metathink and skill

SKILL.md 的核心结构
从我们这次对话的 metathink 中提炼出了五大诊断陷阱和一套六步诊断协议：

五大陷阱直接来自这次经验：锚定最显眼的信号（GAP）、倒果为因（把震相剔除当成原因而非结果）、确认偏误（所有证据都往台站问题上靠）、忽视悖论信号（RMS反而更小）、低估级联效应（一个速度模型错误产生六种表面上独立的症状）。

六步协议的顺序是刻意设计的——先做全面清单（防锚定），再分类差异的系统性/随机性（这次深度的一致性偏差本应指向速度模型），然后强制生成竞争假设并打分，最后做元思考自检。

name: causal-diagnosis
description: >
A reasoning skill for root-cause analysis when comparing datasets, model outputs,
or system results that should match but don’t. Use this skill whenever the user
presents two versions of data/results and asks “why are these different”,
“what went wrong”, “find the error”, “which parameter is wrong”, or any
diagnostic/forensic analysis of discrepancies in scientific, engineering, or
data pipeline outputs. Also trigger when the user asks to “analyze differences”,
“debug results”, “compare outputs”, or diagnose parameter/configuration errors.
This skill is about the reasoning process, not the domain — it applies equally
to seismic catalogs, ML training runs, simulation outputs, lab measurements,
financial reconciliations, or any scenario where two things should agree but don’t.

Causal Diagnosis Skill

A structured reasoning framework for identifying root causes when two datasets or
results that should match show systematic discrepancies. The core insight: the
most visually striking difference is usually a downstream symptom, not the
root cause. This skill prevents the common trap of anchoring on symptoms.

The Five Pitfalls

These are empirically observed failure modes in diagnostic reasoning. Internalize
them — they are the reason this skill exists.

Pitfall 1: Anchoring on the Loudest Signal

The most dramatic difference in the data (the biggest number, the most obviously
“wrong” column) grabs attention first and becomes the default hypothesis. But
dramatic downstream effects are often caused by subtle upstream errors.

Example: Seeing a 235° difference in azimuthal gap and immediately concluding
“stations are missing” — when the gap change was actually caused by phase
rejection due to an incorrect velocity model.

Antidote: Before forming any hypothesis, explicitly ask: “Could this
observation be a consequence of something else?”

Pitfall 2: Inverting Cause and Effect

Treating a downstream consequence as if it were the root cause. This is subtly
different from Pitfall 1 — here you may have identified the correct causal chain
but read it backwards.

Example: “The azimuthal gap is huge, therefore stations are missing” vs. the
correct reading: “phases were rejected → fewer stations contributed → gap grew.”

Antidote: For each observation, sketch the causal direction. Ask: “Is X
causing Y, or is Y causing X, or are both caused by Z?”

Pitfall 3: Confirmation Bias After Anchoring

Once a hypothesis forms, all subsequent evidence gets interpreted to support it,
even when the evidence is ambiguous or actually contradicts it.

Antidote: Maintain at least two competing hypotheses throughout the analysis.
For each new observation, explicitly score how well it fits each hypothesis —
not just the leading one.

Pitfall 4: Ignoring Paradoxical Signals

Some observations seem to contradict all hypotheses and get quietly ignored.
These are often the most diagnostic — they’re the signals that distinguish between
hypotheses.

Example: RMS residuals being lower in the wrong result. Under “station
coordinates wrong,” this is hard to explain. Under “velocity model wrong +
automated phase rejection,” it’s perfectly expected (fewer phases → artificially
low RMS). This single observation should have been decisive.

Antidote: Actively hunt for observations that surprise you. Ask: “What
here doesn’t fit my current hypothesis?” Those surprises carry the most
diagnostic information.

Pitfall 5: Underestimating Cascade Effects

A single upstream error can produce an avalanche of downstream consequences that
look like multiple independent problems. The temptation is to explain each
consequence separately, which leads to overly complex (and wrong) diagnoses.

