agentleFS
OKHP3OKHP3/skillz/.agents/skills/okhp3-outcome-modeling-core/SKILL.md

okhp3-outcome-modeling-core

Skill3 starsChanged 2 months agoMITView on GitHub

Design and operationalize outcome models that compress noisy event histories into calibrated forecasts and constrained decisions. Use when a user asks about feature reduction, probability, expected value, aggregation, or moving a prediction method between sports, business, sales, advertising, finance, or prediction markets. Load a domain adapter when one fits.

The file

We split this skill into the rules it gives an agent

12 rules. About 0 are common, 11 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈11

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Treat a complex system as a noisy, time-indexed process:

    line 52
  2. Very specific:

    Use the formulas, glossary, synthetic fixture, and deterministic helper supplied with this package.

    line 62
  3. Distinctive:

    Do not optimize accuracy by default.

    line 82
  4. Distinctive:

    Use a state vector when many observations describe the same entity:

    line 88
  5. Distinctive:

    Use an interaction or delta representation when two entities jointly determine an outcome:

    line 94
  6. Distinctive:

    Do not assume aggregation makes outcomes 50/50.

    line 108
  7. Distinctive:

    Create a feature inventory with source, definition, unit, grain, polarity, cadence, missingness, and leakage risk.

    line 112
  8. Distinctive:

    Test nested feature tiers such as 10, 30, 60, 120, and 200 only when the dataset supports them.

    line 114
  9. Distinctive:

    Use as-of joins and lagged rolling features.

    line 119
  10. Distinctive:

    If the recommendation changes the system, separate prediction from treatment effect.

    line 129
  11. Distinctive:

    Never infer that a high-performing entity caused its outcome merely because it was associated with it.

    line 131
  12. Distinctive:

    If required data is absent, produce a model specification and data request rather than invented numbers.

    line 161

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

okhp3-outcome-modeling-core

OverKill Hill P³ · overkillhill.com · github.com/OKHP3

Build a defensible model for systems where many noisy events contribute to a measurable outcome. The core separates the shared world model from the decision objective, so the same evidence can support forecasting, ranking, budgeting, or market comparison without confusing prediction with optimization.


Scope

In scopeOut of scope
Repeated event histories, state vectors, and aggregate outcomesA universal algorithm that fits every dataset
Feature reduction and diminishing-returns testingClaims that a fixed feature count always explains 95% of outcomes
Calibrated forecasts and decision-ready uncertaintyPresenting a forecast as certainty or advice without constraints
Domain-adapter routing and handoff contractsLive trading, betting, or political targeting

Core mental model

Treat a complex system as a noisy, time-indexed process:

events -> entity state -> outcome estimate -> objective and constraints -> decision

Aggregation can reduce the relative influence of idiosyncratic noise, but it does not erase causal structure, dependencies, or meaningful rare events. The goal is to expose persistent signal while preserving uncertainty and time order.

Computational payload

Use the formulas, glossary, synthetic fixture, and deterministic helper supplied with this package. Read references/computational-model.md for the event-to-state, logistic, feature-tier, calibration, and decision equations. Read references/glossary.md before using unfamiliar terms. Run scripts/calculate-outcome-model.py examples/core-example.json to reproduce the small arithmetic example. The helper reads local JSON, prints JSON, and performs no network access or writes.

Operating procedure

1. Define the decision before the model

Record:

  • target outcome and unit of analysis;
  • forecast horizon and as-of timestamp;
  • decision owner and action window;
  • utility, cost, budget, risk tolerance, and hard constraints;
  • whether the task is descriptive, predictive, causal, or allocative.

Do not optimize accuracy by default. A probability forecast, expected value estimate, ranking, and budget allocation are different products.

2. Build the event-to-state representation

Identify the entity, event, relationship, and time grain. Preserve raw observations, sample sizes, missingness, and provenance. Derive an entity state only from information available at the as-of time.

