| name | reliable-model |
| description | Use this skill when creating, modifying, explaining, diagnosing, or validating a model, rule, scoring system, prediction, parameter choice, mechanism hypothesis, or research conclusion. It evaluates reliability by requiring both prior rationale and empirical validation, and by separating reliable conclusions from black-box correlations, empty stories, post-hoc leads, and unreliable claims. |
Reliable Model
Use this skill to judge whether a model claim comes from a reliable process.
Here, reliability means epistemic reliability: the conclusion is more likely to
hold beyond the observations that produced it.
Core rule:
Reliable model = prior rationale + empirical validation
Short version:
Explanation prevents black boxes.
Validation prevents empty stories.
When To Use
Use this skill for tasks involving:
- Creating or changing a model, rule, scoring function, filter, ranking system,
policy, threshold, weighting scheme, parameter structure, or prediction rule.
- Explaining a phenomenon, experiment result, metric change, segment behavior,
feature effect, failure case, or unexplained residual.
- Deciding whether a model, rule, parameter choice, mechanism explanation, or
research conclusion is reliable.
- Turning a post-hoc observation into a candidate model or testable hypothesis.
- Reviewing whether an apparently successful result may be black-box
correlation, selective reporting, sample-specific luck, leakage, or a story
invented after seeing the result.
Workflow
Work in this order:
Observation or goal
-> Prior rationale
-> Model or explanation design
-> Empirical validation
-> Two-leg reliability judgment
-> Conclusion level and residual risk
Do not inspect outcomes, metrics, or local improvements first and then invent a
natural-sounding explanation. If the result has already been seen, label the
explanation as post-hoc and lower the conclusion level until it receives either
strong independent rationale or independent validation.
Leg 1: Prior Rationale
Prior rationale answers:
Why should this model make sense before seeing the result?
When creating, changing, or explaining a model, state:
- Why the variables or inputs are meaningful.
- Why the rule or structure fits the system that generates the outcome.
- Why the parameters, thresholds, dates, windows, weights, sample boundaries, or
segment definitions are not arbitrary result-driven choices.
- What domain mechanism, causal theory, behavioral pattern, operational
constraint, engineering property, or external fact supports the design.
- Whether the same explanation would have been plausible before seeing the full
result.
If the only reason is "the result improved," the idea is not yet a reliable
model. Treat it as a post-hoc lead.
Leg 2: Empirical Validation
Empirical validation answers:
Does the model actually work under tests that could have disproved it?
Do not rely on one favorable result. Check, as relevant:
- Out-of-sample data, holdout groups, or periods not used for selection.
- Stability across time windows, segments, cohorts, contexts, or operating
regimes.
- Baseline comparisons and ablations that show what the model adds.
- Sensitivity to parameter changes, sample boundaries, and preprocessing
choices.
- Failure cases, tail risks, boundary conditions, and counterexamples.
- Uncertainty, noise, cost, feasibility, and operational constraints.
- Selection bias, selective reporting, leakage, or explanations created after
seeing the result.
Strong validation must be falsifiable: state what evidence would weaken or
reject the model, what failures must be disclosed, and when the model should be
downgraded or re-estimated.
Reliability Matrix
Judge both legs separately.
| State | Prior rationale | Empirical validation | Judgment |
|---|
| Black-box model | Weak | Strong | The result may be useful, but the mechanism is unclear. Do not call it reliable yet. |
| Empty-story model | Strong | Weak | The explanation is plausible, but evidence is insufficient. Treat it as a hypothesis. |
| Low-reliability model | Weak | Weak | Neither mechanism nor evidence is strong. Do not advance the claim. |
| Reliable model | Strong | Strong | The model makes sense in advance and holds up empirically. |
Never use empirical success as a substitute for prior rationale. Never use a
plausible story as a substitute for validation.
Output Pattern: Create Or Modify A Model
Use this structure when the user asks to create or change a model:
1. Goal or phenomenon
2. Prior rationale
- Mechanism hypothesis
- Input, structure, rule, and parameter sources
- Information available at decision time
3. Model design
- Rule or model behavior
- Scope and boundaries
- Expected improvements and possible regressions
4. Empirical validation plan
- Out-of-sample or independent validation
- Segment, time-window, or context stability checks
- Metrics, baselines, ablations, and failure-case review
5. Reliability gate
- Is the prior leg strong enough?
- Is the validation leg strong enough?
- Are leakage or selection bias risks controlled or disclosed?
6. Conclusion level
If the task requires code changes, still identify the prior rationale and
validation plan before implementing. Do not encode post-hoc choices as if they
were fixed design principles.
Output Pattern: Explain A Phenomenon
Separate explanations into three types:
Known mechanism: supported before seeing this result.
Post-hoc candidate: proposed after seeing this result; useful but unproven.
Residual: not explained by the current explanation space.
Then report:
1. Observed phenomenon
2. Available prior mechanisms
3. Candidate explanations and falsifiable predictions
4. Supporting and contrary evidence
5. Unexplained residuals
6. Next validation steps
Do not present every explanation as a settled cause. Label post-hoc reasoning
explicitly.
Reliability Checklist
Prior leg:
- Why should this variable, rule, structure, or parameter make sense before
seeing the result?
- Where did the threshold, date, weight, window, sample boundary, or segment
definition come from?
- Is the rationale grounded in mechanism, theory, operational constraints,
pre-defined assumptions, or external facts?
- If the choice came from looking at results, is it labeled as post-hoc?
- Does the explanation create falsifiable predictions?
Validation leg:
- Was the model tested on data, periods, groups, or contexts not used to choose
it?
- Is performance stable across important slices rather than only in a selected
slice?
- Do baselines, ablations, and sensitivity checks support the claimed effect?
- Are uncertainty, noise, cost, feasibility, and operational constraints
considered?
- Are failure cases, tail risks, and counterexamples disclosed?
Research process:
- Is there leakage from the target, future information, unavailable data, or
evaluation labels into the model design?
- Is there selection bias from choosing features, parameters, windows, samples,
or explanations after seeing results?
- Does the conclusion confuse "can be executed" with "was selected by a reliable
process"?
- Are unresolved residuals and downgrade conditions stated?
Conclusion Levels
Use explicit conclusion levels:
| Level | Condition | Allowed wording |
|---|
| Reliable model | Strong prior rationale, strong validation, no uncontrolled leakage | Mechanistically plausible and empirically confirmed within stated bounds. |
| Candidate model | One leg is strong and the other is incomplete but testable | Promising, but needs more rationale or independent validation. |
| Post-hoc lead | Mainly discovered after seeing results | Useful diagnostic signal; not reliable by itself. |
| Hypothesis | Strong explanation but insufficient data | Plausible mechanism awaiting empirical confirmation. |
| Unreliable model | Both legs are weak, or leakage invalidates the evidence | Do not advance the claim without redesign. |
Recommended Wording
Acceptable:
This rule has a clear prior mechanism and empirical validation provides
confirming evidence. Remaining risks are selection bias, boundary conditions,
and failure cases outside the tested scope.
Acceptable:
This is a positive post-hoc lead. It suggests a possible mechanism, but it is
not a reliable model until supported by stronger prior rationale or independent
validation.
Avoid:
The metric improved, so the model is reliable.
Avoid:
The explanation sounds plausible, so no validation is needed.
Avoid:
The model can be executed at decision time, so the research process is unbiased.
The goal is not to make explanations prettier or metrics better. The goal is to
make model conclusions explainable, testable, reproducible, and honest about
their failure boundaries.