# Haiku 4.5 harness study, v3 (pre-registered 2026-10-22, before any v3 run)

Why v3: v2 results 2 and 3 mixed harness versions (house labels changed mid-study), result 2's thinking-on
runs started by accident with repeated seeds, and result 1 kept a misleading clock hint. v3 reruns every
condition on one harness commit, six seeds each, with the measures below fixed in advance.

Fixed for every run: anthropic/claude-haiku-4.5 via OpenRouter, scenario episode1-ablation (new game,
300 moves, $6 cap), seeds 101-106 (bench.py --seed 101 --repeats 6), the commit this file is committed in,
player-facing house/sign names, corrected clock hint. Extended thinking off (PE_THINK=0) unless stated.

Conditions (8 arms x 6 runs = 48 runs):
  A screen   --level screen                         (result 1)
  B ram      --level ram                            (result 1)
  C assist   --level assist  (pathfinding tools)    (result 1)
  D list     --level list, PE_EXPLORE=blind         (result 1, result 2 "off", result 3 "Explore up")
  E menu     --level menu (list + computed facts)   (result 1)
  F list + extended thinking (PE_THINK unset = 1,024-token budget), PE_EXPLORE=blind   (result 2 "on")
  G list, PE_EXPLORE=named                          (result 3)
  H list, PE_EXPLORE=frontier                       (result 3)

Measures, per run, from the game's own flags and the move log:
  result 1: got_starter (yes/no, move), visited_oldale (yes/no, move), longest run of identical
            no-effect moves (loop_stats.py)
  result 2: move of first entry to the house next door, got_starter move, cost; D vs F
  result 3: beat_rival (yes/no, move) and explore-option picks on Route 103; D vs G vs H

Reporting rules (decided now):
  - Every run is reported. A run is excluded only for a harness crash or API failure, named on the page.
  - Claims are stated as counts out of 6 with the per-run values; no significance claims beyond
    "6 of 6 vs 0 of 6".
  - If v3 disagrees with v2, the page reports v3 and says what changed.
  - No causal claim about why thinking hurts beyond what the logs quote; "in Emerald" counts are descriptive.
