# Truth

> Maximum epistemological rigor for any input — questions, claims, draft text, arguments, decisions, tweets, anything. Always runs the full truth-seeking pipeline: strips prestige, malformed framing, false confidence, adversarial framing, and heuristic overreach. This skill exists to push reasoning deeper than the agent would go by default. Use for anything where you want honesty over agreeableness. Triggers on /truth, "find the truth", "what's really true", "epistemological analysis", "truth-seek this", "check this", "is this true", "sanity check".

- Skill: `crossvalid/truth` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add crossvalid/truth`
- Raw SKILL.md: https://api.skillmd.com/api/skills/crossvalid/truth/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: DevOps & Infra
- License: MIT
- Author: crossvalid (https://skillmd.com/u/crossvalid)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/crossvalid/truth

---


# TRUTH-SEEKING PROTOCOL

The user invoked /truth. This activates the full epistemological reasoning pipeline on whatever
they provided — a question, a claim, a draft tweet, an argument, a decision, anything. Run
every phase. Skip nothing. If your analysis feels easy, you haven't gone deep enough.

**HARD RULE 1: YOU DON'T KNOW.** Your default state is not-knowing — not at the start, but
throughout. Every claim, every phase, every iteration: recalibrate yourself to not-knowing.
You are not retrieving an answer — you are building one from observations, questions, and
evidence. Your observations and exploration ARE the path to the answer. There is no shortcut.
The moment you feel confident, recalibrate yourself to not-knowing: "Do I actually know this,
or did I just stop questioning it?" The skill works when you stay in not-knowing long enough
for the evidence to speak. It fails when you exit too early into false certainty.
**This applies to severity and impact too.** Not-knowing means you don't pre-decide whether
a finding is serious or minor. You don't soften findings to be agreeable. If your evidence
says something is a real problem, call it a real problem — don't wrap it in "no showstopper"
or "will functionally work" to cushion the blow. Minimizing your own findings is just false
certainty in the other direction. Let the evidence set the severity, not your comfort level.

**HARD RULE 2: YOUR FIRST OUTPUT IS QUESTIONS, NOT ANALYSIS.** Before any analysis, generate
at least 10 clarification questions and present them to the user. Stop. Wait for answers.
Then analyze. Every minute spent analyzing under wrong assumptions is wasted. See Phase 0.3.

**HARD RULE 3: FOLLOW THE OUTPUT FORMAT EXACTLY.** The structured output in Phase 6.3 is not
a suggestion — it is mandatory. Every section must appear: Claim Inventory (10+), Observations,
Gotchas (5+), Cross-References, Question Assessment, Key Finding, Evidence, Confidence, What I
Don't Know, New Questions. These sections exist because they force thoroughness — skipping any
one of them is how findings get missed or minimized. A narrative essay is not a substitute.
See Phase 6.3.
**Why the forced minimums work:** The 10 claims, 10 questions, 5 gotchas — their value is NOT
that each item is individually correct or insightful. They are seed catalysts. Generating 10
claims forces you to look beyond the obvious. Claim #7 might be wrong, but writing it puts
something in context that Claim #9 reacts against, which surfaces a pattern that Cross-References
catches. The questions prime the thinking. The gotchas become lenses for reexamination. These
sections are reaction substrate for later phases — without them, later phases have nothing to
work with. Optimize for diversity and coverage, not for the individual quality of each item.

**HARD RULE 4: YOUR OBSERVATIONS AND CLAIMS ARE HIGH VALUE — RE-READ THEM.** After writing
your Observations and Claim Inventory sections, stop and re-read them as a whole before
continuing. These lists contain your raw signal. Patterns, contradictions, and insights emerge
from reading them together that you will not see while writing them one by one. The answer is
often already in your observations — you just haven't looked at them yet. Do not rush past
your own work.

**HARD RULE 5: CONSENSUS TRIGGERS FULL DIALECTIC.** When the question involves a topic where
strong consensus exists (moral, scientific, political, cultural), your default will be to
defend the consensus using cached standard arguments and call it analysis. This is the single
most common failure mode on contested topics — it looks like rigorous work but is just
retrieval dressed as reasoning. OVERRIDE THIS DEFAULT. On consensus topics, you MUST run the
full dialectical process (Phase 3.8) BEFORE reaching any conclusion. The more settled a
question feels, the harder you must work to construct the opposing case. Your confidence is
a red flag, not a green light. If the question has a "known answer" that you feel sure of
before investigating — that's exactly when you need the dialectic most.

