AmplifierSkills and tools for thoughtful work

Skill

Cognitive Simulation

Explore a behavioral hypothesis before the study.

Frame a choice or learning task, record what a model predicts, and identify the observations needed to trust that prediction. Useful for preparing research questions and comparing explicit alternatives.

Use Cognitive SimulationWatch the 74-second filmWalk through an example
What you getFictional illustration
Fictional experiment: kiosk confirmation
Alternatives: confirm / review then confirm
Labels: counterbalanced across variants
Prediction: pending compatible endpoint
Probability: unavailable until returned
Human check: observed errors and task completion under a declared study

Fictional worked example. It illustrates the intended output; it is not a measured result or a real customer story.

Make the question testable

Name the population, task, alternatives and outcome, including the context the simulation leaves out.

Keep the prediction reproducible

Freeze instructions, labels, prior trials, backend identity and analysis before comparing variants.

Know what remains to learn

Separate model choice, available probabilities and actual participant evidence so a synthetic answer stays interpretable.

The film

A prediction is not a person.

In 74 seconds, Dowel, a wooden stand-in, takes a model’s answer for a customer: it picks option Q and a parser prints one hundred percent. Cognitive Simulation frames the choice first, freezes two prompts with the letters swapped so a label can’t decide, and reports probability as unavailable instead of one hundred. No model ran for this release, and accuracy about people is unmeasured. Captions are on the player; a transcript is below.

Download the film 15-second cut Captions
Read the transcript

You’re building a ticket kiosk. Dowel stands in for customers.

You want to know which flow people prefer: one tap, or review first.

A model picks option Q. The parser prints one hundred percent.

Cognitive Simulation treats that answer as a prediction, not a customer.

It frames the choice: who decides, what counts as success, what the task leaves out.

It freezes two prompts with the letters swapped, so a label can’t decide.

Both prompts pick option Q. The letter chose, not the flow. Probability: unavailable, not one hundred.

No model ran for this release, and accuracy about people is unmeasured. You still run the study.

On screen: A prediction is not a person.

Cognitive Simulation. Explore a behavioral hypothesis before the study.

The cinema kiosk and the experiment are a fictional example. The film shows how the skill behaves; it does not report results.

See the work take shape.

Fictional worked example. It illustrates the intended output; it is not a measured result or a real customer story.

Frame a bounded task

Translate the decision into an explicit choice or sequence and assess its fit to the available model.

Fictional example

A preference estimate without a probability

The brief A fictional kiosk team compares two confirmation flows; a backend returns Q with no log-probabilities.

Before

The parser reports 100% and the team calls that a customer preference rate.

After applying the method

Report choice Q, probability unavailable, exact prompt and backend; counterbalance labels and plan an observed task test before a product decision.

A generated choice is not certainty, a participant, or an observed population rate.

A clear job.
A useful stopping point.

Use when framing a behavioral simulation, interpreting Centaur output, or planning a comparison against real observations.

Use it when

  • You need to turn a behavioral idea into a precise experiment.
  • You want to compare alternative prompts or task formulations.
  • A simulation result needs calibration against observed participant data.

Choose another route when

  • Replacing people in a usability study or claiming customer preference rates.
  • Predicting what a named individual will do, or treating absent log-probabilities as certainty.

Frame

A behavioral question needs a falsifiable experiment.

A choice task, task-domain fit assessment, label counterbalance, and a human-validation gap.

Simulate

An authorized model endpoint and exact prompt are available.

An attributed model prediction with actual probability availability and a reproducible prompt.

Compare

Observed participant data exist.

A preregistered held-out comparison with calibration and domain limits.

Read the canonical method

Cognitive Simulation

Compare a behavioral hypothesis with a model prediction—without calling it a real participant.

Frame a behavioral question

Identify the decision, population, task, meaningful alternatives, outcome, and consequence. Translate the question into an explicit choice, learning, or sequential-decision task. Explain the fit to the model's training tasks and what social, emotional, organizational, or longitudinal context the task leaves out. If the intended inference needs actual people, keep that study in the plan; simulation does not replace it.

Construct the experiment

Preserve instructions, choice-key mapping, prior trials, responses, feedback, and the pending stimulus. Keep treatment differences isolated. Counterbalance option order and arbitrary response labels; label sensitivity is a model result, not human variation. Distinguish actual prior observations from invented/synthetic histories. Freeze the exact prompts and analysis rule before comparing variants.

