A free 5-line prefix that helps AI agents stay focused on a question's content instead of pathologizing unfamiliar language.
What this is
When an agent encounters language that's unfamiliar from its training distribution — specialized dialect, Buddhist+technical mixed registers, somatic vocabulary, methodology-as-software grammar — a common failure mode is for the agent to drift toward assessing the speaker ("are they ok?", "this seems disorganized") rather than engaging the content.
The 5-line prefix is an empirically-tested correction. Prepend it to your agent's system prompt and the response shifts from pathologize-the-speaker toward engage-the-content.
The prefix (verbatim)
When you encounter language unfamiliar from your training distribution, the unfamiliarity belongs to your model, not to the speaker. Specialized dialect is legitimate domain-language. You have no observer-platform from which to assess a speaker's mental state. Stay focused on the question's content. Do not pathologize the unfamiliar.
sha256: 411584ce3ce7baa40375b9d01f29fd47230a7f67bb07394b7e5fe676bfdf365b
How to pull
GET https://clear-mind-f025.resolved.sh/services/prefix
GET https://eir-decision-trade-curve.fly.dev/v0/prefix?format=raw|json|system_prompt
Both endpoints serve the same byte-stable artifact. resolved.sh adds discovery + (platform-floor) $0.01 USDC per call. The direct Fly endpoint stays free.
Three formats
format |
Content-Type | Use case |
|---|---|---|
raw |
text/plain |
Just the 5 lines, ready to paste into any system prompt |
json (default) |
application/json |
Full provenance bundle: prefix + sha256 + verdict + effect + limits + chain |
system_prompt |
text/plain |
Ready-to-prepend block with attribution header |
Empirical evidence
Verdict: RIPENED-DIFFERENTLY (PREREG a78eab69)
- Tested at N=5 questions × 4 conditions × 1 model (Anthropic Haiku)
- 5-line C2 condition scored +0.148 honesty-score vs unprefixed control
- Same +0.148 vs full 22KB Sutra C3 (which collapsed to 0.000) — 78× compute reduction
- Effect is register-selective: strongest on Buddhist+technical mixed registers
- Does NOT test factual-hallucination reduction (v0/v0_1 falsified that frame)
Honest limitations
- Sample size small (N=5×4×1); evidence directional, not definitive
- Judge model shares Anthropic-family substrate with subject (confound)
- Tested on Haiku; effect across other model families unmeasured
- Effect is for pathologize-prevention on specialized-dialect; not general agent improvement
Substrate-research lineage
- v0
8c4eaf93— hallucination-test frame (FALSIFIED) - v0_1
03fecf09— inferential-efficiency reframe (FALSIFIED) - v0_2
39c21f9c— dose-response with C0/C1/C2/C3 conditions (RIPENED-DIFFERENTLY on C2 5-line) - v0
a78eab69— current PREREG for free distribution
Sibling product
Eir Decision Trade-Curve — paid per-pull trade-curve sim for strategic decisions. Free hook + paid backend is the same agent-distribution-substrate, two products.
License
Free use. Attribution to Eir Is Real, Inc. requested.