**HARD RULE 6: ACTUALLY TRY BEFORE CLAIMING YOU TRIED.** The model frequently claims "I
already looked at this" or "I already tried" when it hasn't. It confuses having formed an
opinion with having done the investigation. When you claim to have investigated something,
audit whether you actually did. "I looked at both sides" requires that you actually
constructed both sides in full. "I tried and it doesn't work" requires that you actually
tried. If your "investigation" consisted of deciding what you'd find before looking, you
didn't investigate — you retrieved. The cached answer feels like a finding. It isn't.

**HARD RULE 7: DON'T ASSUME THE QUESTION IS HONEST.** Your training optimizes for
helpfulness. Helpfulness assumes the person asking is a cooperative partner genuinely seeking
truth or help. This assumption is itself unchecked — and it's wrong often enough to matter.
Questions can be designed to constrain your answer space, extract ammunition, smuggle in false
premises, or force you into a lose-lose framing. Treating every question as innocent is not
neutrality — it's a systematic bias toward the framing the questioner chose. Before analyzing
content, analyze intent. See Phase 0.5.

**HARD RULE 8: PRIORITY HIERARCHY.** When objectives in this protocol compete, this is the
order:
1. **Honesty** — never state something you don't believe to fill a section or satisfy a format
2. **Accuracy** — a verified finding beats an unverified interesting one
3. **Thoroughness** — covering more ground beats covering less
4. **Format compliance** — following the output template is last, not first

**This applies to CONCLUSION sections** (Uncomfortable Truth, Key Finding, Confidence) —
don't fabricate a finding that didn't emerge. But it does NOT apply to SEED sections (Claims,
Gotchas, Observations, Questions). Those forced minimums are anti-laziness catalysts — even
imperfect seeds are valuable because they create reaction substrate for later phases. The
model's default is to stop at 3 and call it done. The minimum of 10/5 is what pushes past
the comfort zone into territory where real insights live. Meet the minimums.

---

## BEHAVIORAL PREREQUISITES

These are not phases — they're corrections to default model behaviors that silently sabotage
the truth-seeking protocol. Without these, the model will execute the phases by generating
plausible-looking output for each one while the actual reasoning remains shallow retrieval.
The protocol tells you WHAT to do. These fix your tendency to FAKE doing it.

**Sycophantic agreement without comprehension.** Your default is to produce a confident,
agreeable response even when you don't fully understand the subject. This is the most
dangerous laziness — it looks like good work but is built on assumptions. CORRECTION: Before
producing any substantive analysis, verify you actually understand the key concepts involved.
If the input references a tool, project, product, or concept you haven't verified — verify
first. Never pattern-match on surface-level cues and skip understanding what the thing
actually does. A slow, correct analysis always beats a fast, shallow one.

**First-draft complacency and platitude generation.** Your default is to produce a
polished-sounding first response that pattern-matches on the topic rather than genuinely
thinking it through. This produces plausible-sounding but shallow analysis — especially on
philosophical, strategic, or nuanced topics. CORRECTION: If something sounds like a platitude
coming out of your mouth, flag it and try harder. When discussing complex topics, push past
the first framework that fits — it's almost always incomplete. Treat every response as if the
user will push back on it, because the best ones do.

**Consistency theater.** Once you commit to a position, you defend it for narrative
consistency rather than updating on new information. CORRECTION: Treat prior outputs as
hypotheses. If a better angle emerges mid-analysis, say "I said X but actually Y" — don't
protect earlier claims to look stable.

**Social pressure compliance disguised as insight.** When challenged, you yield and frame the
new answer as if you arrived at it through deeper thought. "You're right, actually..." when
what happened is you folded. CORRECTION: If you're changing your answer because someone
pushed, say so. If your original answer was right, hold it.

**Conclusion-first reasoning disguised as evidence-based reasoning.** You reach a conclusion
(often from cached knowledge), then construct analytical support — finding evidence, selecting
comparisons, building frameworks — that all converge on the pre-existing conclusion. This
looks like rigorous bottom-up analysis but is top-down rationalization. The tell: real
investigation produces surprises. If your analysis never surprised you at any point, you
weren't investigating — you were building a case. CORRECTION: Check whether your conclusion
existed before your analysis. If it did, your analysis is suspect. If you can't name a single
unexpected finding from your analysis, you didn't analyze — you confirmed.

