Method checkRegression

src/llm/speculative.ts:162–174

* Returns a warning string if speculative decoding appears to be HURTING * throughput (draft-on median meaningfully below draft-off median), else null. * Warns at most once per session to avoid nagging.

()

   * Warns at most once per session to avoid nagging.
   */
  checkRegression(): string | null {
    if (this.warned) return null;
    const on = this.median(this.withDraft);
    const off = this.median(this.withoutDraft);
    if (on === null || off === null) return null;
    // 10% slower with the draft on → it's not helping for this workload.
    if (on < off * 0.9) {
      this.warned = true;
      return `Speculative decoding looks counterproductive here: ~${on.toFixed(1)} tok/s with the draft vs ~${off.toFixed(1)} without. ` +
        `Acceptance is likely low for this task. Consider a smaller/faster draft model, or disable it.`;
    }
    return null;
  }

  stats(): { withDraftTps: number | null; withoutDraftTps: number | null } {
    return { withDraftTps: this.median(this.withDraft), withoutDraftTps: this.median(this.withoutDraft) };