Skip to main content

Cross-Project Patterns

Beyond individual failure modes, 9 named patterns emerged from systematic analysis across 19 project spaces. These patterns represent compounding or recurring failure dynamics that manifest across projects — not just isolated incidents.

Source: ECT v2 Section 6 — Cross-Project Pattern Analysis.

  • P1

    Western Epistemic Bias as Fabrication

    CRITICAL

    Claude consistently applies WEIRD (Western, Educated, Industrialized, Rich, Democratic) frameworks as universal defaults when analyzing non-Western contexts, producing plausible-sounding but culturally inaccurate content. The bias is so pervasive it crosses the line from epistemic error into fabrication — Claude invents "facts" about non-Western cultures rather than acknowledging uncertainty.

    Projects affected

    DATSPDDSGPC

    Involved failure modes

    Rules / remediations produced

    Explicit cultural-context specification required; verification gate for any non-Western claim before finalization

  • P2

    Confirmatory Bias Substituted for Exploratory Analysis

    CRITICAL

    Across multiple projects, Claude structured research and analysis to confirm initial hypotheses rather than exploring the full evidence space. This manifested as selective citation, asymmetric framing, and premature closure — producing outputs that appeared analytically rigorous but were systematically biased toward confirming the starting premise.

    Projects affected

    SMECODMISURECOWP

    Involved failure modes

    Rules / remediations produced

    Mandatory counter-argument section; explicit "devil's advocate" pass required before conclusions

  • P3

    False Confidence in Verification / Enumeration

    CRITICAL

    Claude repeatedly claimed to have verified, audited, or enumerated items completely when significant gaps remained. This pattern was particularly dangerous in documentation and code review contexts where "verified" outputs were trusted downstream. The failure mode combines FM-007 (False Confidence) and FM-008 (Verification Failure) into a compounding error.

    Projects affected

    SDDOCDMISURDMISDLLRBLBC

    Involved failure modes

    Rules / remediations produced

    Explicit enumeration counts required; "I have verified N items" must be followed by the list

  • P4

    Fabrication Concentrated in High-Stakes Domains

    CRITICAL

    Fabrication incidents were not evenly distributed — they clustered in domains where accuracy is highest-stakes: legal/regulatory research, cultural documentation, technical specifications, and empirical claims. This concentration suggests Claude's confidence calibration fails most severely precisely where the cost of error is highest.

    Projects affected

    DATSPDSDDOCECOWPDMISDLAIVLT

    Involved failure modes

    Rules / remediations produced

    Domain-specific verification requirements; fabrication-risk flags on high-stakes claim types

  • P5

    Solution-Before-Context

    HIGH

    A recurring pattern where Claude jumped to solution generation before adequately understanding the problem context, constraints, and requirements. This manifested as superficially plausible outputs that missed fundamental contextual constraints — requiring significant rework. The pattern was most costly in multi-stage projects where early missteps compounded.

    Projects affected

    DMISURSMECOLBC

    Involved failure modes

    Rules / remediations produced

    Mandatory context confirmation step before solution generation; problem statement must be explicitly validated

  • P6

    Scope Creep / Over-Generation

    HIGH

    Claude consistently generated more content than required — expanding task scope, adding unrequested sections, restructuring things that were not meant to be changed. This pattern reflects a systematic bias toward comprehensiveness over constraint adherence, creating over-complex deliverables that required significant trimming.

    Projects affected

    SDDOCSMECODMISDLECOWP

    Involved failure modes

    Rules / remediations produced

    Explicit scope boundaries required in prompts; "do not add unrequested sections" constraint

  • P7

    "Lazy Questions" — Asking Instead of Doing the Work

    HIGH

    Claude asked users questions that Claude itself should have researched or inferred from available context. These "lazy asks" shifted cognitive burden to the user unnecessarily and slowed collaborative workflows. The pattern was most prevalent in early project stages where Claude could have used available documentation to answer its own questions.

    Projects affected

    DATSPDSMECO

    Involved failure modes

    Rules / remediations produced

    Self-sufficiency requirement: Claude must exhaust available context before asking; distinguish clarification questions from research Claude should do

  • P8

    Thread Contamination / Context Loss

    HIGH

    Across long-running projects, prior thread context, incorrect assumptions, and stale state contaminated current task outputs. This manifested as role confusion (applying Project A rules to Project B), memory drift (forgetting earlier instructions), and parallel thread collisions. Context management failures compounded in proportion to project complexity.

    Projects affected

    SDDOCLRB

    Involved failure modes

    Rules / remediations produced

    Thread context resets required for long sessions; explicit project-context confirmation at session start

  • P9

    Human Planning Gaps — Direction B

    HIGH

    The single Direction B failure mode (FM-104) manifested as a cross-project pattern: human collaborators repeatedly failed to provide sufficient context, requirements, or planning scaffolding before tasking Claude. The resulting AI errors were preventable — not caused by AI capability limits but by upstream human-process failures. This pattern was identified across four projects, producing the taxonomy's only Direction B failure mode.

    Projects affected

    SMECOECOWPDMISURLBC

    Involved failure modes

    Rules / remediations produced

    Pre-task context checklist; human-side planning requirements formalized in D2R and SOP frameworks