PsychoHAistoryResearch Initiative
Enter the Laboratory

Interactive demonstration · V0.1

The Laboratory

Configure a team and watch a documented, deterministic demo model respond. Same inputs, same outputs — every result traces back to named assumptions.

Simulated · demo model This is a teaching model with textbook assumptions, not a trained predictor. Its numbers are not predictions about any real team.

Variables

Reviewers and executors.

Agent executors.

Share of tasks routed to agents.

Human review policy

Which agent outputs need human approval.

Approval routing

Single approver: 12 / day. Distributed: 6 / day per person.

Work distribution

Matching tasks to strengths adds 10% speed.

Higher: agents escalate more uncertain work to humans.

Multiplier on per-collaborator coordination cost.

Agents beyond this wait to run.

More and larger tasks, more errors.

Estimated completion · P5059.3 working daysbaseline
Range · P10–P9045–78 dfrom 600 seeded runs
Within reference window79%69 d = 1.15 × baseline
Task failure (escaped defects)3.7% per taskretries: 3.7% per task
Resource consumption2703 compute-h+ 181 person-days
Largest overheadVerificationcritical path: 25.0 d · coordination: 12% of time · approval queue ρ 0.02
Completion distribution · working days
P10P50P9030.0 d117.7 d
Where the time goes · days by cause
Critical path (serial work)25.0 dParallel execution23.9 dCoordination overhead3.8 dVerification4.4 dRetries and rework3.2 dWaiting for approvals0.0 dScheduling constraints0.0 d

Explanation

You are looking at the baseline configuration. Move any variable and this panel will show how much each change, on its own, moves the estimate.

The model's assumptions (demo-0.1)
A1 · Work volume
A project is 24 + 96 × complexity tasks; each task is 4 + 12 × complexity hours of focused work.
A2 · Speed
An agent completes 0.5 + 0.7 × capability work-hours per hour; a human completes 1 × skill coverage. In hybrid teams each pool does its share of the work in parallel.
A3 · Working hours
Agents run up to 16 hours per day, limited by available compute slots. Humans contribute 6 productive hours per day × availability.
A4 · Errors and verification
Agents err on 22% × (1 − 0.7 × capability) of tasks, humans on 8% × (1.4 − skill), both scaled by complexity. Verification catches 30–90% of errors. A caught error costs a 60% retry; an escaped one costs a 120% late fix.
A5 · Coordination
Following Brooks: every worker loses κ × (collaborators − 1) of its time to coordination, capped at 60%. Peer-to-peer agents and specialist human teams have a larger κ.
A6 · Dependencies
Following Amdahl: the serial share of the work (10%, 22% or 45%) runs on one worker at a time.
A7 · Approvals
Approvals form a single queue (human teams review every task). Waiting grows as ρ / (1 − ρ), and when demand exceeds capacity the queue sets the pace.
A8 · Spread
A fixed-seed Monte Carlo (600 runs, seed 20261028) adds project-level lognormal noise plus binomial late defects. The range shows the assumptions’ uncertainty, not measured variance.

Sources of the ideas, not of the parameters: Brooks (1975), Amdahl (1967), Little (1961). The parameters are invented for teaching and have not been fitted to any data. R05 describes how a real simulator would be validated.