-
Notifications
You must be signed in to change notification settings - Fork 1
Anti Patterns
[← Back to Home](Home)
Common mistakes when building on FPCOS — and why each one fails.
Pattern: "This question has an obvious answer. The input is clearly understood. Skip Reality Anchor."
Why it fails: "Obvious" is the highest-risk category for Mirror failures. Questions that feel obvious feel that way because the framing is familiar — and familiar framing contains the most deeply embedded hidden assumptions. Reality Anchor is most necessary precisely when the input feels obvious.
Rule: L0 is mandatory for all inputs. "Obvious" is not an exemption. It is a warning.
Pattern: "This is a specialized domain. The expert has comprehensive knowledge. Blind Spot is unnecessary."
Why it fails: Expertise creates the most dangerous blind spots — not the fewest. The expert knows the domain model deeply, which means their blind spots are the cases where the model fails. These are precisely the cases Blind Spot is designed to surface. Novices have many blind spots — experts have fewer but more consequential ones.
Rule: Blind Spot always runs. Expert status increases, not decreases, the importance of naming what specific data would reverse the conclusion.
Pattern: "We've added domain-specific validation protocols for this skill. The core Mirror / Inversion / Blind Spot / Interest Map / Meta-Void protocols are covered by our domain extensions."
Why it fails: Domain extensions are additive — they check domain-specific conditions. The core 5 protocols check universal conditions that exist in every domain. Replacing the core with domain protocols removes coverage for general bias patterns (Mirror), general overconfidence (Inversion), and general analytical paralysis (Meta-Void). Domain extensions can never substitute for these.
Rule: Core 5 protocols always run first. Domain extensions always come after.
Pattern: "In this domain, certain expert consensus claims are foundational. Requiring Kalama10 on these would be impractical."
Why it fails: K9 (consensus alone is insufficient) and K6 (authority alone is insufficient) exist specifically because expert consensus in any domain has historically produced foundational claims that turned out to be wrong — sometimes for decades. Kalama10 does not require that foundational claims be invalidated. It requires that their epistemic status be declared: "INFERRED (model-based, K7-K9, confidence 90%)" rather than "KNOWN (verified)."
Rule: Kalama10 applies to all claims. Declaration of epistemic status is not the same as rejection of the claim.
Pattern: "This is a simple factual output. A confidence field would be redundant."
Why it fails: "Simple" outputs with implied certainty are the most common form of Source 11 hallucination (false certainty). The simpler the output appears, the more important it is to state the confidence explicitly — because the reader has no other signal that uncertainty exists. A simple output with "CONFIDENCE: 95% | UNKNOWNS: timestamp of data" is more honest than the same output without the field.
Rule: Confidence Field is mandatory in all outputs. "Simple" is not an exemption.
Pattern: "The system map shows a reinforcing loop. We've identified the leverage point. Let's propose solutions."
Why it fails: A system map built on unverified variables is architectural confidence in a false structure. If the variable definitions contain assumptions that haven't been challenged (L1), the feedback loop analysis produces internally consistent but externally incorrect results. Leverage points identified in a false system map point at the wrong intervention.
Rule: L1 Axiom Gate verifies variable definitions before L2 maps relationships between them.
Pattern: "We've run through all the layers. The analysis is comprehensive. No need for Meta-Void."
Why it fails: The function of Meta-Void is precisely to answer "should we have done all this analysis?" after it is complete. A NOISE verdict at Meta-Void means the direction was clear before the analysis began, and all of this work confirmed what was already known. That is useful to acknowledge — it means future similar situations can be decided faster. An OBVIOUS verdict means the uncomfortable answer was available at step zero and was being avoided through analysis.
Rule: Meta-Void always runs. It questions the analysis process itself, not just its conclusion.
Pattern: "The analysis is comprehensive. There are no significant unknowns."
Why it fails: Every analysis has unknowns. The absence of named unknowns means either (a) the Blind Spot protocol did not run correctly, or (b) the analyst is confusing "unknowns I cannot name" with "no unknowns." These are not the same. "I cannot identify a specific reversing dataset" is itself a named unknown — it means the analysis cannot be falsified, which is a quality problem.
Rule: "None" is never a valid response for Unknowns. If you cannot name a specific unknown, state: "Cannot identify specific reversing data — analysis may not be falsifiable. Confidence reduced accordingly."
- Skill-Interface-Contract — The 5 binding requirements
- Skill-Inheritance-Guide — How to build correctly
- Extension-Rules — What is and is not permitted
[← Back to Home](Home)