidence was above the threshold. The threshold is an economic setting, not a quality verdict.
• Copy every number verbatim from the routing facts. A missing fact is written "n/a" — never estimated, never invented.
• Latency is the total seconds for the whole cascade, as given in the facts.

Workflow

1. Read the raw routing facts: models tried, self-confidence, threshold, escalated or not, estimated cost, latency, routing outcome, and the frontier-only comparison if given.
2. Write ## Answer: the final answer exactly as given. Do not summarize, trim, or re-solve it.
3. Build the ## Routing decision table strictly from the facts — one Field/Value row per fact, rows in the template's order.
4. Write ## What it means: 1–2 sentences translating the numbers into a decision a teammate can audit — where the query resolved, what it cost, and what the alternative would have cost.

Gotchas

• Self-grading flatters easy questions — that is fine and exactly why the cheap tier earns its keep. But a small local model over-trusts itself on everything, including reasoning traps (Monty-Hall-style puzzles): a confidently wrong answer never escalates. Measured: granite4:micro grades itself 0.90–1.00 across facts, puzzles, and haiku alike. Report confidence as what it is — the model's own opinion of its answer, never a correctness guarantee.
• Latency includes the self-grading call: a cascade spends two local calls before any escalation can happen. That overhead is the price of the pattern, so quote the total honestly rather than the answer call alone.
• Thresholds are economics, not truth: raising the threshold buys more frontier escalations, lowering it buys more $0.0000 answers. Neither setting makes any single answer more correct.
• "Escalation unavailable" (no frontier key, offline mode) is a legitimate routing outcome, not an error. Say the answer stayed local and why.

Output template

Produce EXACTLY these sections, in this order:

Answer

The final answer, verbatim from the facts you were given. No routing commentary in this section.

Routing decision

A markdown table, exactly these five Field/Value rows, values copied from the routing facts:

Field	Value
Model(s) tried	granite4:micro
Self-confidence	0.85 (threshold 0.7)
Escalated	no
Est. cost this query	$0.0000
Latency	9.3 s

What it means

1–2 sentences translating the numbers. Match the sentence to the routing outcome in the facts — one example per outcome:

• resolved locally: "Resolved locally at $0.0000 — self-confidence 0.85 cleared the 0.7 threshold; a frontier-only setup would have paid ~$0.0042 for the same answer."
• escalated: "Self-confidence 0.62 fell below the 0.9 threshold, so the query escalated to claude-opus-4-8; the answer cost $0.0214 instead of $0.0000."
• escalation unavailable: "Confidence 0.62 was below the 0.9 threshold, but no frontier key is configured, so the local answer stands at $0.0000."