Glossary
The model-downgrade glossary
Plain definitions for the words rightmodeler uses.
The decision
The words for calling a swap safe.
- Evidence-backed model downgrading
- Proven on your traces, not benchmarks.
- Model downgrade audit
- The pass that finds the safe swaps.
- Downgrade-safe
- Quality holds, with evidence to spare.
// steps/summarize.ts
export const summarize = step({
// model: "gpt-5.6", · replaced
model: "gpt-5.4-mini", // 85% cheaper
floor: 0.90, // Q 0.94 clears it
});
The guardrails
Where the audit stops.
- Quality floor
- 0.90 by default. Below it, no swap.
- Cascade risk
- One cheap step degrades the next.
- Abstain
- Weak evidence, no recommendation.
The evidence
What gets measured, and by whom.
- Trace
- One agent run, recorded step by step.
- Reference evidence
- Graded against what you shipped.
- LLM-as-judge
- Nothing grades its own work.
// rightmodeler.config.ts
export default audit({
traces: "./traces/*.jsonl", // 214 runs
reference: "shipped", // as shipped
judge: "cross-family", // no self-grade
abstain: "weak-evidence", // say no
});