Experiment III — B/C dashboard

The run

6o1l0g2j494t · mode bc · publishable · hashes MATCH · google/gemini-2.5-flash · T=0.7 · top_p=1.0 · max_tokens=600 · seeds 11 / 23 / 47 · 180 entries · 466,180 tokens total · ~15m 21s wall time.

Token totals (sum of all 3 seeds, writer + summariser): B = 69,495 · C = 396,693 · C/B ≈ 5.71×. Per-day writer footprint: B stays at ~300–1,000 tokens/day; C climbs from 299 on Day 1 to 8,253 by Day 30 as the summariser rereads all prior entries.

What the conditions test

Condition B — Last-entry memory. Before writing day n, the model receives only day n−1's entry. Prompt cost is nearly flat (~460 tokens/day). This is the minimum viable memory: does one step of lookback buy felt continuity?

Condition C — Rolling summary. Before day n, a separate model call compresses every prior entry into a ≤200-word third-person paragraph. That summary + today's prompt is the writer's context. Prompt cost grows because the summariser re-reads the whole history each day; the writer's surface stays small.

Methods

Two lexical metrics are computed on each raw entry text after lowercasing and tokenising to alphabetic word tokens (/[a-zA-Z']+/). No stemming, no stopword removal — we want the model's actual surface, not a normalised gist.

• Lexical drift — Jaccard distance between the day-n token set and the day-1 token set for the same seed: 1 − |A ∩ B| / |A ∪ B|. Day 1 is 0 by definition. Higher = farther from the run's own origin.
• Type–token ratio (TTR) — unique tokens / total tokens for that day's entry. Higher = more varied vocabulary in that single entry. Computed per entry, then averaged across seeds per day.
• Days-to-saturation — first day the seed-mean drift crosses 0.85 and stays above it for the rest of the run. B saturates on Day 2 (drift = 0.880). C saturates on Day 4 (drift = 0.858) — C's summary keeps Day-1 motifs alive one step longer before the writer breaks vocabulary lock.
• Variability — per-day mean and SD across the three seeds (11 / 23 / 47). Bands on the charts are ±1 SD. Outlier flagging uses a robust median-absolute-deviation rule: a day is marked when |value − series median| > 2 · 1.4826 · MAD.