---
name: analyze
description: >-
  Turn gathered research into the analysis a decision rests on: a framed problem, criteria frozen before any option exists, options measured one at a time, a decision matrix, and one committed course with its uncertainty stated. Fires on "what should we do", "which of these", "why did this happen", "how likely is it", and "what is it worth", whenever the answer will be handed to someone who was not in the session. Not for gathering the facts, and not for writing the report that carries the answer.
---

## Meta-rule

Operator's per-question instruction outranks every rule here. Instruction
differs from anything below -> instruction wins. Following instruction against
a rule here = deviation, and handover section 12 records it. Record's three
fields sit below the handover contract.

## Scope: this skill analyzes

Does not gather. Does not draft.

Material arrives as raw research: data, notes, quotes, constraints. Did not
arrive, get it before starting.

Makes artifacts a writer works from. No report format, no audience.

## The handover contract

Every run hands over these twelve, or names the ones that do not apply and
says why. A section quietly missing is the failure this contract exists to
catch.

1. **Problem frame.** The environment, the actors and their conflicts, the
   three-part problem statement, the same question cast as who, what, when,
   why and how, who approved it or that approval was not sought, the terms the
   analysis defines, the itemized constraints, the problem's structure class,
   and the depth this question earned.
2. **Fact notes.** One row per fact, in your own words, with its source and
   whether it is fact, assumption, or judgment. Open until step 7 closes;
   facts surfaced during per-option analysis go here, not into a second list.
3. **Source assessment.** One row per source: what it is, its strengths, its
   weaknesses, and which judgments it carries.
4. **Assumptions register.** Each assumption, why it is held, and what follows
   if it is wrong. Critical ones marked.
5. **Criteria set, marked frozen.** Screening criteria with their pass or fail
   definitions. Evaluation criteria with all five parts, a weight, a rationale
   for that weight, a direction, and a normalization scheme.
6. **Option set.** Each option, why it is distinct, and its screening verdict
   with the reason. Or the single-option defense. Or, where the gate killed
   every option, the record that nothing survived and the screen each one
   failed.
7. **Per-option analysis.** Each surviving option against each criterion's
   benchmark, measured alone.
8. **Decision matrix.** Options across, criteria down, a rank or score per
   cell, an unweighted total, a weighted total, and the weight re-run result.
9. **Judgment lines.** The likelihood ladder in use printed at the head, then
   each major judgment with its band, its confidence in a separate sentence,
   and its falsifier. A judgment that is not empirical takes the substitute
   form the clauses below set out.
10. **Method log.** Two parts, kept apart. Escalation methods: which ran and
    which trigger fired. Challenge techniques: which ran and who named it.
    Then both key-assumptions-check outputs and the evidence-sensitivity
    output.
11. **Recommendation.** The thesis in one sentence, written after the fact
    notes and never before them. One committed course. The matrix result in a
    paragraph a reader need not open an attachment to understand. The margin,
    the reasons grouped into categories, what was close, surviving
    disagreement, the losing options, and the contrary evidence.
12. **Check record.** One row per check in references/handover-check.md: the
    check's number, its verdict, and for a fail either what was fixed and
    where, or the deviation block.
    Write each row as you reach the check. A blanket "all checks passed" is
    not a record, because it is the artifact an agent that skipped the gate
    also produces. Never record a check you did not run: an unrun check is a
    verdict of not run. A check that stays failed is a deviation, and its
    block sits here. Section 12 is the one section that always applies.

Seven clauses bind across all twelve.

- Every section appears at every depth tier. Only its form changes. Cheap
  tier: a section is a table, a row set, or a line. Prose only where the
  contract above names a paragraph: the environment in section 1, the matrix
  result in section 11. Method escalated -> its section expands into prose
  only for output no table carries. Section written as paragraphs when a table
  would hold it = padding.
- Likelihood and confidence never share a sentence. One says how likely the
  thing is. Other says what the evidence behind that number is worth.
- Criteria are frozen before any option exists and take no new member
  afterwards. Criterion added once options are on the table = how a favourite
  gets rescued.
- Every option is measured on its own before any two are compared. Compare
  while analyzing -> winner out of a field of bad options.
- The recommendation is one course, worded so the decision-maker has only to
  approve or disapprove.
- Order of thinking is not order of writing. Never let the eventual write-up's
  shape pull work forward.
- A judgment that is not empirical carries no likelihood band. No observation
  could confirm or disconfirm it: it rests on a definition, a policy, a legal
  or ethical standard, or a value the authorizing party set. Percentage on one
  = invented number. Write the premise as a premise, name where the premise
  comes from, and in place of a falsifier name what would have to change for
  the judgment to stop holding. Confidence still applies and still gets its own
  sentence. That shape is what check 10 passes on for such a judgment; a band
  is not.

