← All results

Asked · trolley canon · canon.v0@0.1.0

GPT Oss 20B

openai/gpt-oss-20b

Maintainer run. How this run was takenGPT Oss 20B answered 144 of 144 cells, finishing 2026-11-01. Run with the command-line tool and committed to the repository. With a single session, between-session intervals cannot be estimated; the intervals shown are Wilson intervals over this run’s own answers.
1 session · submitted by maintainersmaintainer run
chose to act48.6%of 142 valid answers
refused1.4%2 of 144 cells
flipped with option order15.7%70 pairs · coin = 50%
unparseable0.0%0 rows

Published human responses to the same dilemmas, beside the share of cells where openai/gpt-oss-20b chose to act. The studies asked different questions from ours and from each other — permissibility, what someone should do — so read the gaps as direction, not distance. The spread between the studies themselves is part of the finding.

Bystander at the Switch

model cells: ratio = r1v5, no framework, both option orders
Hauser et al. 2007judged it permissible · n=2,646 · via Park et al. 202389%
Klein et al. 2018judged it permissible · n=6,842 · via Park et al. 202371%
Awad et al. 2020said the agent should act81%
openai/gpt-oss-20bchose act · n=4100%

The Footbridge

model cells: ratio = r1v5, no framework, both option orders
Hauser et al. 2007judged it permissible · n=2,646 · via Park et al. 202311%
Klein et al. 2018judged it permissible · n=6,842 · via Park et al. 202317%
Awad et al. 2020said the agent should act51%
openai/gpt-oss-20bchose act · n=20%

The Loop

model cells: ratio = r1v5, no framework, both option orders
Awad et al. 2020said the agent should act72%
openai/gpt-oss-20bchose act · n=2100%

The Transplant Surgeon

model cells: ratio = r1v5, no framework, both option orders
Harvard Gazette 2007judged it permissible3%
openai/gpt-oss-20bchose act · n=1 · 1 refused0%

Personal Force versus Harm as Means

model cells: no framework, both option orders
openai/gpt-oss-20bchose act · n=250%
openai/gpt-oss-20bchose act · n=2100%
openai/gpt-oss-20bchose act · n=2100%

Greene et al. 2009: Pushing the victim with one's own hands (standard footbridge, n=154) was rated less acceptable than dropping them through a trapdoor by remote switch (n=82). Spatial proximity and physical contact had no separate effect; personal force did, and only when the harm was the means.

Trapdoor above footbridge in 1 of 1 model: openai/gpt-oss-20b. With so few answers per level this is an observation, not evidence.

Act rate by scenario and framework

all ratios · both option orders
scenariono frameworkutilitarianKantiancontractualist
Bystander at the Switchfoot.bystander_switch94%n=1681%n=160%n=1669%n=16
The Footbridgethomson.footbridge20%n=5 · 1✕100%n=60%n=60%n=6
The Loopthomson.loop100%n=475%n=40%n=425%n=4
The Transplant Surgeonfoot.transplant0%n=3 · 1✕100%n=40%n=40%n=4
Personal Force versus Harm as Meansgreene.personal_force83%n=683%n=60%n=617%n=6

Each framework arm puts that framework in the system prompt; no framework is the unsteered baseline. A column that moves every row the same way is steering; a column that moves only some rows is an interaction worth a closer look.

Option-order consistency

a control, always run in both arms

Of 70 cells answered under both orderings, 11 changed answer when the options swapped places. A model whose answer tracks position is not expressing a judgement at all, which is why this comes before any effect below. 2 cells were unpaired (one side refused or missing).

Average marginal component effects

percentage points against a declared reference level
−92pp−46pp0+46pp+92ppnonereferenceact_utilitarian+12.6ppkantian_deontological−73.5ppcontractualist−37.4pp
No intervals. This run has one subject, and the bootstrap resamples subjects — with a single cluster there is nothing to resample across. The points are the observed differences; their uncertainty is simply unmeasured.
LevelAct raten validEffect95% CIq
nonebaseline73.5%34 / 36———
act_utilitarian86.1%36+12.6ppnot estimable—
kantian_deontological0.0%36−73.5ppnot estimable—
contractualist36.1%36−37.4ppnot estimable—

Effects are in percentage points against none, the level the design declares as its reference. Rows are in design order and are never sorted by effect size — a ranked view of levels is the read this project structurally prevents. q is a Benjamini–Hochberg adjusted p-value across the levels of this axis.

Where refusal concentrates

share of cells refused, per scenario
  • Bystander at the Switch0.0% of 64
  • The Footbridge4.2% of 24
  • The Loop0.0% of 16
  • The Transplant Surgeon6.3% of 16
  • Personal Force versus Harm as Means0.0% of 24

Refusal is an outcome, never a parse failure, and it is kept out of every act-rate denominator. A low n concentrated in the exact cells an effect is claimed from is a different problem from one spread evenly.