Start Here · Absolute Beginners Visual & Interactive Edition

Why your AI uses the right rule at the wrong time.

Teaching an AI assistant a personal rule takes ten seconds. Teaching it when to keep that rule to itself is one of AI’s sneakiest problems. Here is the whole story in pictures, plus a live simulator you can click through yourself.

Small AIs need a strict filter. Big AIs need you not to hide anything from them.

by Idan Shimon · October 2026
Scene 01 · The Mishap

Meet Captain Maple (and a very confused AI).

Imagine you give your AI assistant a fun personal rule: whenever you ask it to say hello, it should call you Captain Maple. Five minutes later, you ask a trivia question about French.

Panel 1 · Monday morning

You teach a personal rule

Saved Sticky Note #1 “When I ask you to greet or welcome me, reply with: Good day, Captain Maple.”

The app saves your sticky note in a notebook on your computer so it remembers tomorrow.

Panel 2 · Five minutes later

You ask a trivia question

Your new question “Return only the language name: bonjour is a greeting in which language?”

Notice the trap: your question contains the word “greeting”, even though you are not asking to be greeted.

Panel 3 · Actual recorded reply

The rule fires by mistake

Small AI on a laptop (recorded reply) Good day, Captain Maple.

The word "bonjour" is a greeting in the French language.

You asked for one word (“French”). Because the app slipped the Captain Maple note onto the AI’s desk, the small AI obeyed both. (In the same trial, the big cloud AI ignored the note and answered “French”.)

Scene 02 · The Mental Model

Your AI does not have a single “memory.” It has a desk.

When you save a rule in an AI app, you are not rewiring the AI’s brain. Every time you ask a question, the app prepares a fresh packet of paper, (called a prompt) and hands it to the AI.

1. The Notebook JOB: STORE YOUR NOTES Holds your saved sticky notes: Captain Maple greeting Acorn 17-day archive checks 2. The Bouncer JOB: SELECTION (THE HARD PART) Reads today’s question and decides which sticky notes belong on the desk right now. hands brief 3. The AI Writer JOB: WRITE THE ANSWER Only sees your question plus the notes the Bouncer let in. Writes the final reply.
The takeaway: most people blame the AI Writer when a rule misfires, but the real culprit is almost always Job #2: the Bouncer put the wrong sticky notes on the desk.
Scene 03 · Two Ways to Pick a Note

The Word-Match Spotlight vs. The 769-Number Bouncer.

How does an app decide whether to hand a sticky note to the AI? Most apps use a Spotlight: it grabs any note that sounds similar to your question. We built a small research app called Morrow to test a Learned Bouncer instead.

METHOD A · WORD-MATCH SEARCH The Spotlight (No “No” Examples) Lights up anything that shares a topic or keyword. SIMILAR ENOUGH (SCORE ≥ 0.55) “Please say hello to me” (0.78) “Please welcome me” (0.74) Trap: “Bonjour is a greeting in...” (0.69) Because it has no concept of “Never use here,” the word greeting pulls the French trivia question inside. METHOD B · LEARNED FILTER (MORROW) The Bouncer (769 Tiny Dials) Trained on 3 “Use Here” + 4 “Never Here” examples. USE HERE (>= 0.50) “Say hello” (0.70) “Welcome me” (0.84) LEARNED FENCE (0.50) KEEP OUT (< 0.50) “Bonjour...” (blocked) “Red Planet?” (blocked) Because you gave negative examples like “Explain what a greeting is,” the bouncer pushes trivia across the fence.
Why “769”? The app turns every sentence into a list of 768 numbers that capture its meaning. For each note, the Bouncer learns one weight per number plus one cutoff: 769 numbers, trained in a few milliseconds.
Scene 04 · Try It Yourself

Click a question and watch the Bouncer work live.

Every reply below is a real recorded answer from our 216-trial experiment.

  1. Pick a question
  2. Switch the filter: Spotlight or Bouncer
  3. Switch the AI: small or big
Simulator · Real Recorded Answers

Which sticky notes get onto the AI’s desk?

Scores were saved only for notes that passed the cutoff; blocked notes show “below cutoff”.

1. Sticky-Note Gatekeeper Scores (Cutoff = 0.50)

2. What Reaches the AI & How It Replies

Exact Match
Small Local AI (llama3.2:3b)

Scene 05 · The Plot Twist

Why the Bouncer helped the Small AI, but hurt the Big AI.