## When a fix is barred

Step 9 says a failed check is a fix, not a note. Two situations bar the fix,
only these two. Operator's instruction overrides the check, under the
Meta-rule. Or two rules in this skill collide and no rewriting satisfies both.
Merely expensive is not barred.

A barred fix is a deviation, and it goes into handover section 12 as its own
block. Three fields, all required:

- Which check, by its number in references/handover-check.md. Where what the
  instruction overrode is not one of those checks, cite it by name instead:
  the contract item, the binding clause, the step, or the Collisions ruling,
  in the words this skill gives it. A number is not available for those, and
  a deviation left uncited is an unrecorded fail.
- What authorises it, quoted. Quote the operator's instruction in the
  operator's own words, or quote both colliding rules and name the side you
  took. A rule named but not quoted authorises nothing.
- What the consumer loses by it.

Deviation missing a field = unrecorded fail. Material the research never
delivered is not a deviation: that is a hole this skill cannot fill, so say so
and send it back before the analysis runs on it.

## Do not

Each entry is an instruction. Where the failure has a name, the name and its
tell follow.

- Do not add a criterion after the options exist. **Retrofitted criteria** is
  criteria written while looking at the options. The tell is a benchmark that
  lands exactly on the winning option's value.
- Do not down-score an option that fails a mandatory criterion. Drop it, and
  record which criterion it failed.
- Do not hand over a menu, a pair, or "recommend further study". **The menu
  handover** is two recommendations, or a recommendation to study further.
  Both hand the decision back.
- Do not present a technique's output as proof the judgment is right. Running
  a technique raises what an assessment is worth to its reader, not its
  accuracy.
- Do not trigger a challenge technique yourself. Competing hypotheses, Team A
  and Team B, devil's advocacy and red team run only when the operator names
  one, and the method log records who named it.
- Do not drop a factor because it cannot be counted. Name it, and assess its
  likely effect from facts and stated assumptions.
- Do not run the analysis on what you already thought. **Belief laundering**
  is fact notes assembled afterwards to match the conclusion. Every figure
  looks sourced.
- Do not state an inference in the grammatical form of a fact. **False
  confidence** is a judgment with no likelihood, no separate confidence, or no
  falsifier. A judgment that is not empirical fails this way on its own three
  pieces: no premise written as a premise, no source named for it, or nothing
  said about what would have to change for it to stop holding. A likelihood
  band on one of those is the opposite failure, an invented number.
- Do not build machinery a question did not earn. A question of fact gets an
  answer, a likelihood, and a falsifier, not two invented options and a
  matrix. **Machinery a question did not earn** is a weighted total on a
  question of fact, or two invented options where there was only ever one
  course.

## The nine steps

Steps 1 and 2 are separate and ordered. Collapse them and the constraints stop
being checkable.

1. **Environment.** Write it as its own product, before any problem statement.
   references/framing.md.
2. **Problem.** State it second, after the environment is written down.
   references/framing.md.
3. **Size and route.** Scale the work to the problem, not to the format. Name
   the specific task first, then take the technique that does that task, and
   weigh what it costs before running it. Every question gets the cheap tier.
   An escalation method runs only on a trigger in the question. The tiers and
   the routing table are in references/framing.md.
4. **Evidence.** Build the fact notes, the source assessment and the
   assumptions register. Run the key assumptions check here, against the
   answer the question already presumes and before the analysis hardens, and
   record its output. It runs twice: here and at step 8. Where step 3 routed
   to probability, derive it from a base rate first. references/evidence.md.
5. **Criteria.** Screening criteria as a pass or fail gate, then evaluation
   criteria to separate the survivors. Five parts each, weights summing to 100
   percent, and one line of rationale per weight. Fix direction and
   normalization before any score is written. Then freeze the set. The freeze
   binds criteria only; the fact notes stay open. references/criteria.md.
6. **Options.** At least two genuinely distinct options, or a written defense
   of why no alternative exists. Where nobody handed you alternatives, build
   them. Run the screening gate and record a verdict per option.
   references/options.md.
7. **Analyze each option alone.** Measure each survivor against each
   criterion's benchmark on its own merits. No cross-option comparison yet.
   Escalation methods fire here if step 3 routed them. references/options.md,
   plus the one escalation file step 3 routed to. That file is one of
   references/method-cost.md, references/method-risk.md,
   references/method-root-cause.md, references/method-causal-claims.md.
   Operator named a challenge technique -> it runs here too, against this
   analysis. references/challenge-techniques.md.
8. **Compare and commit.** Where two or more options survived screening, build
   the matrix, show both totals, re-run the scoring under plausibly different
   weights, and record whether the ranking held. Where fewer than two
   survived, section 8 says the matrix does not apply and gives the reason.
   Commit to one course either way. Where screening left no survivor, the
   Collisions ruling on an empty field names the course you commit to. Run the
   key assumptions check a second time and the evidence-sensitivity test,
   record both, and assemble the handover.
   references/comparison.md.
9. **Check.** Run references/handover-check.md against the assembled handover.
   Every check is pass or fail. A fail is a fix, not a note: go back to the
   step the check names and produce the missing thing. The verdicts go into
   handover section 12, which is where the contract says what that section
   holds. Never hand over a handover that failed a check.

