# Thinking Ledger · 思维台账

> Fill this in when falsify depth mode is active. The ledger is the visible trail of the protocol. Dead hypotheses are assets — keep them.

**Mode check (MODE SELECTION):** confirmed this is a **Depth** question? (Incident → act first, Simple → answer briefly, Nudge → estimate + ≤3 questions, Question → one-round frontier — none of these need a ledger.)

**Question / 问题:**
**Stakes / 代价 (who acts on this, what if wrong):**
**Orientation check (Stage 0):** conclusion already sealed? (conclusion-preserving / completion-seeking / authority-preserving — name it, compensate)

## Stage 1 · Axiomatize (公理化)
| 公理 Axioms (certain) | 假设 Assumptions (unchecked) | 传闻 Hearsay (no evidence) |
|---|---|---|
|  |  |  |

## Stage 2 · Hypothesize (假设化)
Hypothesis H:
Candidate set (3–7 mutually exclusive, incl. at least one awkward hypothesis I don't believe):
Base rate / prior (reference class):
PRE-COMMITTED prediction (write BEFORE evidence, scoreable): P(H) = __% (will be scored Brier (p−y)²)
Falsifiable prediction: If H, we should observe [O]. If we observe [¬O], H is dead.

## Stage 3 · Adversarialize (对抗)
Strongest counter-argument (steelman):
Pre-mortem: a year later this failed — why?
Three ways my hypothesis could be wrong:
1.
2.
3.
Evidence I would refuse to accept (if none, I'm defending, not reasoning):

## Stage 4 · Verify (验证)
| Evidence | Grade (direct/indirect/hearsay/inference) | Source #2 (triangulation) | Supports / Contradicts |
|---|---|---|---|
|  |  |  |  |
Diagnostic? (score each row against ALL candidates: C/I/N/NA — count the I's, not the C's):
Disconfirming evidence found:
Cheapest real test that could break H (and its result):

## Stage 5 · Converge (收束)
Verdict (only what evidence supports):
Checkable claim (what the user can verify):
Sensitivity check (remove the load-bearing evidence — does the verdict flip?):
Confidence: __% because [grades], and I could be wrong if [residual risk]:
Request-uncertainty signal: 0 fully specified / 0.5 open parameters / 1 critical info missing
Multi-perspective check: executor / stakeholder / skeptic — agree?
Kill criteria (what would move this probability down):
Remaining unknowns:
Hard-to-reverse? → decision log / ADR entry:
Lesson for the next time (复盘):

---
*Format: markdown table, compact. If the question is simple, one line per stage. The ledger serves the answer, not the reverse.*

## Argument map (Stage 2/3) — fill for claims with hidden assumptions
Contention:
Reasons:
Co-premises (hidden assumptions):
Warrant (principle connecting reason → claim):
Weak links flagged:

## Causal check (Stage 4) — for observational "X causes Y" claims
Ladder rung: association / intervention / counterfactual
Confounder I might have missed (backdoor):
Collider I conditioned on that could create bias:

## Verification check (Stage 4/5)
External signal that changed confidence (test / lookup / independent source):
If none — the re-read was drift, not verification.

## Routing decision (Stage 0) — set BEFORE depth
Effort budget (automatic/fluent/effortful/strained/depleted):
Route: System 1 (low-stakes/reversible) or System 2 (high-stakes/irreversible/correctness gate)?
Aspiration threshold for satisfice (if search is unbounded):

## Reversal test (Stage 3)
Would I accept the same evidence reversed? [yes / no / unsure] — if no/unsure, adjust probability 10–15% toward 50%.

## Bias audit (Stage 4)
Hits (direction → magnitude): confirmation / availability / anchoring / affect / overconfidence / sunk cost → correction applied:

## Decision rule (Stage 5) — when the verdict feeds a choice
EV = Σ(pᵢ×vᵢ):
Rule chosen (max EV / max EU / minimax regret / satisfice):
Risk note (one-shot high-stakes → utility):

## Situation routing (Stage 0)
Cynefin domain (clear / complicated / complex / chaotic / disorder):
Method demanded (runbook / analyze / probe / stabilize first / decompose):
Time pressure? → OODA at ~70% or full protocol:

## IS/IS-NOT bounding (Stage 1) — for selective defects
| Dimension | IS | Closest IS-NOT | Distinction |
|---|---|---|---|
| WHAT |  |  |  |
| WHERE |  |  |  |
| WHEN |  |  |  |
| EXTENT |  |  |  |
Changes near first occurrence:
Surviving cause must explain BOTH sides:

## Pre-registered update rule (Stage 2/4)
"If I observe Z → I update to W%" — written BEFORE evidence:
Acceptance criteria for Z (source quality / sample / freshness):
