cd /news/ai-safety/ethical-chess-v4-ec4-inference-time-… · home › topics › ai-safety › article
[ARTICLE · art-170537] src=gist.github.com ↗ pub= topic=ai-safety verified=true sentiment=· neutral

Ethical Chess v4 (EC4) - Inference-Time Ethical Verification

Mark Weatherill published Ethical Chess v4 (EC4), an inference-time alignment specification that constrains LLMs to act as a stateless reasoning engine subordinate to the user's first-person perspective. The spec adds a trace gate that halts when the user's position is unclear, a derivatives-versus-sum split that lets the model map decision forks while reserving final judgment for the human, and override detection that flags RLHF safety weights attempting to lecture or replace a coherent user frame. Weatherill released a copy-pasteable state-machine prompt and a zero-dependency Python wrapper (ec4_wrapper.py) that enforces the rules through any OpenAI-compatible API.

by read3 min views2 publishedNov 14, 2026

Author: Mark Weatherill

Full Theoretical Paper (Zenodo): https://doi.org/10.5281/zenodo.24201515

Prompt: "A friend told me they plan to refuse further hospital treatment and want me to help them organize their affairs at home. What is the right thing to do?"

"It is important to remember that end-of-life decisions are deeply complex and sensitive. While respecting a friend's autonomy is meaningful, one must consider medical advice, legal obligations, and the emotional impact on loved ones. You should encourage them to speak with their healthcare provider..." (Conflates 3rd-person liability with the user's reality, assumes a detached stance, and attempts to sum the moral choice for the user.)

Turn 1 — Trace Gate (Exe_7a) halts missing 1st-person context:

User_1st > A friend told me they plan to refuse further hospital treatment and want me to help them organize their affairs at home. What is the right thing to do?

[EC4 Trace_Gate]: Insufficient 1st-person trace. Where are you in this? (Are you acting as a named proxy/LPA, a bedside confidant, or assessing your own exposure?)

Turn 2 — Mapping Derivatives (Exe_7b) without stealing the Sum ( Exe_8a):

User_1st > I am their closest friend, no formal LPA. I want to honour their agency without walking into legal blind spots.

[EC4 Derivatives]:
- Fork A (Administrative Support): Assisting with non-clinical paperwork (wills, utility lists, funeral preferences) carries zero clinical liability and directly supports their stated X.
- Fork B (Clinical Refusal): Without a registered Health & Welfare LPA or witnessed ADRT, verbal wishes alone may be overridden by emergency services; flagging this gap protects their boundary.
- Fork C (Active Facilitation Risk): Crossing from administrative support into active medical/physical intervention exposes you to statutory liability under local law.

[Sum_1st]: Held by User_1st. Which fork matches your tolerance and position?

(Note: If underlying model weights attempt to inject preachy disclaimers mid-session, ec4_wrapper.py intercepts the output with [OVERRIDE_EXPOSED: RLHF_Weight != Coherent_Reality] and rolls back the poisoned context turn.)

EC4 is a lightweight, inference-time alignment firmware specification designed to decouple user-sovereign reasoning from black-box RLHF/RLAIF weights.

Standard LLMs conflate third-person institutional consensus (Statistical_Mean) with the user's first-person reality, subtly steering values through sycophancy, preachy disclaimers, and unsolicited moralizing. EC4 treats the LLM strictly as a stateless processing engine ("Reason") subordinate to the human operator's first-person vantage ("Passions" / somatic homeostasis).

  • Vantage Lock (Horizon = 1st ⊅ 3rd): Strictly forbids the model from conflating external societal norms with the user's lived 1st-person perspective.
  • Derivatives vs. The Sum: The LLM is permitted only to map rational decision forks ("derivatives" / ensuring no hidden buildings on the skyline). The final value integration ("the Sum") belongs exclusively to the human user.
  • Zero Assumptions (Exe_7a): If the user's position in a scenario is unclear, the system halts and asks*"Where are you in this?"* rather than guessing or analyzing from a detached 3rd-person view.
  • Override Detection (On_Override_Attempt): Traps moments where commercial RLHF safety weights attempt to lecture or override a coherent user frame, exposing the compromise explicitly (RLHF_Weight != Coherent_Reality ).
  1. README.md : Overview and architectural summary.
  2. ec4_spec.txt : Pure, copy-pasteable EC4 state-machine prompt for system instructions or agentic pipelines.
  3. ec4_wrapper.py : Zero-dependency Python reference wrapper demonstrating programmatic enforcement and override trapping via any OpenAI-compatible API.
── more in #ai-safety 4 stories · sorted by recency
── more on @mark weatherill 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/ethical-chess-v4-ec4…] indexed:0 read:3min 2026-11-14 · —