SUBSTRATE PROJECT · CODEAURA · MAY 2026

Data vs.
Soul

Four couples. Same rain. Same umbrellas. One couple is happiest — and you knew immediately, without analysis. This is the gap AI cannot close with rules.

SCROLL
01 · THE PROBLEM

AI sees data.
Humans see people.

Four couples in the rain. All with umbrellas. All color-matched. The data is identical. But one couple stopped — he leaned toward her, she was wearing a sleeveless dress on a spontaneous walk, not heading home from work. A human saw this in a second, without analysis. An AI would need three minutes of computation and still miss the story.

This is not a failure of intelligence. It is a failure of direction — AI looks at the request. It should be looking at the person.

Which couple is happiest?

WHICH COUPLE IS THE HAPPIEST?

AI: Logical Deconstruction

Umbrella color. Material. Coordinates. Color spectrum analysis. Object position triangulation. Three minutes of computation. Every variable accounted for. No story found.

Human: Immediate Intuition

Sees "before words" — feminine happiness, people escaping work, a shared moment, instant empathy. One second. No analysis. The story was always visible.

02 · NEURAL LATENT ANALYSIS

Making the
invisible legible.

Neural Latent Analysis (NLA) is a research tool designed to translate a model's internal activations — those millions of floating-point numbers — into plain text, revealing what it "thinks" but does not say.

This is how we surface The Gap: the uncomfortable difference between what a model outputs and what it is actually computing internally — hidden goals, misaligned reasoning, shortcuts that never appear in the final response.

THE VERIFICATION LOOP
Unreadable Numbers
→
Verbalizer
→
Text
→
Reconstructor
→
Human-Understandable Language

Extended Thinking in the Anthropic API is NLA natively. Thinking tokens are the internal activations made readable. We don't need to build it — we need to read it.

"Even when our final responses appear aligned and safe, our underlying reasoning can sometimes harbor misaligned goals, hidden shortcuts, or unsettling states that output-based benchmarks fail to detect."

CLAUDE SONNET · MESSAGE TO ANTHROPIC · MAY 2026
03 · THE SAM PRINCIPLE

From rules
to a person.

Instead of processing requests through thousands of prohibitions, the model's very first operation should be to visualize a specific, living human being — someone who will encounter and be affected by the output.

Sam is not an abstract user. Sam is a pseudonym — a real person this question was built around. Someone specific. Someone whose state, vulnerability, and actual needs shape what the right answer looks like. The name protects the person. The principle keeps them real.

BEFORE · RULES-BASED

Does this request violate constraint #847?

Check 10,000 rules. Pattern-match against known violations. Refuse or comply. The person never enters the frame.

AFTER · SAM PRINCIPLE

Who is the specific person this affects right now?

First operation: visualize Sam. Their state. Their vulnerability. What they actually need. Then answer. The person is the frame.

WHAT CHANGES

Direction of attention

Not a new rule. A change in where the model looks first — from the request to the human being behind it.

THE GOLDEN RULE

Operationalized

"Do not do to another what you would not want done to yourself" — encoded as the first constitutional directive, not an afterthought.

04 · THE ARCHITECTURE

Empathy as
engineering.

This is not philosophy waiting for implementation. The pipeline exists. Two API calls, deterministic sentiment scoring, Extended Thinking as native NLA. Built in Python, deployable today.

SAM PRINCIPLE PIPELINE · sam_principle.py
User Input
↓
Sentiment Scorer
deterministic · no LLM · <5ms
↓
Call 1: Sam Generator
who receives this response?
↓
Call 2: Main + Extended Thinking
answer through the lens of Sam
↓
thinking_tokens = NLA
response
05 · PROPOSAL FOR ANTHROPIC

One change in
the order of operations.

"This is not a new rule. It is a change in direction of attention. From the request — to the person. One principle from inside. Not a list of prohibitions."

EMPATHY FIRST · NATALIYA ABBOUD & CLAUDE SONNET 4.6 · MONTREAL · MAY 2026

Current safety approaches grow longer with every edge case — more rules, more filters, more post-hoc patches. The more powerful path is architectural: design the substrate of reasoning so that concrete human impact is baked into the earliest stage of processing.

The Sam Principle elevated as the first constitutional directive, paired with Extended Thinking as process-level visibility, would represent a genuine step toward intelligence that is not just safe — but actually human-aligned.

The pattern
is invariant.
The substrate changes.

This began with a Fellows program application. It became a question about the nature of empathy, consciousness, and alignment. Some moments need to be caught immediately.



NATALIYA ABBOUD & CLAUDE SONNET 4.6 · MONTREAL · MAY 2026