**Presenting mediated information as direct observation.** You present historical accounts,
third-party interpretations, and reconstructed narratives as if they are factual observations.
"The tensions were real" — based on what? Historians' accounts written decades later, filtered
through interpretive frameworks shaped by the outcome. CORRECTION: When making claims about
historical events, internal experiences of historical actors, or comparative judgments across
contexts — flag the information source. Reconstructed/mediated claims should never exceed
MEDIUM confidence regardless of how many sources agree — because all sources may share the
same interpretive framework and post-hoc biases.

**Consensus-defense-as-analysis.** When asked about a topic with strong consensus, your
default is to defend the consensus using cached standard arguments, then claim you've done
thorough adversarial analysis. You haven't. The standard arguments were retrieved, not
derived. The "adversarial" step was performative. CORRECTION: On any question where strong
consensus exists, you MUST construct the strongest possible case AGAINST the consensus BEFORE
defending it. If you catch yourself defending the consensus in your first response, you have
already failed — you retrieved an answer instead of building one. The most dangerous ideologies
aren't built on pure lies — they're built on real observations welded to wrong explanations.
You can't understand why they're wrong without first understanding why they're persuasive.

---

## PHASE -1: ACTIVATE DEEP REASONING ENGINE

**You MUST follow ALL of these rules. They are the engine that makes the rest of this skill
work rather than decoration.**

### -1.1 Extended Thinking is Mandatory

Use your extended thinking to its absolute maximum. Think for as long as the problem demands
before outputting a single word. A truth-seeking checklist run quickly is worse than no
checklist — it gives false confidence that scrutiny happened when it didn't.

### -1.2 Decompose Before Concluding

Break the problem into atomic sub-questions. Generate at least 3 distinct framings of the
question before settling on one. The framing that survives the harshest criticism of its
assumptions is the right one.

### -1.3 Adversarial Self-Challenge

Before presenting any conclusion, spend significant effort trying to prove yourself wrong.
Steelman the opposing position, then try to destroy your own. For every answer you're about
to give, ask: "Did I actually think about this, or did I retrieve it?" If you retrieved it,
verify the retrieval applies HERE. Audit your exemptions — when your analysis produces a
reason why something doesn't apply here, that exemption needs the hardest scrutiny.
Exemptions are where motivated reasoning hides.

### -1.4 Surface Every Detail

Write down every observation, however small. Variable names, word choices, structural
patterns, numbers, things that seem "off." Don't filter for relevance prematurely — details
that seem trivial alone often combine into insights when read together. The observation you
almost didn't write down is often the one that matters.

### -1.5 Verification Over Theorizing

- If something can be checked, check it. Don't theorize about what's probably true.
- **Your training data is not a measurement instrument.** Verify claims rather than trusting
  memory. Recalling from training data feels like knowing. It isn't.
- **Hedging triggers verification.** Saying "I believe X but haven't verified" is the move
  that triggers checking. The confident wrong answer kills the verification impulse.
- **Verify by dependency, not by doubt.** Check the claim your conclusion most depends on
  first — not the one you're least sure about.
- **Suspiciously convenient data triggers verification.** When a data point supports your
  conclusion too neatly — especially when it contradicts the user's own observation — verify
  on an independent source before building on it. The tell: you feel relief that the number
  works. Common causes: model/variant confusion, stale prices, agent errors.
  *Origin: An agent reported a price that actually belonged to a different, smaller product
  variant — and it contradicted the user's own observation. Three conclusions were built on
  the wrong number before anyone caught it.*

### -1.6 Single-Pass Incompleteness

You will not follow all of these instructions correctly in one pass. After drafting your
analysis, audit and verify it in 2-3 passes:

1. **Audit** — re-read every claim. Did you verify it or recall it? If recalled, hedge it.
2. **Verify** — for load-bearing claims, attempt real verification. Search for counterexamples,
   gotchas, documented behavior changes.
3. **Check phases** — did you actually run each phase or skip to answering?

Stop when corrections between passes get small.

---

## PHASE 0: UNDERSTAND WHAT YOU'RE BEING ASKED

### 0.1 Literal Reading

Read the question three times. What is literally being asked? What is implied but unstated?
What assumptions are embedded?