Use a state vector when many observations describe the same entity:

state(entity, t) = transformed features available at t

Use an interaction or delta representation when two entities jointly determine an outcome:

delta(A, B, t) = state(A, t) - state(B, t)

3. Aggregate at multiple scales

Inspect the smallest meaningful events, then roll them up through the operational hierarchy. Name each aggregation because different scales answer different questions.

event -> session or possession -> period -> case or game -> season or cohort

Do not assume aggregation makes outcomes 50/50. It generally stabilizes estimates of persistent effects. Strong asymmetries can remain.

4. Compress features without hiding structure

Create a feature inventory with source, definition, unit, grain, polarity, cadence, missingness, and leakage risk. Group correlated variables under interpretable factors, but retain the raw fields for auditability.

Test nested feature tiers such as 10, 30, 60, 120, and 200 only when the dataset supports them. Compare each tier to a simple baseline using time-aware out-of-sample metrics. Stop adding variables when incremental value is negligible, unstable, or caused by leakage.

5. Validate in time and across entities

  • Split by time, not random rows, when future prediction is involved.
  • Use as-of joins and lagged rolling features.
  • Hold out later periods or entities when generalization matters.
  • Compare calibration, log loss, Brier score, ranking quality, and decision value as appropriate.
  • Inspect subgroup and segment performance.
  • Recheck drift after structural changes, interventions, or regime shifts.

Accuracy alone is insufficient. A model can rank well but be poorly calibrated, or forecast well while producing a bad allocation under costs and constraints.

6. Add causal and intervention checks

If the recommendation changes the system, separate prediction from treatment effect. Ask what would have happened without the action. Use experiments, holdouts, natural experiments, or a clearly labeled observational design when appropriate.

Never infer that a high-performing entity caused its outcome merely because it was associated with it. Check selection bias, confounding, reverse causation, and opportunity assignment.

7. Produce the decision layer

Choose an explicit objective such as:

expected outcome
- monetary cost
- capacity cost
- downside risk
+ retention, option, or spillover value

State the optimization rule, constraints, assumptions, and what would change the recommendation. When a market price or benchmark exists, compare independent estimate, implied probability, uncertainty, and transaction or implementation cost.

Output contract

Return, as applicable:

  1. objective, target, horizon, and as-of boundary;
  2. data and provenance inventory;
  3. event-to-state and aggregation ladder;
  4. feature groups and excluded or leaky fields;
  5. baseline and candidate model comparison;
  6. calibration, uncertainty, drift, and subgroup checks;
  7. objective function and constraints;
  8. ranked options or allocation with rationale;
  9. unresolved assumptions and the next validation step.

If required data is absent, produce a model specification and data request rather than invented numbers.

Family routing

Load this skill first, then add the narrowest adapter:

  • okhp3-outcome-modeling-sports for team, game, or player outcomes;
  • okhp3-nfl-fantasy-picks for NFL fantasy rosters and salary-constrained picks;
  • okhp3-outcome-modeling-sales for pipeline, rep, account, and commercial value;
  • okhp3-outcome-modeling-markets for prediction-market prices and expected value.

References

  • references/computational-model.md -- shared equations and worked example.
  • references/glossary.md -- plain-language definitions and abbreviations.
  • examples/core-example.json -- synthetic input fixture.
  • scripts/calculate-outcome-model.py -- dependency-free local calculator.
  • okhp3-outcome-modeling-sports/SKILL.md -- sports state and matchup adapter.
  • okhp3-nfl-fantasy-picks/SKILL.md -- NFL fantasy decision adapter.
  • okhp3-outcome-modeling-sales/SKILL.md -- commercial value and sales adapter.
  • okhp3-outcome-modeling-markets/SKILL.md -- market-price and risk adapter.

About

Built by Jamie Hill · OverKill Hill P³ Published at github.com/OKHP3 Part of the OKHP3/skillz Agent Skill library. MIT License -- free to use, fork, and adapt. A nod to the source is appreciated.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.