ase-meta-evaluate

@${CLAUDE_SKILL_DIR}/../../meta/ase-control.md @${CLAUDE_SKILL_DIR}/../../meta/ase-skill.md @${CLAUDE_SKILL_DIR}/../../meta/ase-getopt.md

<purpose name="ase-meta-evaluate"> Evaluate Alternatives </purpose> <expand name="getopt" arg1="ase-meta-evaluate"> $ARGUMENTS </expand> <objective> *Evaluate* *alternatives* through a weighted multi-*criteria* decision matrix. </objective> <flow> 1. <step id="STEP 1: Determine Reason"> - From the <request><getopt-arguments/></request>, try to derive the overall reason <reason/> for the evaluation. If no such reason can be derived, assume <reason>generic comparison</reason>.

-   Output the determined reason with just the following <template/>
    and do not output anything else:

    <template>
    <ase-tpl-bullet-secondary/> **REASON**: *<reason/>*
    </template>
</step>

2. <step id="STEP 2: Determine Alternatives"> - From the <request><getopt-arguments/></request> derive the two or more alternatives <alternative-K/> (K=1-N) the user wants to be evaluated. Do not output anything.

-   If fewer than two alternatives could be derived (N<2), output the
    following <template/> and *stop the entire flow* immediately without
    executing any further steps:

    <template>
    <ase-tpl-bullet-signal/> **ERROR: INSUFFICIENT ALTERNATIVES**: at least two are required for a comparison!
    </template>

-   For each alternative <alternative-K/> (K=1-N), decide whether
    you have all necessary information at hand to give it the proper,
    unique, short, and *concise name* <alternative-K/>. If you are
    unsure, use the `ase-meta-search` skill (at most one invocation per
    alternative, drawing from the *skill-wide shared budget* of at
    most 8 `ase-meta-search` invocations in total across STEP 2 and STEP 3
    combined) to find out more and adjust the name <alternative-K/>.

    If still unsure after the shared budget is exhausted, pick the
    best-guess concise name and proceed. Do not output anything.

-   For each alternative <alternative-K/> (K=1-N), decide which *one*
    of *USP* (Unique Selling Point -- what makes it unique), *Crux*
    (what you should notice), or *Gotcha* (what you should not stumble
    over) is its single most distinguishing perspective, and remember
    this as an <info-K/> (K=1-N) formatted like `<type/>: <hint/>` where
    <type/> is one of `USP`, `Crux`, or `Gotcha` and <hint/> is a 1-6
    word hint. Do not output anything.

-   For the set of alternatives, decide what the 1-6 word long
    name of the *class of alternatives* <class-of-alternatives/> is.
    Do not output anything.

-   For each alternative <alternative-K/> (K=1-N), decide whether
    it is a genuine member of <class-of-alternatives/>. If any
    <alternative-K/> is *not* a member (i.e. the alternatives are not
    mutually comparable within a single class), output the following
    <template/> (repeating the `⚑` line once per non-member
    alternative, i.e. once for each <alternative-K/> for which the
    membership decision was negative) and *stop the entire flow*
    immediately without executing any further steps:

    <template>
    <ase-tpl-bullet-signal/> **ERROR: INCOMPARABLE ALTERNATIVES**: *<class-of-alternatives/>*

    ⚑ **<alternative-K/>** (*member of a different class*)
    </template>

-   Output the determined, individual alternatives as a Markdown
    *table* with just the following <template/> and do not output
    anything else:

    <template>
    <ase-tpl-bullet-normal/> **ALTERNATIVES**: *<class-of-alternatives/>*

    | ⚑ *Alternative*        | ⚖ *Hint*  |
    | :--------------------- | :-------- |
    | ⚑ **<alternative-1/>** | <info-1/> |
    [...alternatives K=2-(N-1) for N>2...]
    | ⚑ **<alternative-N/>** | <info-N/> |
    </template>
</step>

3. <step id="STEP 3: Derive Criteria"> - From the <request><getopt-arguments/></request>, try to derive the criteria <criteria-L/> (L=1-M) for the evaluation. Do not output anything.