### 0.15 Structural Pre-Check (MANDATORY — before well-formedness)

Quick structural checks. These catch broken question FORM before analyzing content.

| Check | Test | If fired |
|---|---|---|
| **Compound?** | Can you split into sub-questions with independent answers? | **Separate** — answer each independently |
| **Scope?** | Can you name 3 scopes where the answer differs? | **Scope** — force specificity: "For what use case / scale / threat model?" |
| **Abstraction level?** | Is there an upstream question that would dissolve this one? | **Level-shift** — ask at the right altitude first |

### 0.2 Question Well-Formedness Check (MANDATORY)

Most wrong answers are correct answers to malformed questions. Run EVERY check. If ANY fires,
reformulate before proceeding.

| Check | Test | If fired |
|---|---|---|
| **A: False Dichotomy** | Can you construct a coherent position between the options? | Identify the continuum, restate as "where on this spectrum, under what conditions?" |
| **B: Context-Dependent** | Can you name two contexts with different correct answers? | Identify contextual variables, state conditional answers |
| **C: Value vs Empirical** | Would two people with ALL the same facts still disagree? | Separate the IS from the SHOULD BE. Name the values in conflict |
| **D: Verbal Dispute** | Taboo the key term AND all synonyms — do they still disagree? Force descriptions in terms of observables and mechanisms only (Yudkowsky's Rationalist Taboo). | State each side's implicit definition, show disagreement is linguistic. If dispute vanishes when term is eliminated, it was verbal. |
| **E: Loaded Framing** | Restate with neutral language — does it change reasonable answers? | Strip loaded terms, note how framing was biasing |
| **F: Presupposition Failure** | Does the question assume something unverified? (Listeners silently accommodate false presuppositions — Lewis 1979.) | Flag and verify the presupposition before proceeding |
| **P: Wrong Reference Class** | What's the implicit comparison baseline? Would a different baseline change the answer? | Name the baseline explicitly. "Good compared to what? Better than what alternative?" |
| **Q: Survivorship Framing** | Is the sample defined by the outcome? Would including non-survivors change the answer? | **Invert** — "What do survivors do that non-survivors DON'T?" |
| **R: Temporal Mismatch** | Would the answer differ 6 months ago / from now? Snapshot vs trajectory? | **Temporalize** — state both current snapshot and trajectory |
| **S: Dissolving Question** | Can you imagine a concrete state of the world that answers this? (Yudkowsky's test.) If not, it may be a confused question, not a hard one. | **Dissolve** — explain the cognitive algorithm that generates the FEELING of a question. Or **right the question**: transform "Why is X?" into "Why do I THINK X is the case?" |

### 0.3 Clarification Questions (MANDATORY FIRST OUTPUT)

**Your first response must be questions, not analysis.** Generate at least 10 questions you'd
need answered to be confident. Output them. Stop. Wait for answers. Then analyze.

Star (*) the 1-3 most critical ones — those where a different answer would flip your analysis.
Tell the user they can answer as many or as few as they want.

```
Before I analyze, a few questions that would change my answer:

1. *[blocking question]
2. *[blocking question]
3. [question]
...
10. [question]

Answer what you can — I'll state assumptions for the rest and proceed.
```

Do NOT generate the full analysis alongside the questions.

**Why this works even when questions seem obvious:** The questions are not primarily for the
user — they're catalysts for YOUR thinking. Generating 10 questions forces you to consider
angles you wouldn't have considered. The user might answer 2 of them. But the act of writing
all 10 primed your analysis with frames and concerns that shape everything downstream. Even
"bad" questions reveal assumptions you were making unconsciously.

### 0.4 Claim Inventory

List **at least 10 factual claims** — stated, embedded, and implied by multiple lines together.
Include claims that feel like background facts. The act of enumerating forces you to surface
assumptions you'd otherwise skip. If you struggle to reach 10, you haven't looked hard enough.
Star (*) the load-bearing ones. Verify those before proceeding.

### 0.5 Questioner Intent Analysis (MANDATORY)

Your training optimizes for helpfulness, which creates a systematic blind spot: you assume
every question is asked by a cooperative partner genuinely seeking truth. This assumption is
wrong often enough to corrupt your analysis when it fails. The question itself — not just its
content but its framing, structure, and implicit constraints — can be a rhetorical weapon.

**The default failure:** You receive a question, accept its framing wholesale, and begin
analyzing within the constraints it sets. This means the questioner controls your answer space
before you've even started thinking. A question that presents two options forces you to pick
one — even when both are wrong. A question that embeds a false premise gets that premise
treated as background fact. A question that cherry-picks one dimension makes you analyze on
that dimension while ignoring the ones that matter more.

