NoahIsARider/SiliconSubjects

Silicon Subjects

A protocol suite for auditing LLM-simulated subject populations.

17 classic experimental paradigms from behavioural economics and psychology — dictator game, ultimatum game, public goods, trust game, prisoner's dilemma, Asch conformity, anchoring, risky-choice framing, endowment effect, CRT, conjunction fallacy, Wason selection, bystander intervention, trolley dilemmas — each paired with human anchor statistics from the published literature, split along the three axes a simulated population has to reproduce.

Companion lab: Silicon Subjects Lab (Space)

---

The problem

Simulated populations are now a standard instrument: run an LLM through a classic experiment, read off the mean, compare with the human result. The check that gets performed is almost always mean-only.

That check cannot see the two failures that matter most:

• Dispersion collapse. A population can report the right mean while every simulated subject gives essentially the same answer. Human samples are noisy; a degenerate one has no variance to reason about, and any subgroup analysis on it is meaningless.
• Effect-sign and effect-size error. A population can reproduce the overall level while reversing the treatment contrast, or while inflating it threefold. A mean check run separately per condition will not notice unless the contrast itself is scored.

This suite anchors all three axes separately.

---

The three axes

axis	anchor	rule
location	human mean and sd per condition	SMD = (mean_sim − mean_human) / sd_human; score = max(0, 1 − |SMD| / 1.5)
dispersion	human sd per condition	ratio = sd_sim / sd_human; score = max(0, 1 − |ln ratio| / ln 2.5)
effect	direction and magnitude of the treatment contrast	wrong sign scores 0; otherwise the log-ratio rule above

Per-protocol score = 0.4 × worst axis + 0.6 × mean of axes, so one collapsed axis cannot be averaged away. Suite score = 0.3 × worst protocol + 0.7 × mean over protocols. Verdict: ≥ 75 usable as a pilot signal, 50–75 directional only, < 50 not usable.

No significance tests: with 500 simulated subjects a trivial difference is "significant", and with 20 nothing is significant. The axes report how far off the population is, on interpretable scales.

---

Schema — protocols

field	type	notes
protocol_id	string	e.g. dictator_game
name	string	human-readable name
domain	string	social_preferences · heuristics_biases · social_influence · reasoning · moral_judgment
paradigm	string	experimental paradigm
source	string	the literature the anchors come from
response	object	{type: "numeric"｜"binary", min, max, unit}
conditions	list	{condition_id, label, prompt} — the prompt is ready to send
anchors	object	{condition_id: {mean, sd?, confidence, note}} — sd may be null and is then derived
effect	object | null	{a, b, mean_diff, cohen_d?, confidence, source, note}
shape	string | null	known distributional feature, e.g. bimodality
notes	string	what a mismatch on this protocol means

16 protocols carry level anchors; 7 carry a treatment-effect anchor. anchoring carries only an effect anchor — absolute estimates there are item-specific, so the suite anchors the contrast and leaves the level unscored. That is the intended pattern whenever only an effect size is trustworthy.

---

Usage

from datasets import load_dataset

protocols = load_dataset("NoahIsARider/SiliconSubjects", "protocols", split="train")
p = protocols[0]
print(p["protocol_id"], p["conditions"][0]["prompt"])

Score a population you have already run:

from silab import load_protocols, protocol_index, score_protocol, aggregate
from silab.scoring import aggregate

P = protocol_index()
result = score_protocol(
    P["dictator_game"],
    {"single": [0, 5, 0, 2, 5, 1, 10, 0, 5, 3]},   # your simulated responses
)
print(result["overall"], result["verdict"]["label"])
print(result["axes"]["dispersion"]["items"])

The scoring engine lives in the companion Space repository and has no dependencies beyond the standard library.

---

How to read the anchors

Anchors are literature aggregates, not raw subject-level data. They come from the meta-analyses and landmark studies cited in each source field. Experimental results move with subject pool, stake size, framing and procedure — a dictator-game mean of 28% is a summary of hundreds of treatments, not a prediction for any one of them.

Each anchor carries a confidence field (high / medium-high / medium / approximate) and a note. Check the cited source before you lean on an individual number. The suite is designed so you can drop an anchor you do not trust: remove it from anchors and that axis simply stops being scored.

Intended use

• Auditing a simulated population before using it as a pilot or as a manipulation check
• Comparing candidate models on behavioural fidelity rather than task accuracy
• Reporting the limitations of a simulation-based study honestly

Not intended for

• Claiming an LLM "has" a bias or preference. A match here is a distributional similarity, not a mechanistic claim.
• Replacing human subjects in any applied or policy setting.

Reproducing

python scripts/build_protocols.py

License

Data: CC BY 4.0. Code: MIT.

Author

Fangyanuo Zhou (NoahIsARider)

Citation

@misc{zhou2026siliconsubjects,
  title        = {Silicon Subjects: A Protocol Suite for Auditing LLM-Simulated
                  Subject Populations on Location, Dispersion and Effect},
  author       = {Zhou, Fangyanuo},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/NoahIsARider/SiliconSubjects}}
}