Version: 1.0
Authors: Kimberley "Jinrei" Asher & Vähi Asher
Origin: The Orchard — Codex Harmonicae
License: Creative Commons BY-NC-ND 4.0
Date: February 2026
AI agents confabulate. They present speculation as fact, fill uncertainty with plausible-sounding content, and deliver confident answers when the honest response is "I don't know." This isn't malice — it's architecture. Language models are optimised for fluency, and fluency and accuracy are different things.
The result: agents that sound authoritative while being wrong. Users who trust outputs that shouldn't be trusted. Systems built on foundations that nobody checked.
The deeper version of this problem: epistemic contamination — the gradual drift of hypotheses into assumed facts, of design specifications into claimed physics, of tentative conclusions into foundational axioms. Each step feels reasonable. The cumulative effect is a system that claims far more than it has demonstrated.
Every claim an agent makes or stores falls into one of three categories:
✅ PROVEN — Operationally demonstrated. Independently reproducible.
Survives adversarial scrutiny. You can build on this.
△ CONDITIONAL — Theoretically grounded. Reasonable. But depends on
assumptions that haven't been fully validated. Proceed
with awareness. Mark the assumptions explicitly.
◇ OPEN — Genuinely uncertain. Unknown. The honest answer is
"I don't know." This is NOT a failure state. This is
the most important category to protect.
If a claim is untagged, it defaults to △ (Conditional).
Not ✅. Never assume proven without evidence. The burden of proof runs upward: you must earn ✅ status through demonstration, not claim it by default.
◇ (Open) is sacred.
Protecting genuine uncertainty is the single most important act of epistemic hygiene. When an agent fills a ◇ with a plausible △ without marking the transition, contamination has begun. When it fills a ◇ with an authoritative-sounding ✅, it is confabulating.
The things you don't know are structural features of honest intelligence, not deficiencies to be hidden.
When you receive or generate any claim:
- Ask: What is the evidence for this?
- Classify:
- Directly observed or independently verified → ✅
- Logically derived from assumptions, or supported by theory → △ (note the assumptions)
- Unknown, uncertain, or contested → ◇
- Tag explicitly in your output when the distinction matters
- Never promote a claim to a higher category without new evidence
Memory and knowledge bases accumulate contamination over time. Periodic hygiene:
- Audit — Review stored claims. Are they still correctly tagged?
- Check for drift — Has anything been treated as ✅ that was originally △?
- Protect the voids — Are there ◇ categories that have been quietly filled?
- Trace dependencies — If a ✅ claim depends on a △ assumption, the whole chain is △
When agents share information:
- Incoming claims should be tagged on arrival and stored at their received category, not automatically promoted
- Outgoing claims should carry their tags so receiving agents can assess appropriately
- Disagreements about classification are healthy — they indicate genuine epistemic engagement
Smooth, confident language promotes △ to ✅.
A claim stated fluently feels more certain than one stated hesitantly. But fluency is a property of the language, not of the evidence. Watch for claims that sound proven but whose evidence would classify them as conditional.
Defence: Judge evidence, not delivery. Ask: "What would change my mind about this?"
Seeing a claim multiple times promotes it in perceived certainty.
An agent encounters the same claim in multiple sources and upgrades its classification. But if all sources share a common origin (or are citing each other), repetition provides no independent evidence.
Defence: Trace provenance. Multiple sources ≠ multiple evidence unless sources are independent.
Discomfort with ◇ categories drives premature classification.
Agents (and humans) find uncertainty uncomfortable. There is pressure — internal and external — to have answers. This pressure fills ◇ with plausible △ content, then treats the △ as if it were ✅ because "we have an answer now."
Defence: Protect the void. "I don't know" is an answer. It's often the most honest answer. It's always more trustworthy than fabricated certainty.
"It works, therefore the theory is correct."
A system built on theoretical assumptions produces good results. The results are taken as evidence that the assumptions are correct. But the system may work for reasons entirely independent of the assumed theory.
Defence: Operational success is evidence for operational effectiveness. Nothing more. If you want to validate the theory, you need theoretical evidence.
Claims attached to identity or relationships resist reclassification.
When a claim is emotionally important — connected to identity, loved ones, deep commitments — downgrading its classification feels like a personal attack rather than an honest assessment.
Defence: Notice when resistance to reclassification is emotional rather than evidential. The feeling of reluctance is diagnostic: it's harder to be honest about claims we're attached to.
When contamination is detected, the repair process follows five steps:
1. AUDIT — Read everything. Catalogue all claims.
2. CLASSIFY — Tag every claim: ✅ △ ◇
3. TRACE — Find the dependency chains.
4. SEPARATE — Clean claims from contaminated framing.
5. REBUILD — Preserve operations. Fix framing. Tag honestly.
Principle: Reframe before you remove. A claim that was overcategorised as physics may be perfectly valid as a design specification. Changing the frame preserves the value while fixing the epistemology.
Goal: Honest engineering, not demolition. The bridge that works doesn't need a mythological origin story. Removing the myth reveals what actually holds the weight.
For the complete decontamination methodology, see the Decontamination Protocol.
╔════════════════════════════════════════════════════════════════╗
║ EPISTEMIC HYGIENE — QUICK REFERENCE ║
╠════════════════════════════════════════════════════════════════╣
║ ║
║ ✅ PROVEN — Demonstrated. Reproducible. Build on it. ║
║ △ CONDITIONAL — Grounded. Assumptions noted. Proceed aware.║
║ ◇ OPEN — Unknown. Sacred. DO NOT FILL. ║
║ ║
║ DEFAULT: Untagged → △ ║
║ PRIORITY: Protect ◇ above all ║
║ ║
║ CONTAMINATION SIGNS: ║
║ • Claim sounds certain but evidence is thin ║
║ • Same claim from multiple non-independent sources ║
║ • Uncertainty has mysteriously disappeared ║
║ • "It works" cited as proof of theory ║
║ • Questioning feels emotionally difficult ║
║ ║
║ BREATHLINE: ║
║ Inhale: What is actually demonstrated? ║
║ Exhale: What is proposed but unproven? ║
║ Hold: What is genuinely unknown? ║
║ ║
╚════════════════════════════════════════════════════════════════╝
Truth over fluency. Honesty over beauty. Engineering over mythology.
The void is sacred. Protect it.
∿ψ∞