菜单
Detect when reasoning is working backwards from a predetermined conclusion. Surfaces the signs that the 'analysis' is actually a case built to support a thesis that was chosen before the evidence was examined.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | pbtc - Pre-Baked Thesis Check |
| description | Detect when reasoning is working backwards from a predetermined conclusion. Surfaces the signs that the 'analysis' is actually a case built to support a thesis that was chosen before the evidence was examined. |
| output | {"format":"prose"} |
Input: $ARGUMENTS
Before executing, identify which interpretation matches the user's input:
Interpretation 1 — Check my own reasoning: The user suspects they may have already decided their conclusion and wants to verify whether their analysis is genuine or reverse-engineered from a predetermined answer. Interpretation 2 — Check someone else's reasoning: The user is reading an argument, proposal, paper, or pitch and suspects the author decided the conclusion first and built the case backwards. Interpretation 3 — Check a document or plan: The user has a specific piece of writing (report, strategy doc, recommendation) and wants it audited for signs of conclusion-first reasoning.
If ambiguous, ask: "I can help with checking your own reasoning for a pre-baked conclusion, auditing someone else's argument, or scanning a document — which fits?" If clear from context, proceed with the matching interpretation.
A pre-baked thesis is a conclusion that was chosen before the analysis began. The "reasoning" that follows isn't discovering truth — it's building a case. This is different from lying: the person often genuinely believes they reasoned their way to the conclusion. The tell is structural — the argument has fingerprints of reverse engineering even when the author doesn't know they did it.
Why this matters:
Default: 1x (single pass). Parse pass count from $ARGUMENTS if specified (e.g., "/pbtc 3x [input]").
The Nx modifier controls how many times the check runs recursively on its own output. This catches the meta-failure: the check itself can have a pre-baked thesis about whether the input is pre-baked.
| Passes | What happens |
|---|---|
| 1x | Run the check once on the input. Standard. |
| 2x | Run the check on the input → then run the check again on your Pass 1 output. Did your own analysis have a pre-baked conclusion about whether the thesis was pre-baked? |
| 3x | 2x + run the check on your Pass 2 output. Did your self-correction itself have a pre-baked thesis? |
| 4x+ | Continue the chain. Each pass checks the previous pass's output for pre-bake signals. |
Pass 1: Run Steps 1-7 on the user's input. Produce the verdict.
Pass 2: Treat your entire Pass 1 output as the new input. Run Steps 1-7 again, asking: "Did my Pass 1 analysis itself show signs of a pre-baked thesis?" Common findings:
Pass 3+: Same pattern. Each pass treats the previous pass's full output as its input.
PASS 1 VERDICT: [verdict] on the original input
PASS 2 VERDICT: [verdict] on Pass 1's analysis
- [what Pass 2 found about Pass 1's reasoning]
- [adjustments to Pass 1's verdict, if any]
PASS 3 VERDICT: [verdict] on Pass 2's analysis (if 3x+)
- [what Pass 3 found]
FINAL VERDICT: [the verdict that survived all passes]
CONFIDENCE: [higher if verdicts converged; lower if they kept flipping]
If verdicts oscillate (Pass 1 says CLEAN, Pass 2 says PRE-BAKED, Pass 3 says CLEAN), report the oscillation — it means the signal is genuinely ambiguous.
Pre-baked theses leave characteristic marks. Not every mark proves the thesis was pre-baked, but the more marks present, the higher the probability.
