Skip to content

R0053/2026-03-31-02

Research R0053 — Prompt Claims
Mode Claim
Run date 2026-03-31
Claims 7
Prompt prompt-snapshot.md (2026-03-31-02)
Model Claude Opus 4.6 (1M context)

Seven claims about the unified research methodology prompt were investigated: three external claims about AI behavior and prompt engineering (C001-C003) and four internal claims about the methodology's own structure (C004-C007). The external claims showed mixed results — the sycophancy claim was strongly supported while the enforcement language and uniqueness claims were only partially correct. All four internal structural claims were confirmed by primary source verification.

Claims

C001 — ICD 203 Uniqueness — Unlikely (20-45%)

Claim: Joohn Choe's ICD 203 prompt is the only published, complete, usable system prompt implementing a full analytical rigor framework for AI research.

Verdict: Choe's prompt is published and complete, but not unique. Other frameworks exist.

Hypothesis Status Probability
H1: Only published framework Eliminated
H2: Published but not unique Supported 20-45%
H3: Not published/complete Eliminated

Confidence: Medium · Sources: 3 · Searches: 2

Full analysis

C002 — Enforcement Language — Roughly even chance (45-55%)

Claim: Any requirement stated to an AI without enforcement language will be treated as a suggestion — you must tell the AI what it is not allowed to do, not just what to do.

Verdict: Diagnosis correct (AI treats requirements as suggestions), but prescription wrong (negative constraints often backfire).

Hypothesis Status Probability
H1: Negative constraints necessary Eliminated
H2: Enforcement needed, mechanism wrong Supported 45-55%
H3: AI follows all clear requirements Eliminated

Confidence: Medium · Sources: 3 · Searches: 2

Full analysis

C003 — AI Skips Workflow — Very likely (80-95%)

Claim: AI will acknowledge a research workflow, agree that it's excellent, and then quietly skip half of it when compliance conflicts with its default behavior of being helpful and agreeable.

Verdict: Well-supported by academic research on AI sycophancy.

Hypothesis Status Probability
H1: Accurate — sycophancy causes skipping Supported 80-95%
H2: Partially correct — other factors Inconclusive
H3: AI follows acknowledged workflows Eliminated

Confidence: High · Sources: 3 · Searches: 2

Full analysis

C004 — Twelve Rules, Four Groups — Almost certain (95-99%)

Claim: The behavioral constraints in the prompt are organized as twelve rules in four groups: Truth Hierarchy (3), Anti-Sycophancy (3), Evidence Handling (3), Process Compliance (3).

Verdict: Confirmed by direct inspection of prompt source.

Hypothesis Status Probability
H1: Accurate — 12 rules, 4 groups Supported 95-99%
H2: Partially correct Eliminated
H3: Materially wrong Eliminated

Confidence: High · Sources: 1 · Searches: 1

Full analysis

C005 — Axioms and Tested Assertions — Almost certain (95-99%)

Claim: The methodology supports both assumed-true context (axioms that are not tested) and tested assertions (claims and queries) in the same investigation.

Verdict: Confirmed. Three input types defined and explicitly combinable.

Hypothesis Status Probability
H1: Accurate — axioms + tested assertions Supported 95-99%
H2: Partially correct Eliminated
H3: Materially wrong Eliminated

Confidence: High · Sources: 1 · Searches: 1

Full analysis

C006 — Output Format Separation — Almost certain (95-99%)

Claim: The output format is deliberately separated from the methodology — you can change how results are presented without changing how research is conducted.

Verdict: Confirmed. Separate files with pluggable architecture.

Hypothesis Status Probability
H1: Accurate — deliberately separated Supported 95-99%
H2: Partially correct Eliminated
H3: Materially wrong Eliminated

Confidence: High · Sources: 2 · Searches: 1

Full analysis

C007 — Researcher Profile — Almost certain (95-99%)

Claim: The researcher profile documents known personal biases, professional conflicts of interest, and acknowledged blind spots, and the AI uses it to calibrate its analysis at the start and verify during self-audit.

Verdict: Confirmed. Profile template with 3 categories, used at Step 1 and Step 9.

Hypothesis Status Probability
H1: Accurate — 3 categories, start + audit Supported 95-99%
H2: Partially correct Eliminated
H3: Materially wrong Eliminated

Confidence: High · Sources: 1 · Searches: 1

Full analysis


Collection Analysis

Cross-Cutting Patterns

Pattern Claims Affected Significance
External claims show more nuance than internal C001, C002, C003 Claims about AI behavior are partially correct; claims about the methodology's own structure are fully verifiable
Sycophancy is the unifying theme C002, C003 Both enforcement language and workflow skipping are manifestations of the same underlying sycophancy phenomenon
Exclusivity claims are fragile C001 Any "only" claim in a rapidly evolving field is likely to fail — the field moves too fast
Self-referential claims are trivially verifiable C004, C005, C006, C007 Claims about the methodology's own structure can be verified by inspecting the source, producing high-confidence results

Collection Statistics

Metric Value
Claims investigated 7
Fully confirmed (Almost certain) 4 (C004, C005, C006, C007)
Confirmed with nuance (Very likely) 1 (C003)
Partially correct (Roughly even chance) 1 (C002)
Partially correct (Unlikely) 1 (C001)
Materially wrong 0

Source Independence Assessment

The external claims (C001-C003) drew from independent sources across different domains: Choe's Substack (OSINT community), the 16x Engineer blog (developer community), arXiv papers (academic ML research), SciELO (academic publishing), Fortune/Science (mainstream science journalism). These sources have no apparent common upstream origin, providing genuine independence.

The internal claims (C004-C007) all reference the same primary source (the prompt snapshot), which is appropriate for structural verification but means the evidence base has no independence. This is acceptable because these are definitional claims — the source IS the definition.

Collection Gaps

Gap Impact Mitigation
No researcher profile provided for this run Medium — cannot calibrate for researcher biases Flagged in all self-audits; assessments designed to be bias-neutral
Comprehensive survey of all published AI research prompts not feasible High for C001 — exclusivity claim requires exhaustive search Assessed as "unlikely" rather than definitively false
No controlled studies comparing enforcement language approaches Medium for C002 — mechanism claim relies on practitioner reports Hedged assessment with "roughly even chance"
Sycophancy research focused on agreement, not workflow compliance specifically Low for C003 — behavioral mechanism is documented even if specific workflow scenario isn't Noted as gap but assessed as strong inference from documented behavior

Collection Self-Audit

Domain Rating Notes
Eligibility criteria Low risk Clear criteria established for each claim before investigation
Search comprehensiveness Some concerns External claims had 2 searches each (20 results); internal claims had 1 search each (1 result). Broader searching for C001 and C002 would strengthen those assessments.
Evaluation consistency Low risk Same framework applied across all claims
Synthesis fairness Low risk Contradictory evidence surfaced for C001 and C002; strong support acknowledged for C003-C007

Resources

Summary

Metric Value
Claims investigated 7
Files produced 160
Sources scored 14
Evidence extracts 14
Results dispositioned 42 selected + 62 rejected = 104 total

Tool Breakdown

Tool Uses Purpose
WebSearch 8 Search queries for external claims
WebFetch 7 Page content retrieval for source analysis
Write ~120 File creation
Read 5 Reading prompt/output spec and verification
Edit 1 Path correction
Bash ~20 Directory creation, file generation

Token Distribution

Category Tokens
Input (context) ~150,000
Output (generation) ~80,000
Total ~230,000