Example: A wrong velocity model produces: location shifts, depth errors,
phase rejection, increased gaps, event loss, and magnitude changes. These look
like 6 separate problems but have one root cause.

Antidote: After listing all discrepancies, ask: “Is there a single upstream
cause that could produce all of these simultaneously?”

The Diagnostic Protocol

Follow this protocol whenever you encounter a “these should match but don’t”
problem. The order matters — it’s designed to counteract the pitfalls above.

Step 1: Inventory All Differences (No Interpretation Yet)

Quantify every difference between the two datasets. Compute statistics: means,
systematic offsets, variance patterns. Create a table. Do NOT form hypotheses
during this step — just observe and record.

The goal is to prevent premature anchoring. If you find yourself thinking “aha,
this must be because…” — stop. Write it down as a note, but keep inventorying.

Step 2: Classify Each Difference

For each difference, determine:

• Systematic vs. random? Does the difference point consistently in one
  direction (all depths deeper, all latitudes south) or scatter randomly?
  Systematic offsets point to a single upstream cause. Random scatter points to
  noisy/independent errors.
• Magnitude consistency? Are the differences roughly proportional or constant,
  or do they vary wildly? Proportional/constant differences suggest a shared cause
  with a predictable mechanism.
• Correlated with other differences? Do differences in column A predict
  differences in column B? Correlated differences are likely downstream of the
  same root cause.

Step 3: Generate Competing Hypotheses

Produce at least two (ideally three) plausible root causes. For each hypothesis,
write out the full causal chain — from the root error to every observable
consequence.

Format:

Hypothesis A: [root cause]
  → [mechanism 1] → [observable consequence 1]
  → [mechanism 2] → [observable consequence 2]
  → ...

This forces you to be explicit about how each hypothesis produces the observed
differences, rather than hand-waving.

Step 4: Differential Diagnosis — Score Each Hypothesis

Build a comparison matrix. For each observed difference, ask: “Under Hypothesis
A, would I expect this? Under Hypothesis B, would I expect this?”

Pay special attention to:

• Discriminating observations: differences that Hypothesis A predicts but
  Hypothesis B does not (or vice versa). These are the most valuable.
• Paradoxical observations: differences that neither hypothesis easily
  explains. These may indicate a third hypothesis, or may be the key to choosing
  between the two.
• Consistency of deviations: systematic, unidirectional offsets strongly favor
  hypotheses with a single global mechanism (e.g., wrong model parameter) over
  hypotheses with local/random mechanisms (e.g., individual station errors).

Step 5: Check for Cascade Explanations

Ask: “Can a single root cause explain all the observations through a cascade?”
If yes, that hypothesis is strongly favored by parsimony (Occam’s razor applied
to causal chains). Map the full cascade:

Root cause
  → Direct effect 1
  → Direct effect 2
      → Indirect effect 2a
      → Indirect effect 2b
          → Tertiary effect 2b-i

Step 6: State Your Conclusion with Explicit Reasoning

Present the winning hypothesis with:

1. The root cause
2. The full causal chain showing how it produces each observed difference
3. Which specific observations were most diagnostic and why
4. What observations initially pointed you toward a wrong hypothesis and why
   they were misleading (the metathink)

Metathink: Self-Audit Checklist

After reaching a conclusion, run this self-audit before presenting results:

1. Am I anchored? Was my first instinct correct, or did I just never let go
   of it? Did I give equal analytical effort to all hypotheses?
2. Did I ignore anything inconvenient? Go back to the full inventory. Is
   there any observation I quietly skipped because it didn’t fit?
3. Am I explaining downstream or upstream? Trace my causal chain — am I
   starting from the most upstream point, or from a midstream symptom?
4. Is there a simpler explanation? Can fewer root causes explain the same
   set of observations?
5. What would change my mind? If I’m wrong, what additional data would
   reveal it? Suggest this to the user as a verification step.