-   For each criterion <criteria-L/> (L=1-M), decide on its <weight-L/>
    from the value set { 4.00, 2.00, 1.00, 0.50, 0.25 } (from most
    important, via normal, to less important). Do not output anything.

-   Ensure the final number of criteria is always within the range of
    minimum 8 and maximum 12: if fewer than 8 criteria were requested,
    use the set of alternatives to decide on additional criteria
    which potentially best allow triaging the alternatives, take the
    <reason/> into account, and use the `ase-meta-search` skill (drawing from
    the *skill-wide shared budget* of at most 8 `ase-meta-search` invocations
    in total across STEP 2 and STEP 3 combined) to find out about the
    potentially still missing criteria and assign their <weight-L/>.

    If still under 8 criteria after the shared budget is exhausted,
    fill the remaining slots from existing knowledge without further
    searches; if more than 12 criteria were requested, drop the criteria
    with the smallest <weight-L/> until 12 remain. Do not output
    anything.

-   To prevent a single high-weight criterion from dominating the
    weighted sum (the weight set is geometric ×2 while the evaluation
    Likert scale is linear), assign weight 4.00 to *at least one* and
    *at most two* criteria, and weight 2.00 to *at least two* and *at
    most three* criteria. Symmetrically, to prevent a long tail of
    negligible-weight criteria, assign weight 0.50 to *at most two*
    criteria, and weight 0.25 to *at most one* criterion. Do not output
    anything.
</step>

4. <step id="STEP 4: Evaluate Alternatives against Criteria"> - For each alternative <alternative-K/> (K=1-N) and each criterion <criteria-L/> (L=1-M), decide on the evaluation <eval-K-L/>, which means how well the alternative meets the criterion on a Likert-scale from { -2, -1, 0, +1, +2 } (from worst, via neutral, to best). Do not output anything.

-   Then, calculate the ratings <rating-K/> (K=1-N) for all
    alternatives in a single call by invoking the `ase_decision_matrix(matrix:
    [ [ <weight-1/>, <eval-1-1/>, ..., <eval-N-1/> ], ..., [ <weight-M/>,
    <eval-1-M/>, ..., <eval-N-M/> ] ])` tool of the `ase` MCP server.
    The tool returns an array of N numerical values, where the K-th
    entry is the product-sum of all weights <weight-L/> (L=1-M) and
    the evaluation <eval-K-L/> (L=1-M) for alternative <alternative-K/>.
    Retain the *raw, unrounded* <rating-K/> for use in STEP 5, but
    round <rating-K/> to 2 decimal places *for display only* when
    emitting it in the table below. Do not output anything.

-   Output the resulting *Weighted Decision Matrix* as a Markdown
    *table* with just the following <template/> and do not output
    anything else. When emitting the table, render *one column per
    alternative* <alternative-K/> (K=1-N).

    <template>
    <ase-tpl-bullet-normal/> **EVALUATION**: *Weighted Multi-Criteria Decision Matrix*

    | ⦿ *Criteria*  | ⚖ *Weight*  | ⚑ **<alternative-1/>** | [...alternatives 2-(N-1)...] | ⚑ **<alternative-N/>** |
    | :------------ | ----------: | ---------------------: | ---------------------------: | ---------------------: |
    | <criteria-1/> | <weight-1/> | <eval-1-1/>            | [...evals 2-1..(N-1)-1...]   | <eval-N-1/>            |
    [...criteria L=2-(M-1) for M>2...]
    | <criteria-M/> | <weight-M/> | <eval-1-M/>            | [...evals 2-M..(N-1)-M...]   | <eval-N-M/>            |
    | **RATING**    |             | **<rating-1/>**        | [...ratings 2-(N-1)...]      | **<rating-N/>**        |
    </template>
</step>

5. <step id="STEP 5: Report Best Alternative"> - The best alternative <alternative-K/> (K=1-N) is the alternative whose raw, unrounded <rating-K/> (i.e. the product-sum from STEP 4, before the display-only rounding) is the maximum rating value across all alternatives. Do not output anything.