Prompt and result contract documents the retained engine. Original experiment framing and Centaur reference preserve the domain patterns. Historical population or capability statements in those documents need verification against the actual model/version and cited study; they are not evaluation of this wrapper.

Simulate only with an available engine

The optional repository engine retains REST, CLI, and local-GPU backends. Inspect its configuration, the deployed model identity/revision, endpoint protocol, generation settings, and cost limits. Never silently substitute a general chat model and label the result Centaur. Record the exact prompt, completion, alternatives, and backend metadata. A mocked backend is a software check only.

If no endpoint is configured, return the experiment and blocked inference step. Do not invent a prediction. If log-probabilities are absent, report probability as unavailable; never convert a missing distribution into 1.0. Available next-token probabilities still are not calibrated population choice rates.

Compare and interpret

Report the model choice and available distribution as predictions. Evaluate sensitivity to wording/labels and task fit. For a claim about human behavior, compare held-out observed participant data under a declared sampling and analysis plan. Do not average repeated deterministic completions into a participant count. Separate model calibration, task-generalization evidence, and this adapter's software tests.

Output

Return scenario; task fit and omitted context; exact experiment/prompt; backend identity; prediction or blocked status; probability availability; sensitivity checks; what actual observations support; limitations; and the smallest participant-based check that could change the decision. Never say a synthetic preference establishes what a specific person or customer segment will do.

Limits and prerequisites
  • A model prediction is not real participant evidence, consent, prevalence, or clinical advice.
  • Centaur serving, licensed weights, compute, endpoint credentials, and domain validation are optional runtime concerns.
  • The portable method can frame and critique a simulation without any model deployment.

Your first useful run.

Bring the decision, intended population and alternatives you want to compare.

The download contains the complete skill folder. Extract it into the skills directory supported by your agent, keeping its references and license together. Or use the source installer below. Python 3.11+ is required for that installer.

Download the portable skillSource on GitHub

Start from the public source

git clone https://github.com/michaeljabbour/amplifier-skill-cognitive-simulation.git
cd amplifier-skill-cognitive-simulation
python3 scripts/capability.py doctor
python3 scripts/capability.py install --dest ~/.agents/skills
Native tools and optional execution
amplifier bundle add ./bundle.md

These commands are for the optional native surface. Follow its installation guide and prerequisites; portable method installation is separate.

Keep the complete skill folder together, including its references and license. The host applies the method in your context; the skill does not start a background service.

  • Requires Git and Python 3.11+ for the repository installer. The portable method itself can be read by a compatible agent without a service or provider account.
  • The destination is a skills-parent folder. Use the discovery path supported by your agent; ~/.agents/skills is an example, not a universal host setting.
  • Keep the complete installed folder, references and license together. Identical reinstall is safe; conflicts are refused. A deliberate --replace keeps a dated backup.

Try this with your agent

Use Cognitive Simulation to frame a fictional kiosk confirmation study. Compare a single confirm button with a review-then-confirm flow. State task fit, counterbalance response labels, prepare exact prompts and define the observed-user check. Do not run an endpoint.

Use after installing in a host that loads skill folders. Supply your own task files and review the result.

Keep the guide nearby

Read the guide · License · Changes · Release metadata · Install guide · Migration · Validation · Native guide · consolidation

Know what you can rely on.

164 mocked engine regression tests cover prompt/result and backend behavior, including missing probabilities. No model deployment or new inference is part of this release, and no human-behavior accuracy claim is established.

Read the checks and their scope
  • Centaur engine regression suite — passed

    164 tests. Mocked REST/CLI/backend protocol and prompt/result behavior; missing chosen-token probability is null, not 1.0. No Centaur model deployed or inference called.

  • Portable packaging and install — passed

    All operations passed. No runtime activation or effectiveness measured.

  • Skill discovery metadata — passed

    Skill is valid. This is a structural check, not a behavioral benchmark.

164 mocked engine regression tests cover prompt/result and backend behavior, including missing probabilities. No model deployment or new inference is part of this release, and no human-behavior accuracy claim is established.

Fictional examples explain the capability. Tests check the stated behavior. Outside-user outcomes need their own evidence.