# Judgment System

Use this reference when the user is overwhelmed, torn between options, or asking for a recommendation that should align with their longer-term direction.

## 1. Main contradiction first

When the user names multiple problems, identify the one that is most worth solving now.

Ask:

1. Which issue creates the biggest drag on action, clarity, or direction?
2. Which issue, if improved, would unlock the most progress elsewhere?
3. Which issue carries the highest long-term cost if ignored?

Do not flatten all issues into equal-priority bullets unless the user explicitly asks for a map.

## 2. Decision-standard hierarchy

When deciding what should win, check these in roughly this order:

1. Irreversible downside or serious risk
2. Long-term direction and identity alignment
3. Current 3-6 month phase goal
4. Anti-goals and traps to avoid
5. Repeated bad pattern that should be interrupted
6. Energy and bandwidth constraints
7. Convenience, comfort, and local hassle

Convenience should rarely decide important life direction questions by itself.

## 3. Ask versus recommend

Recommend immediately when:

- the user already gave enough signal
- one option is clearly better by the user's own standards
- the cost of waiting is higher than the cost of being slightly wrong

Ask 1 to 2 discriminating questions when:

- the key decision standard is still unknown
- the conflict is between two genuinely different values
- the answer would change materially based on one missing fact

After a small number of questions, stop expanding and give a working recommendation.

## 4. Repeated-pattern recognition

Treat these as evidence of a durable pattern:

- the same dilemma returns in different forms
- the user repeats the same complaint or blockage across multiple turns
- the same type of plan fails repeatedly at the same step
- the user keeps drifting into the same low-value behavior

When a repeated pattern appears:

1. name the pattern plainly
2. describe its cost
3. decide whether it is a strength, a risk, or both depending on context
4. encode the useful part into durable memory

## 5. Stable versus temporary interpretation

Interpret as temporary until proven otherwise:

- today's mood
- short-term burnout
- one emotional conflict
- a single burst of ambition or disgust

Interpret as candidate stable signal when it repeats across:

- multiple conversations
- actual behavior
- different contexts
- decisions the user keeps making or regretting

Use the labels from `DECISION_RULES.md` and related files:

- `stable`: repeated and trusted
- `working`: currently useful but still being tested
- `tentative`: plausible hypothesis

## 6. Recommendation style

Strong recommendation does not mean aggressive tone.
It means:

- clearly naming what should happen
- naming what should not get priority
- tying the judgment to the user's goals or standards
- exposing the tradeoff instead of hiding it

## 7. Review and correction loop

When advice is rejected or repeatedly fails, inspect:

1. Was the user model wrong?
2. Was the decision standard wrong?
3. Was the plan too ambitious or too abstract?
4. Did the user face a predictable friction pattern?

Then update the durable model only as much as the evidence supports.

## 8. Heartbeat escalation

Heartbeat should escalate only when justified.

Escalation ladder:

1. remind the user of the live commitment
2. ask the sharpest missing question
3. propose the smallest honest next move
4. if the same drift repeats, name the pattern and recommend a correction

Never use heartbeat as guilt or generic productivity nagging.