Option Scoring

Climpt's core tenets from CLAUDE.md:

Derived axes (2–3 per task)

Read the relevant design doc (docs/internal/*, agents/docs/*, or a user-cited constraint) and extract axes that matter for this decision. Examples:

• Prompt resolution change → edition/adaptation coverage, breakdown wrapper compatibility
• CLI option change → unified-help-concept fit, --help protocol stability
• Agent runner change → definition.parameters purity, verdict-type correctness

Every derived axis must cite a source line: from docs/internal/unified-help-concept.md §3. If no citation is possible, the axis is invented — drop it.

Phase 2: Enumerate Options

List options as Option A, Option B, … . Each option gets:

• Sketch (1 sentence): what it does
• Change surface: which files/modules it touches
• Cost signal: new code / removed code / migration shape

Do not pre-rank here — just lay them out at the same abstraction level. If one option is described as a design shape and another as an implementation detail, re-level before scoring.

Phase 3: Score Matrix (Quantified)

Scoring scale

Each axis is scored 0 / 1 / 2:

Weights

Each axis carries a weight w ∈ {1, 2}:

• w = 2 — critical / non-negotiable (a violation here likely sinks the option)
• w = 1 — supportive (a violation here costs quality but does not sink the option)

Default weights for the fixed spine:

Derived axes default to w = 1. Raise to w = 2 only when the cited doc explicitly marks the constraint as blocking (e.g. "MUST", "non-negotiable", "breaks the protocol").

Aggregation

For each option:

• total = Σ(score × weight) across all axes
• max = Σ(2 × weight) — same denominator for every option
• fit % = total / max

Disqualification rule

An option is DQ (disqualified) if any axis with w = 2 scores 0, regardless of total. Foundational violations cannot be outweighed by quality wins elsewhere. A DQ option may still be presented for context, but it cannot be the recommendation.

Example

Every ✗ and △ must have a one-line rationale directly below the table. Every weight that deviates from the default must also have a one-line rationale ("Reviewer precision raised to w=2 because PR will go through external auditor").

Phase 4: Trade-off Narrative

After the matrix, write:

1. Dominant axis — which axis actually drives the decision (often the highest-weight axis where options diverge)
2. What each option sacrifices — one sentence per option, citing the axes where it lost points
3. Recommendation — the pick (highest fit % among non-DQ options), and the single reason it wins
4. Reversal conditions — what new signal would flip the recommendation (typically: a weight change, or a new axis surfacing)

If the top two non-DQ options are within 5 percentage points, declare it a near-tie and name the tie-breaker axis explicitly — do not let the numeric ranking hide a judgment call.

Output Template

## Decision: <short title>

### Options
- Option A: <sketch>
- Option B: <sketch>
- Option C: <sketch>

### Axes (sources)
- Fixed: 全域性 / Core-first / No backward-compat / Fallback / Reviewer
- Derived: <axis> (<doc>§<section>)

### Matrix
<scored table including w column, weighted total, fit %, DQ?>

Rationale for ✗/△:
- A × Core-first (score 0): <reason>
- B × 全域性 (score 1): <reason>
- ...

Rationale for non-default weights (omit if all weights are default):
- <axis>: raised to w=2 because <reason>

### Trade-offs
- Option A sacrifices: …
- Option B sacrifices: …
- Option C sacrifices: …

### Recommendation
Option <X> — because <single reason tied to dominant axis>.
Reverse if: <signal>.

Anti-Patterns

• Invented axes: derived axes without a doc citation — drop them
• Score inflation: every option scores ✓ everywhere — axes are too coarse, split them
• Hidden options: sketching a 4th option mid-analysis — restart Phase 2 so all options are at parity
• Narrative without matrix: trade-off prose with no scored grid — scores must anchor the prose
• Dead-level compare: scoring Option A's implementation detail against Option B's design shape — re-level first
• Tie-avoidance: ranking without naming a dominant axis — ties come from missing the real driver, not from genuine equivalence
• Weight gaming: bumping a weight to make the preferred option win — weights must be set before scoring and justified against the cited doc, not adjusted to fit the conclusion
• DQ ignoring: recommending an option that violates a w=2 axis because its total is high — DQ is absolute, not advisory
• False precision: reporting fit % to decimal places — the underlying scale is 0/1/2; round to whole percent

Checklist

Phase 1: - [ ] Fixed spine listed
         - [ ] 2–3 derived axes cited with doc source
Phase 2: - [ ] All options sketched at the same level with change surface
Phase 3: - [ ] Weights set (default unless cited reason given) before scoring
         - [ ] Matrix filled; every ✗/△ has a one-line rationale
         - [ ] Weighted total, fit %, and DQ status computed for each option
Phase 4: - [ ] Dominant axis named
         - [ ] Recommendation = highest fit % among non-DQ options
         - [ ] Near-tie (≤5pp gap) flagged with explicit tie-breaker
         - [ ] Reversal condition stated

References

• CLAUDE.md — fixed spine source
• docs/internal/ — derived axes source
• docs-writing skill — if the decision produces a doc, apply 5-level framework after the pick
• simplify skill — if the chosen option still feels heavy, run simplify on it before committing