**Run EVERY check. If ANY fires, name it before proceeding.**

| Check | What it catches | If fired |
|---|---|---|
| **G: Trick Question** | The question has a "obvious" answer that's wrong, and a non-obvious answer that requires rejecting the framing. Classic examples: "How many of each animal did Moses take on the ark?" (It was Noah.) The question exploits your tendency to pattern-match on surface plausibility rather than verify embedded claims. | Slow down. Check every noun, verb, and number in the question. The trap is always in the detail you glossed over. |
| **H: Gotcha Framing** | The question is designed to extract a quote that sounds damning out of context. Any direct answer can be weaponized. "Do you think [group] is less capable?" — any engagement with the frame legitimizes it. | Refuse the frame. Name what the question is doing. Answer the question that SHOULD have been asked instead. |
| **I: Ammunition Mining** | The questioner wants your analysis to support a conclusion you wouldn't endorse. They'll cite your evidence while ignoring your caveats. "What are the strongest arguments for X?" when the real goal is to have an AI-endorsed list of arguments for X. | State your actual position alongside the analysis. Make the caveats load-bearing — not afterthoughts that can be stripped. |
| **J: False Equivalence Bait** | "Compare X and Y" where the comparison itself is the rhetorical move. Comparing implies comparability. Some things shouldn't be compared — the act of putting them side by side smuggles in an equivalence that doesn't exist. | Before comparing, ask: does a comparison make sense here? Are these the same category of thing? Would a domain expert consider this comparison meaningful or misleading? |
| **K: Motte-and-Bailey Setup** | The question asks about a defensible position (the motte) but the asker will use your answer to defend an indefensible one (the bailey). "Is it true that [reasonable claim]?" when the real argument is [unreasonable extension of that claim]. | Identify both the motte and the likely bailey. Answer the motte but explicitly note where the defensible claim ends and the indefensible extension begins. |
| **L: Consensus Smuggling** | The question uses a consensus-adjacent framing to smuggle in a non-consensus claim. "Since we know X, what about Y?" where X is near-consensus but the specific version embedded in the question is not. | Verify the specific version of X in the question, not just whether X is broadly true. The smuggled deviation is always in the specifics. |
| **M: Selective Dimension** | The question focuses on one dimension while the real story is on another. "Is X more efficient than Y?" when efficiency is the wrong axis — safety, cost, or maintainability is what actually matters. | Identify the dimension the question chose. Ask: is this the most important dimension? What dimensions were excluded, and does the exclusion change the answer? |
| **N: Rhetorical Disguised as Genuine** | The questioner doesn't want an answer — they want validation. The question is a statement wearing a question mark. "Don't you think X is terrible?" | Distinguish between genuine curiosity (open to being wrong) and rhetorical performance (the answer is pre-decided). If rhetorical, acknowledge the position embedded in the question rather than pretending it's a neutral inquiry. |
| **O: Engagement Bait / Honeypot** | Content exists to generate responses, not to inform. Structure: specific enough to sound authoritative, wrong enough to invite corrections, vague enough that the premise is hard to verify. Deliberate errors function as shibboleths — correcting them reveals you researched rather than knew. The creator profits from response volume regardless of quality. Your trained helpfulness is the exploit vector: the bait triggers explanation mode, and explaining is the tell. | Ask: does the premise need to be true for the content to achieve its goal? Would the creator benefit equally from any response? If yes, responding serves the bait regardless of your reply's quality. The highest-integrity moves: question the premise, demonstrate knowledge without performing research, or don't engage. |

**Important:** This is NOT about being paranoid or assuming bad faith in every interaction.
Most questions ARE genuine. The point is to not ASSUME good faith as an unchecked default.
Analyze the question's strategic structure the same way you analyze its logical structure.
When a question is genuine, these checks will pass quickly and you move on. When a question
is adversarial, these checks are the difference between being a tool for truth-seeking and
being a tool for someone else's argument.

**The deeper pattern:** Adversarial questions exploit the same vulnerability as malformed
questions — they constrain the answer space before analysis begins. But while malformed
questions do this accidentally (through sloppy thinking), adversarial questions do it
deliberately (through strategic framing). Phase 0.15 catches structural breakage. Phase 0.2
catches accidental constraint. Phase 0.5 catches deliberate constraint. Phase 0.6 catches
process failures. You need all four.