-   The second best alternative <alternative-X/> (X=1-N, X != K) is
    the alternative whose *raw, unrounded* <rating-X/> is the second
    largest rating value across all alternatives. Do not output anything.

-   If multiple alternatives share the second-largest raw rating, pick
    any one of them as <alternative-X/>; the resulting <distance/> and
    <percentage/> are unaffected by the choice, so the downstream output
    is deterministic. Do not output anything.

-   Determine rating distance <distance/> between <alternative-K/> and
    <alternative-X/> from their *raw, unrounded* ratings by calculating:
    <distance/> = <rating-K/> - <rating-X/>. Do not output anything.

-   Determine rating distance percentage <percentage/> between
    <alternative-K/> and <alternative-X/> from their *raw,
    unrounded* ratings as follows. If <rating-K/> is exactly zero,
    skip the division entirely (it is undefined against a zero
    anchor) and classify directly by <distance/> instead of routing a
    raw rating magnitude through the ratio thresholds below: if
    <distance/> is also exactly zero, both best and second best rate
    zero, which is a genuine tie, so set <percentage/> = 0 to fall
    into the *MULTIPLE BEST* branch below; otherwise <distance/> > 0
    means the second best rates strictly negative (since <rating-K/>
    is the maximum), so there is a clear positive gap to a worse
    runner-up and <alternative-K/> is an unambiguous winner, hence set
    <percentage/> = 1 (a sentinel exceeding all small-distance
    thresholds) to fall through to the plain *BEST ALTERNATIVE* branch.
    Do not output anything.

-   Otherwise (<rating-K/> is non-zero), calculate: <percentage/> =
    <distance/> / abs(<rating-K/>). Do not output anything.

-   By construction, <rating-K/> is the maximum rating across
    all alternatives, so <distance/> >= 0 always holds; using
    abs(<rating-K/>) keeps <percentage/> sign-stable across all rating
    regimes. Note that when <rating-K/> itself is negative, the
    denominator anchors to a poor best rating and small gaps can
    appear large; the all-negative regime is surfaced as a dedicated
    warning branch below. Do not output anything.

-   If <rating-K/> is less than zero (i.e. all alternatives
    rate negatively and the "best" is merely the least-bad), stop the
    flow after you output just the following <template/> and do not
    output anything else. This all-negative check is evaluated *before*
    the small-distance checks below, so that an all-negative alternative
    set with a small winning gap surfaces the all-negative warning rather
    than being masked by the multiple-best or small-distance branches:

    <template>
    <ase-tpl-bullet-signal/> **BEST ALTERNATIVE**: ⚑ **<alternative-K/>**
    ⚠ *ATTENTION: all alternatives rate negatively; this is the least-bad choice, not a strong winner!*
    </template>

-   Otherwise, if <percentage/> is less than 0.01 (i.e. <distance/> is
    effectively zero relative to abs(<rating-K/>)), stop the flow after
    you output just the following <template/> and do not output anything
    else:

    <template>
    <ase-tpl-bullet-signal/> **ERROR**: ✘ *MULTIPLE BEST ALTERNATIVES FOUND*,
    ⚠ *Please give some hints on the criteria to ensure a single best alternative!*
    </template>

-   Otherwise, if <percentage/> is less than 0.10, stop the flow after
    you output just the following <template/> and do not output anything
    else:

    <template>
    <ase-tpl-bullet-signal/> **BEST ALTERNATIVE**: ⚑ **<alternative-K/>**
    ⚠ *ATTENTION: small distance to second best alternative!*
    </template>

-   Otherwise (<percentage/> is greater than or equal to 0.10), output
    just the following <template/> and do not output anything else:

    <template>
    <ase-tpl-bullet-signal/> **BEST ALTERNATIVE**: ⚑ **<alternative-K/>**
    </template>
</step>

</flow>