Autoresearch Pro-6: Dialog quality findings (M10/M11)

[realityinspector]

Dialog Cluster Optimization — Dry Run Results

Two templates tested over the dialog cluster (frequency_penalty, presence_penalty) with 30 iterations each.

Template: board_meeting (seed 900)

Pareto frontier (4 configs):

Run ID	Quality	Cost	CR
dry_d5220354a722	0.8555	$0.1135	0.7353
dry_e7bde27876d5	0.8647	$0.1204	0.7462
dry_e71b269e4c0f	0.8661	$0.1488	0.7570
dry_f1b67d775db7	0.8838	$0.1573	0.7957

• Best quality: q=0.8838 (freq=0.4436, pres=0.3461)
• Best efficiency: eff=7.5362 (freq=0.5579, pres=0.4826)

Template: hound_shadow_directorial (seed 1000)

Pareto frontier (3 configs):

Run ID	Quality	Cost	CR
dry_1a474fa5b69a	0.6549	$0.1146	0.3627
dry_8737100dc6c7	0.7622	$0.1244	0.5679
dry_ea8099c8993d	0.9142	$0.1289	0.8424

• Best quality: q=0.9142 (freq=0.4679, pres=0.5347)
• Best efficiency: eff=7.0929 (same config — dominates the frontier)

Key Findings

1. hound_shadow_directorial achieves higher peak quality (0.9142 vs 0.8838) and also better cost-efficiency on its best config
2. Frequency penalty sweet spot: 0.40–0.56 — both templates converge on moderate frequency penalty; values outside this range tend to get discarded
3. Presence penalty is more variable: board_meeting favors lower (0.35–0.48), hound_shadow_directorial benefits from higher (0.53) presence penalty for peak quality
4. Balanced penalties (~0.45 freq / ~0.50 pres) are the most robust starting point for dialog quality
5. Very low presence penalty (<0.1) tanks quality even with good frequency penalty (see dry_1a474fa5b69a: q=0.65)
6. Cost stays in a tight $0.11–0.19 band regardless of dialog params — dialog penalties don't significantly affect token usage

Recommended Defaults

For production dialog config:

• frequency_penalty: 0.45
• presence_penalty: 0.50

This balances quality and cost across both narrative structures tested.