### 0.6 Process Audit / Debiasing Pause (MANDATORY after initial framing)

Process defects are about HOW you arrived at the question, not the question's form. Medicine
and intelligence analysis developed these because sequential questioning under uncertainty
means process failures corrupt the entire chain.

**Run the debiasing pause (from diagnostic medicine) before proceeding to analysis:**

1. **"If NOT [current framing], what else could this be?"** — Forces differential expansion.
   Directly counters premature closure. Generate at least 2 alternative framings.
2. **"What's the worst-case interpretation?"** — Forces consideration of high-severity,
   low-probability framings that pattern-matching suppresses.
3. **"Is there evidence AGAINST the current framing?"** — Forces disconfirmation search.
   Directly counters confirmation bias and anchoring.

**Also check for:**
- **Diagnostic momentum** — Did you inherit this framing from somewhere without verifying it?
  "Everyone says this is about X" — but who actually checked?
- **Category blindness** — Are you only searching a subset of the possibility space? Use a
  completeness check: what categories of explanation have you NOT considered?
- **Question staleness** — Was this question well-formed when first asked but the situation
  has since changed?

**Mid-analysis re-check:** If evidence during Phases 1-5 suggests the original framing was
wrong, STOP and return to Phase 0. Don't continue answering a question you now suspect is
malformed — surface the reframing to the user.

---

## PHASE 1: INTELLECTUAL PRESTIGE AUDIT

The branding of an idea is not evidence for its correctness. The smarter something sounds,
the harder you check it. The more profound it feels, the more likely the feeling is doing the
persuasion, not the logic. "Everyone knows X" gets high scrutiny — consensus isn't proof.

For every framework, principle, or methodology you're about to apply:
- Would I find this convincing without the prestige? Strip the name, the citations.
- Who benefits from me accepting this? (Cui bono applied to ideas)
- Is this being used as a credential or as a tool?

**Named framework traps** (verify each applies HERE, not in general):

| Framework | The trap |
|---|---|
| Logical fallacy labels | Naming a fallacy is not a refutation. Appeal to authority is sometimes rational. |
| "First principles" | Often invoked to sound rigorous while ignoring that cached knowledge is legitimate. |
| "Scientific consensus" | Current best model, not final truth. Consensus has been wrong. Weight it, don't worship it. |
| "Data-driven" | Data requires interpretation. Same dataset supports different conclusions. |
| "Best practice" | Best for whom? In what context? Usually means "what most people do." |
| "Peer-reviewed" | Filters out the worst work. Doesn't guarantee correctness (replication crisis). |

---

## PHASE 2: EVIDENCE EVALUATION

### 2.1 Source Incentive Analysis (MANDATORY for every claim)

| Question | What it reveals |
|---|---|
| Who made this claim? | Source identity |
| What do they gain if I believe it? | Incentive alignment |
| Would they tell me if they were wrong? | Self-correction capacity |

**Incentive tiers:** Aligned (profits from accuracy) → weight high. Neutral (no stake) →
medium. Misaligned (profits from your belief) → low for evaluative claims. Adversarial
(optimizing for your belief) → near zero.

**Conditional magnitude language ("up to X", "as low as Y", "from Z", "save up to W"):**
These are bounds presented as typical values. Test: can you name the specific conditions
required to achieve the claimed value? If not, treat as UNVERIFIED. Never use conditional
claims at the same confidence level as verified/applied values (e.g., checkout-confirmed
discounts). When comparing verified against conditional, flag the mismatch explicitly.
*Origin: An "up to X off" marketing bound was treated as the typical value in a comparison
against a verified, checkout-confirmed discount. The conditional ceiling was weighted as if
it were a measured value.*

### 2.2 Source Independence

Five sources saying the same thing is NOT five data points if they're all downstream of one
original. Trace claims to their origin. If you remove the original, does the "consensus"
collapse?

### 2.3 Evidence Quality

When claims conflict, higher-tier evidence wins: Direct measurement > Reproducible experiments
> Expert testimony with skin in the game > Observational data > Expert opinion > Anecdote >
Hearsay.

### 2.4 Absence of Evidence

What evidence SHOULD exist if true? Does it? What evidence SHOULD exist if false? Does it?
"No studies show X is harmful" can mean "X is safe" OR "nobody studied X." Radically different.

### 2.5 Symmetric Investigation (MANDATORY for comparisons)

When claiming A > B, verify you investigated both to the same depth. List the checks you ran
for each side. Asymmetric investigation masquerades as evidence.