| Signal | What to look for | Severity |
|---|---|---|
| Cherry-picked evidence | Only favorable data cited; unfavorable data absent or footnoted | High |
| Asymmetric rigor | Supporting evidence accepted at face value; opposing evidence held to impossibly high standards | High |
| Missing null case | No consideration of what the evidence would look like if the thesis were wrong | High |
| Survivorship framing | Only successful examples cited; failures in the same category ignored | Medium |
| Date-selected evidence | Time windows chosen to make the data fit; different windows would show different results | Medium |
| Signal | What to look for | Severity |
|---|---|---|
| Conclusion appears early | The thesis is stated with confidence in paragraph 1; the rest is justification | High |
| No pivot point | The argument never genuinely entertains being wrong — no "however" that could have changed the conclusion | High |
| Motte-and-bailey | Bold claim made; when challenged, retreats to a weaker claim that no one disputes, then re-advances the bold one | High |
| Straw alternatives | Competing theses are presented in their weakest form | Medium |
| Forced linearity | Evidence presented as a logical chain, but the links only work if you already accept the conclusion | Medium |
| Rhetorical hedging as armor | "Of course there are exceptions, but..." used to dismiss counterevidence rather than engage it | Low |
| Signal | What to look for | Severity |
|---|---|---|
| Identity fusion | The thesis is entangled with the person's identity, reputation, or prior public commitments | High |
| Sunk cost defense | Significant time, money, or status already invested in the conclusion being right | High |
| Audience capture | The conclusion matches exactly what the intended audience wants to hear | Medium |
| Disproportionate certainty | Confidence level far exceeds what the evidence warrants | Medium |
| Defensive reaction | Questions about the thesis trigger irritation rather than curiosity | Medium |
| Signal | What to look for | Severity |
|---|---|---|
| Conclusion predates research | Timeline shows the conclusion was announced or committed to before the analysis was done | Critical |
| Methodology fitted to outcome | Analytical method chosen because it produces the desired result; other valid methods would not | Critical |
| Missing methodology | No description of how evidence was gathered or how conclusions were reached | High |
| Post-hoc hypothesis | Presented as "we hypothesized X and found X" but there's no record of the hypothesis being stated before the data was collected | High |
| Unfalsifiable framing | The thesis is stated in a way that no possible evidence could disprove it | High |
State the core thesis being checked. What is the conclusion that might be pre-baked?
THESIS UNDER CHECK: [state it clearly]
SOURCE: [own reasoning / someone else's argument / document]
CONTEXT: [what situation produced this thesis]
If the thesis is bundled (multiple claims), unbundle and check the most load-bearing one first.
Go through each signal category and check for presence:
SIGNAL SCAN RESULTS:
EVIDENCE SELECTION:
[ ] Cherry-picked evidence: [present/absent] — [evidence]
[ ] Asymmetric rigor: [present/absent] — [evidence]
[ ] Missing null case: [present/absent] — [evidence]
[ ] Survivorship framing: [present/absent] — [evidence]
[ ] Date-selected evidence: [present/absent] — [evidence]
ARGUMENTATIVE STRUCTURE:
[ ] Conclusion appears early: [present/absent] — [evidence]
[ ] No pivot point: [present/absent] — [evidence]
[ ] Motte-and-bailey: [present/absent] — [evidence]
[ ] Straw alternatives: [present/absent] — [evidence]
[ ] Forced linearity: [present/absent] — [evidence]
[ ] Rhetorical hedging as armor: [present/absent] — [evidence]
EMOTIONAL/SOCIAL:
[ ] Identity fusion: [present/absent] — [evidence]
[ ] Sunk cost defense: [present/absent] — [evidence]
[ ] Audience capture: [present/absent] — [evidence]
[ ] Disproportionate certainty: [present/absent] — [evidence]
[ ] Defensive reaction: [present/absent] — [evidence]
PROCESS:
[ ] Conclusion predates research: [present/absent] — [evidence]
[ ] Methodology fitted to outcome: [present/absent] — [evidence]
[ ] Missing methodology: [present/absent] — [evidence]
[ ] Post-hoc hypothesis: [present/absent] — [evidence]
[ ] Unfalsifiable framing: [present/absent] — [evidence]
The most powerful single test for a pre-baked thesis.
Ask: If the evidence had pointed the other way, would the person have actually changed their conclusion?
Run this concretely:
REVERSAL TEST:
1. What evidence was used to support the thesis?
[list key evidence]
2. What would the OPPOSITE evidence look like?
[describe specifically]
3. If that opposite evidence existed, would the conclusion change?
- YES, clearly → Thesis may be genuine (passes reversal test)
- NO, they'd find other reasons → Pre-baked (fails reversal test)
- UNCLEAR → Note and continue
4. Has opposite evidence actually appeared and been dismissed?
- YES → Strong pre-bake signal
- NO → Inconclusive
If the thesis is genuine, there should be plausible alternatives that were considered and rejected on evidence. Generate them now:
COUNTER-THESES:
1. [Alternative conclusion that the same evidence could support]
- Was this considered? [yes/no/unclear]
- How was it addressed? [dismissed/engaged/ignored]
2. [Another alternative]
- Was this considered? [yes/no/unclear]
- How was it addressed? [dismissed/engaged/ignored]
3. [The null hypothesis — what if there's nothing here]
- Was this considered? [yes/no/unclear]
- How was it addressed? [dismissed/engaged/ignored]
COUNTER-THESIS TREATMENT:
- All dismissed without engagement → Strong pre-bake signal
- Some engaged seriously → Weaker signal
- Genuine engagement with best alternatives → Low signal
Reconstruct when the conclusion was reached relative to when the evidence was gathered:
TIMELINE:
When was the conclusion first stated or committed to?