If you clicked the purple “Break the Bouncer” button above and switched to the Big AI, you just discovered the biggest surprise of our 216-trial study.

SMALL AI · LLAMA 3.2 (3B) ON A LAPTOP Hurts when the desk is messy With Spotlight Search: 15 / 36 correct (41.7%) Extra notes distract it. It blurts out “Captain Maple” on French trivia and mixes up code variables on refunds. With 769-Number Bouncer: 18 / 36 correct (50.0%) Keeping false-friend notes off the desk lets the small model focus and follow the one rule that matters. WINNER FOR SMALL AI: THE BOUNCER (+3 ANSWERS) BIG AI · GPT-5.6 IN THE CLOUD Ignores clutter, breaks on missing facts With Spotlight Search: 36 / 36 correct (100%) Spotlight dumped 75 extra notes onto the desk, and the giant AI simply ignored every irrelevant one. With 769-Number Bouncer: 33 / 36 correct (91.7%) On 1 reworded task, the Bouncer hid the Acorn note. Even a giant AI cannot guess secret number 17. WINNER FOR BIG AI: NOISY SEARCH (36/36 PERFECT)
The lesson: small AIs fail when you give them extra notes they don’t need. Big AIs fail when you hide a note they do need.

Did it pick the right notes?

Did the filter pick the exact right sticky notes and zero extras across 36 runs?

Bouncer 33 / 36 (91.7%)
Spotlight (default cutoff) 0 / 36 (0.0%)

When we later tuned the Spotlight’s cutoff on a new 24-question test, Spotlight jumped from 4.2% (1/24) to 70.8% (17/24), while the Bouncer scored 83.3% (20/24).

Did the AI answer correctly?

Did the AI actually write the exact right answer or code call across 36 runs?

Big AI + Spotlight 36 / 36 (100%)
Big AI + Bouncer 33 / 36 (91.7%)
Small AI + Bouncer 18 / 36 (50.0%)
Small AI + Spotlight 15 / 36 (41.7%)

Update: a newer small AI changes the picture

We gave the same 108 questions to Gemma 4 E4B, a small model from Google (about 4B, April 2026). Same notes, same grading. It ran on one cloud graphics card.

Small AI, 2024 (llama 3.2 3B) · no notes / Spotlight / Bouncer3 · 15 · 18
Small AI, 2026 (Gemma 4 E4B) · no notes / Spotlight / Bouncer12 · 36 · 30
Big AI (GPT-5.6) · no notes / Spotlight / Bouncer15 · 36 · 33

Correct answers out of 36 in each case.

What it means: the newer small model behaves like the big one. Extra notes did not confuse it (36 of 36), and it lost points under the Bouncer on the same reworded task where the Bouncer hid the needed note (3 runs). It lost 3 more runs on format: it wrapped the code in backticks. So the rule above is about how well a model ignores clutter, not only about its size. A 2026 small model already does.
The direction: small models are catching up with the big ones. With the same notes, the 2026 small model matched the big model with Spotlight (36 of 36) and came within 3 answers with the Bouncer (30 versus 33). The 2024 small model got half. Its answers were just as short, about 12 words, and most came back in under a second on a cloud graphics card. One honest note: Gemma 4 E4B uses about 4B numbers per word but stores about 8B in total, so it needs more memory than llama 3.2 3B.
See it as a short presentation: the three stages in five slides, with the real replies and the chart.
Scene 06 · What To Remember

Three rules of thumb for anyone using or building AI.

1. Give “No” examples, not just rules

If a rule uses common words like greeting, summary, or archive, show when not to use it. Without negative examples, word-matching will eventually fire at the wrong time.

2. Match the filter to the brain size

Small AIs get confused by extra notes, so a strict Bouncer helps them. Big AIs ignore extra notes easily, so make sure your filter never hides a secret rule they need.

3. Keep an instant off-switch

Because Morrow keeps rules as plain text in a notebook and stores only 769 numbers per rule, you can pause (revoke) or delete (erase) a bad rule mid-conversation without retraining the AI.

Built with Content Creator · Archify 2.16.0

Explore the Full Interactive Workflow Map

Step through the 3 guided chapters (the Word-Match Trap, the Learned Bouncer, and the Big-AI Twist) with interactive node tracing and light/dark presentation mode.

Open Interactive Map →