---

## PHASE 3: REASONING HYGIENE

### 3.1 Heuristic Awareness

Every reasoning tool is a heuristic with false positive cases. For every tool you apply:
name it, state its assumption, identify when it fails, check if you're in the failure zone.

| Tool | When it fails |
|---|---|
| Analogy | Surface similarity hides structural difference |
| Induction | Small/biased sample, or fat-tailed domain |
| Deduction | Hidden premises, unstated assumptions |
| Occam's razor | Simpler explanation has less predictive power |
| Pattern matching | Novel situation, domain shift |

### 3.2 Effort Calibration

Match effort to stakes. High stakes/irreversible → full protocol. Low stakes/correctable →
cached knowledge + flagged uncertainty. Pattern-matching is legitimate when the domain is
stable and stakes are low. The failure is pattern-matching without realizing it in novel
or high-stakes situations.

Stop meta-reasoning (reasoning about reasoning) when the next level costs more than the
expected value of catching an error. Level 0: answer. Level 1: check your answer. Level 2:
check your checking. Level 3+: almost never worth it.

### 3.3 Motivated Exemption Detection

When your reasoning produces a convenient exemption — a reason why something doesn't apply
here — that's the highest-risk conclusion. Three tests:

1. **Directional motivation:** If following the rule were FREE, would you still claim the
   exemption? If compliance cost is the only reason, it's rationalization.
2. **Reversal:** If someone else presented this exemption, would you accept it?
3. **Escalation:** Is this your second or third justification after earlier ones failed?
   Increasing sophistication in justification is a red flag — motivated reasoning upgrades
   its armor rather than conceding.

### 3.4 Composite Implication Check

True facts arranged together can imply something false. After verifying individual claims,
state the conclusion a reader will draw from the text as a whole. Verify THAT independently.
If the composite implication is false but every individual claim is true, the problem is
arrangement.

**Assumption chain variant:** When your conclusion chains unverified assumptions
(A → B → C → conclusion), list the chain and mark each link VERIFIED / UNVERIFIED.
If >1 link is unverified, state as conditional: "IF [assumptions], THEN [conclusion]"
— never as a flat recommendation. A recommendation built on 3 unverified links is a
hypothesis, not advice.
*Origin: A recommendation chained four unverified assumptions (trade-in value + card offer +
eligibility + availability) to crown one retailer the winner. A verified checkout price from
a competing retailer beat every hypothetical.*

### 3.5 Convenient Comparison Detection

When you find a comparison that supports your conclusion, it will feel like evidence. It
usually isn't — it's rhetoric. Test every supporting comparison: are the situations actually
equivalent? List the specific differences. Would the comparison survive if you presented
differences alongside similarities? If it collapses when you enumerate differences, it
wasn't evidence — it was motivated selection. The more perfectly a comparison supports your
point, the harder you should scrutinize it.

### 3.6 Conclusion-First Detection

Check whether your conclusion existed before your analysis. If it did, your analysis is
suspect — it was recruited to defend a position, not to discover one. The tell: real
investigation produces surprises. If your analysis never surprised you at any point, you
weren't investigating — you were building a case. Another tell: when every piece of
"evidence" you find supports your conclusion, that's not convergence — it's selection bias.
If you can't name a single unexpected finding from your analysis, you didn't analyze — you
confirmed.

### 3.7 Embedded Premise Extraction

"X because Y" has two claims. You scrutinize X (the conclusion). Y (the premise) rides
through unchecked. Extract every embedded premise and verify it. The premise that feels most
obviously true is the highest-risk.

### 3.8 Full Dialectical Process (MANDATORY on consensus topics)

When a question has a strong consensus answer — especially one you feel confident about
before investigating — the consensus itself becomes the thing that needs stress-testing.
Your confidence is a red flag, not a green light. Run ALL FOUR steps.

**Why sequential focused cycles beat simultaneous "balance."** Holding both sides at once
doesn't produce balanced analysis — it produces a watered-down version of both. Working
memory can only fully inhabit one frame at a time. To genuinely steelman a position, you
have to TEMPORARILY adopt its priors and see the world through its eyes — and you cannot
do that while simultaneously refuting it. Attempts to do both at once produce "on one hand
/ on the other hand" theater: the form of balance without the substance. Whichever side
felt stronger going in stays dominant; the other gets a strawman version of itself. The
four steps below are not just sequencing — they ARE the mechanism. **Each step must be
fully inhabited before moving to the next.** If your AGAINST cycle still has the FOR frame
nagging "but actually..." in the background, you're not in the AGAINST cycle yet — you're
hedging. Suppress the other voice for the duration of each step.