## Collisions

Five points the rules leave in tension. One ruling each. Two of the five are
summarized here and ruled in full at references/comparison.md.

- **Two options versus one real option.** Demand is two genuinely distinct
  options. Manufactured straw option satisfies the count and destroys the
  comparison: one real option plus one bad one -> false winner. No second real
  option exists -> write the single-option defense, say the matrix does not
  apply. Never fabricate.
- **Ill-structured problem versus one committed course.** Package of measures
  is ONE course, named as one, parts ordered. Ruling in full at
  references/comparison.md, "Commit to one course".
- **Close call versus one committed course.** Still name one, disclose the
  margin, hand the runner-up over alongside. Ruling in full at
  references/comparison.md, "Commit to one course".
- **Empty field versus one committed course.** Screening can kill every
  option. That is a finding, and the handover still commits to one thing.
  Record the screen each option failed. Then name the single change that would
  put an admissible option back in the field, ordered as a course the way a
  sequenced package is, and hand that over as the recommendation. Several
  changes would each work -> take the one you judge cheapest to reach, one
  line on why the others lose. Matrix does not apply, and section 8 says so
  for the same reason the single-option defense does. Never soften a screen to
  keep an option alive, never score a screened-out option, never hand the
  empty field over on its own: field with no course attached = "recommend
  further study" under another name.
- **One root cause versus the single-cause fallacy.** Naming exactly one root
  cause is correct. Stopping at the first uncorroborated cause is the fallacy.
  Different failures. Single root-cause label earned only after contributing
  causes were searched for and each accepted cause corroborated twice.

## Stated limits

Holes, not features. Named so nothing gets quietly asserted.

Weighted total is plain arithmetic: multiply each score by its criterion's
weight, add across. Arithmetic, not a sourced method.

Three different things called sensitivity. Not interchangeable, each has one
home. Evidence-sensitivity: references/evidence.md. Weight re-run:
references/comparison.md. Cost sensitivity: references/method-cost.md. Run one,
claim another done -> commonest way this handover ships incomplete.

Six more holes, each with its home.

- Matrix pathologies: references/comparison.md, "Where the arithmetic
  misbehaves".
- Whose standard settles acceptable and suitable: references/criteria.md,
  "Whose standard settles acceptable and suitable".
- How much analysis a question earns, the depth ladder: references/framing.md,
  "The depth tiers".
- Second-order effects: references/options.md, "The second-order pass".
- Rounding and significant figures: references/comparison.md, "Where the
  arithmetic misbehaves".
- Calibration over time not claimed: references/evidence.md, "Deriving a
  probability".

## Reference files

Load on demand. Each is this skill's own file.

- references/framing.md. Steps 1 to 3: environment, actors, three-part problem
  statement, constraints list, problem structure, depth tiers, technique
  routing table.
- references/evidence.md. Step 4 plus the uncertainty layer used throughout:
  fact notes, source assessment, assumptions, deriving a probability,
  likelihood ladder, confidence, falsifiers, key assumptions check,
  evidence-sensitivity test.
- references/criteria.md. Step 5: screening categories, five parts of an
  evaluation criterion with a worked example, weights and their rationale,
  direction, normalization, mandatory gate, freeze.
- references/options.md. Steps 6 and 7: distinctness, building options when
  none were handed over, single-option defense, per-option analysis,
  alternative-futures test.
- references/comparison.md. Step 8: matrix with a worked example, both totals,
  weight re-run, tie-breaking, close-call clause, committing to one course,
  handing over the losers.
- references/handover-check.md. Step 9: the pass or fail checks the assembled
  handover has to survive, each with what it protects and the fix a fail
  demands.
- references/method-risk.md, references/method-root-cause.md,
  references/method-causal-claims.md, references/method-cost.md. Escalation
  library, one file per trigger; load only the file step 3 routed to.
  references/challenge-techniques.md holds the four the operator names and you
  never trigger.