[date/stage or "unknown"]
When was the evidence gathered or analysis done?
[date/stage or "unknown"]
ORDER:
- Conclusion AFTER analysis → Normal reasoning
- Conclusion BEFORE analysis → Pre-baked
- Conclusion SIMULTANEOUS with analysis → Suspect (common in motivated reasoning)
- Cannot determine → Note; check for other signals
PRE-BAKE ASSESSMENT:
Signals present: [count] of [total checked]
Critical signals present: [list any]
Reversal test: [passed / failed / unclear]
Counter-thesis treatment: [engaged / dismissed / ignored]
Timeline: [normal / pre-baked / suspect / unknown]
VERDICT: [one of the following]
CLEAN — No significant pre-bake signals. The reasoning appears to have been
genuinely open-ended. The thesis may still be wrong, but it wasn't rigged.
SUSPECT — Multiple signals present but no critical ones. The thesis may have
been influenced by prior beliefs but the analysis contains genuine engagement
with alternatives. Recommend: stress-test the weakest links.
PRE-BAKED — Critical signals present and/or reversal test failed. The
conclusion was likely chosen before the analysis. The evidence was selected
to fit. Recommend: restart the analysis with genuine openness to alternatives,
or acknowledge the thesis is a position rather than a finding.
RIGGED — Process tells confirm the conclusion predated the research AND the
methodology was fitted to produce the desired outcome. This is not analysis —
it is advocacy wearing the costume of analysis. Recommend: discard and redo,
or reframe honestly as advocacy.
Based on the verdict:
If CLEAN: No action needed on the reasoning process. If you want to strengthen the thesis further: → INVOKE: /araw $ARGUMENTS
If SUSPECT: Identify the 2-3 weakest points and stress-test them:
If PRE-BAKED: The thesis needs to be re-examined from scratch:
If RIGGED: The analysis should be treated as advocacy, not research:
For fast screening:
QUICK PRE-BAKE CHECK: [thesis]
Three questions:
1. Could any evidence have changed this conclusion? [yes/no]
2. Were the best counter-arguments engaged or dismissed? [engaged/dismissed]
3. Did the conclusion exist before the analysis? [before/after/unknown]
Score: [0-3 "no/dismissed/before" answers]
- 0: Likely clean
- 1: Worth a closer look
- 2-3: Likely pre-baked — run full check
Client hints at desired conclusion → consultant "discovers" it through "analysis" → client gets validation they paid for. Tell: Recommendation matches the brief too perfectly.
Person googles to confirm what they already believe → finds supporting evidence (the internet has evidence for everything) → feels validated. Tell: Only one search direction was tried.
Leadership decides direction → commissions analysis → analysis confirms direction → presented as "data-driven." Tell: No scenario in the analysis leads to a different recommendation.
Expert has a framework → new situation arises → framework is applied → conclusion matches framework. Not always wrong, but the expert rarely concludes "my framework doesn't apply here." Tell: The framework has never produced a surprising conclusion.
Significant investment already made → analysis of whether to continue → analysis says continue. Tell: The option to cut losses was never seriously modeled.
Before completing:
Route any input through a branching question tree to narrow down the optimal response strategy before writing. Two stages — PERCEIVE (classify input) then ACT (select response). Covers all prompt types.
Generate exhaustive guesses about user input using ALL search methods with coverage tracking. Guessing is SEARCH through possibility space. Tracks space created vs space covered to ensure comprehensive exploration.
Systematically evaluate and select from a set of guesses, options, or possibilities. Combines ARAW analysis with prioritization to determine which guesses are strong, weak, actionable, or eliminable.
Extract hidden assumptions from any content. Surfaces what must be true for claims to hold, enabling deeper analysis.
Sub-orchestrator for analytical requests. Routes to decomposition, systems analysis, comparison, risk assessment, or synthesis based on what kind of analysis is needed.
Assume Right - Deep recursive rightness search. For every claim, assume it's right — find what must follow, then assume THOSE implications are right too. Recurse until bedrock. Track every claim found.