How It Works

The Computation Pipeline

Every estimate follows this pipeline:

1. Base Rounds (from complexity lookup)
   × Risk Coefficient
   × Domain Familiarity
   = Adjusted Agent Rounds

2. Adjusted Rounds × Minutes per Round (from maturity)
   = Agent Time

3. Agent Time × Integration Overhead %
   = Integration Time

4. Agent Time × Adjusted Fix Ratio (agent-effectiveness-adjusted)
   = Human Fix Time

5. Review Minutes (from depth × complexity lookup)
   = Human Review Time

6. Planning Minutes (from complexity lookup)
   = Human Planning Time

7. Human times × Org Size Overhead
   = Adjusted Human Times

8. Sum of all times = Subtotal

9. Subtotal × Task Type Multiplier = Total

10. Apply Cone of Uncertainty Spread
    (widens min/max range based on definition phase)

11. PERT Expected = (min + 4×midpoint + max) / 6
    SD = (max - min) / 6

12. Total × Confidence Multiplier = Committed Estimate

13. Multi-agent/multi-human adjustments (if applicable)

14-15. Token & Cost Estimation
      Rounds × tokens-per-round × num_agents = total tokens
      Split into input/output by complexity ratio
      Optional cost at model tier pricing

What "Agent Rounds" Means

A round is one agent invocation cycle:

Prompt → Reasoning → Code/Output → Validation

One round might produce a function, fix a bug, or write a test. Complex tasks require many rounds of iteration.

Complexity	Rounds	Example
S	3-8	Add a 404 page, fix a typo in API response
M	8-20	Stripe integration, REST API endpoint
L	20-50	Database migration, test suite setup
XL	50-120	Architecture rewrite, multi-system refactor

Two Output Numbers

Every estimate produces two numbers:

• Expected — The PERT-weighted most likely outcome (50% chance of finishing sooner, 50% later)
• Committed — The risk-adjusted estimate at your chosen confidence level (80% or 90%)

These serve different audiences:

• Expected is for internal planning ("how much work is this?")
• Committed is for external promises ("when will it be done?")

Where the Numbers Come From

Every number in the output traces back to:

1. A lookup table (base rounds, review minutes, planning minutes)
2. A multiplier from a questionnaire answer (risk, domain, org size, task type)
3. A formula defined in formulas.md

Nothing is a guess. If an estimate seems wrong, you can trace it back to a specific input and adjust.