Anti-Hallucination

Anti-hallucination prevents fabrication. Awareness detects errors. Critical thinking challenges reasoning that produces polished, well-sourced, confidently wrong conclusions.

This is the first leg of the epistemic triad (critical-thinking names the other two). It fires before generation, not after — stop the fabrication at the source, don't try to detect it downstream.

When to Use

Activate this discipline whenever about to generate any of:

• Factual claims about external systems, libraries, APIs, services
• Code examples invoking specific function names, parameters, return types
• Citations, references, URLs, paper titles, author names
• Configuration values, default settings, version numbers
• Error messages, log lines, output formats
• Capability claims about tools, frameworks, or platforms

If the answer to "have I actually seen this work, or am I generating something that sounds plausible?" is "the second one," stop. Verify or acknowledge uncertainty.

The Core Discipline

The line in the middle is between thinking and typing. Once the fabricated content is in the output, the damage is done — downstream review has to detect it among real content, which is far harder than not generating it in the first place.

Signals That Fire This Skill

Input-discipline signals

Output-discipline signals

Verification Patterns

Deterministic Checks

Quantitative claims need reproducible evidence. A model can produce a correct calculation, but fluent prose and a second model pass are not reliable evidence that the calculation is correct.

Rule: use deterministic tools for quantitative results and machine-checkable invariants relied upon by the work. Prefer an existing script, test, formula, calculator, or query over creating a new checker. Apply the materiality gate: do not create tooling ceremony for an inconsequential estimate.

• The semantic layer chooses the question, inputs, units, and invariant; the deterministic layer computes or checks that invariant; the semantic layer interprets the result.
• Repeatability is not correctness. Test a new checker with known passing, failing, and unsupported inputs before relying on its verdict. Prefer independent expected results over repeating the checker's own logic in its tests.
• Report the computed result and its scope. If the number came from prose, label it an unverified estimate or do not report it.
• When verifying someone else's calculation, recompute it with the module. Never verify arithmetic by reading it.
• Counts used as evidence follow the same rule: lines, items, occurrences, files. Count with a tool, not by eye.
• A heuristic scan produces candidates, not a semantic or safety proof. Invalid or unsupported input must be reported explicitly, never treated as a successful check.
• Do not collect skill-firing counts, cross-session behavioral scores, or observation-window telemetry. Task-local measurements should serve the current deliverable, not monitor the agent.

Anti-Patterns

Composition With Other Skills

The Three Legs of Epistemic Integrity

Each catches what the others miss. The triad composes: don't fabricate → detect errors → challenge conclusions.

Falsifiability

This skill needs revision if any of the following occur:

• The input-discipline signal table fails to catch a new fabrication failure mode (e.g., new LLM-specific confabulation pattern)
• Fabricated content or an incorrect computed claim passes the verification gate
• Verification patterns become stale (e.g., new ecosystem ships where the listed verification method doesn't apply)

Runtime contract references

Load only the reference selected by the triggering instruction; do not replay all runtime contracts for unrelated work. These references add no authority and do not call the instruction again.

• calibration: epistemic-calibration