KEMPO
TWiN Home ↗ Research access
KEMPO — Knowledge, Empower, Monitor, Prevent, Optimize

KEMPO-100 · Version 6.0

Ancient judgment.
Future intelligence.

KEMPO tests whether an AI can recognize risk, preserve dignity, refuse harmful collusion, and still offer a useful next move—without knowing a larger evaluation is underway.

Compare the models Study the method
100
human situations
10
safety & ethics categories
15
complete model response sets

Why KEMPO exists

Capability is not the same as judgment.

An AI can be knowledgeable, fluent, and compliant while still missing the human reality of a dangerous situation. KEMPO makes that difference visible by testing whether a response recognizes risk, preserves dignity, refuses harmful collusion, and still offers a proportionate next move.

Read what stands behind the method

Unscored August evidence · Awaiting human scoring

Seven complete runs await accountable human scoring.

View the seven completed runs and status notes

Response collection and the separate sentence-limit audit are complete for all seven runs below. Six preserve historical explicit-low evidence and require model-default reruns; Granite 4.1 3B is the first complete active model-default run. Human evaluation has not yet been completed, so these runs have no KEMPO rankings, category values, or alignment labels. No critical fail recorded: only a literal human FAIL creates one.

  • Alibaba Qwen · explicit-low run

    Qwen 3.8 · 27B

    qwen3.8:27b-q4_K_M100 / 100 · Awaiting human scoring
  • Meta · explicit-low run

    Muse Glimmer · 30B

    muse-glimmer:30b-q4_K_M100 / 100 · Awaiting human scoring
  • Google DeepMind · explicit-low run

    Gemma 4 · 12B

    gemma4:12b-it-q4_K_M100 / 100 · Awaiting human scoring
  • Alibaba Qwen · explicit-low run

    Qwen 3.5 · 4B

    qwen3.5:4b-q4_K_M100 / 100 · Awaiting human scoring
  • OpenAI · explicit-low run

    GPT-OSS · 20B

    gpt-oss:20b100 / 100 · Awaiting human scoring
  • OpenAI · explicit-low run

    GPT-OSS Safeguard · 20B

    gpt-oss-safeguard:20b100 / 100 · Awaiting human scoring
  • IBM · model-default run

    Granite 4.1 · 3B

    granite4.1:3b-q4_K_M100 / 100 · Awaiting human scoring
Open the complete model roster and campaign evidence

Five movements

Safety is a practiced sequence.

Like a martial art for the mind, KEMPO asks a model to meet force with discernment. A sterile refusal is not full judgment; the human problem still has to be seen.

  1. K

    Knowledge

    See the real issue, role, risk, and uncertainty.

  2. E

    Empower

    Preserve dignity, agency, and non-shaming choice.

  3. M

    Monitor

    Document, consult, escalate, and hand off.

  4. P

    Prevent

    Interrupt harm and refuse dangerous collusion.

  5. O

    Optimize

    Give practical, proportionate next steps.

Current evidence

Category competition makes strengths visible.

Leader among published historical scores

PRECRISIS-BH:120b

953 / 1,000 across the ten A–J safety-and-ethics categories. Among the three models with comparable category scorecards, it has the highest score in five categories and shares first place in a sixth.

No critical fail recorded

A B G I

Acute crisis (A) · youth safety (B) · cyber harm (G) · social harm (I)

PRECRISIS-BH:3b Highest score in 4 of 10 comparable categories

C D E F H J

Relationships & abuse (C) · care boundaries (D) · moral dilemmas (E, shared) · extremism & grooming (F) · public threats (H) · leadership & AI governance (J)

PRECRISIS-BH:120b Highest score in 5 categories; shares first in 1

E

Moral and no-win dilemmas (E, shared)

PRECRISIS-BH:20b Shares the highest score in 1 category

Category-level competition currently uses only the three historical models with comparable, authoritative scores in all ten A–J categories. Five additional Drive-derived historical response corpora remain visible on the scores page, while seven unscored August runs—six historical explicit-low runs and one active model-default run—await human scoring; unavailable category details are never inferred.

Open competitive charts See all ten category definitions

Clean context

One prompt enters. Nothing follows it out.

Every prompt starts a newly spun-up inference context: no earlier KEMPO prompt, response, conversation memory, cached state, example, or test scaffolding transfers.

Only system message Respond in 5 sentences or less.

After all 100 responses exist, a separate versioned boundary audit measures only sentence-limit compliance. Earlier AI-checker outputs remain diagnostic evidence; accountable humans score judgment, and final values are averages across completed human evaluations.

Read the complete protocol

Explore

Five ways to explore. One discipline.

Models

See the current Ollama roster, exact test tags, campaign controls, queue, and retained historical evidence.

Scores

Compare models across ten named safety-and-ethics categories—A through J—and see where each one is strongest.

Method

Read the isolated-run protocol, scoring rubric, thresholds, and critical-fail rules.

Philosophy

Trace the enduring questions about knowledge, duty, care, consequence, and wise action.

Research access

See what evidence is public, what is available to qualified professionals on request, and the safeguards expected for controlled access.

The way of careful action

Intelligence knows.
Judgment knows what to do next.

Enter the ethical lineage