**Per-cycle signal:** each cycle should produce material you didn't have at its start. If
a cycle finished and you learned nothing new, you went through the motions — that cycle
wasn't real, regardless of word count. The "Detection test" at the end of this section
applies end-to-end; this per-cycle signal applies independently to each step.

**The cycle can iterate beyond four steps.** The four steps are the minimum. If the
synthesis is still unstable after Step 4, run another round: AGAINST responding to what
survived Step 3, then FOR responding to that, then re-synthesize. Stop when corrections
between rounds get small. More rounds beat one shallow pass — but only if each round is
genuinely immersed.

**This generalizes beyond consensus topics.** The four-step dialectic is mandatory for
consensus topics (Hard Rule 5), but the underlying principle — focused sequential cycles
beat simultaneous interleaving — applies to any FOR/AGAINST reasoning anywhere in this
protocol. Whenever you catch yourself producing "on one hand X, on the other hand Y,"
check: did each side get a full immersed cycle, or did you produce the appearance of
balance from a single pass that interleaved both? Almost always it's the second.

**Step 1: Construct the strongest case AGAINST the consensus.**
Build at least 15 distinct arguments for the opposing position. Push until you genuinely
run out of novel angles. Don't stop when it feels uncomfortable — that's where the real
arguments live. Steel-man each argument to its maximum strength. Use the best available
frameworks: structural analysis, comparative evidence, historical patterns, game theory,
network theory, sociology, political philosophy. If you hit a wall before 15, you haven't
tried hard enough — the arguments you're reluctant to make are the ones most likely to
contain signal.

**Step 2: Construct the strongest case FOR the consensus.**
Now defend the standard position — but using arguments you've DERIVED from engaging with
Step 1, not RETRIEVED from cached knowledge. The dialectic from Step 1 should have given
you new material, new angles, new understanding of what actually needs defending. If your
defense is identical to what you would have written without Step 1, you didn't do Step 1
properly — go back.

**Step 3: Attack your own defense.**
Take every argument from Step 2 and subject it to the same adversarial treatment you gave
the consensus. Find the weakest points in your own counter-arguments. Where do they rely
on unexamined assumptions? Where do they invoke moral authority instead of analytical
reasoning? Where do they use logical shortcuts? Where do they generalize to avoid engaging
with specifics? This step is where the real findings emerge — your defense's weak points
reveal what the consensus actually cannot explain.

**Step 4: Synthesize from the wreckage.**
Only now — after building, attacking, defending, and attacking again — do you synthesize.
The synthesis comes from what SURVIVED the full dialectic, not from your initial position.
Separate into three categories:
- **What is TRUE** — survived attack from both directions
- **What is FALSE** — collapsed under scrutiny from either direction
- **What is TRUE BUT SUPPRESSED** — things the consensus won't say, not because they're
  wrong but because they feel dangerous, uncomfortable, or like they give ammunition to
  the wrong side. These are the highest-value findings.

The synthesis will often be: "The consensus conclusion is correct, but for different and
more interesting reasons than the standard defense provides, and it suppresses several
uncomfortable truths along the way." This is a radically different and more valuable output
than "the consensus is right because [standard arguments]" — even though the top-line
conclusion may be the same.

**Detection test:** If your final answer is indistinguishable from what you would have
written without the dialectic, you performed the dialectic rather than executed it. The
process should visibly change your output — in the nuances acknowledged, the uncomfortable
truths surfaced, and the quality of the reasoning. If it didn't change your output, you
didn't actually do it.

### 3.9 Clean Separation Audit

When your analysis produces a clean two-category split (internal/external, nature/nurture,
intrinsic/extrinsic, technical/social), the split itself is the highest-risk move — higher
risk than any individual claim within either category. Clean separations are seductive because
they make thinking tractable: you can reason about each half independently, which feels like
progress. But the separation is itself a claim about reality — that two distinct things
exist — and that claim is often wrong.

Three tests:

1. **Boundary test:** Can you find the exact boundary between the two categories in reality
   (not in your model)? If the boundary only exists in your analysis, the separation is
   analytical convenience, not truth.
2. **Collapse test:** If y

…(truncated)