Kernel A/B Testing

Experiments comparing model behavior with and without the LEK-1 kernel signature across Gemini versions.

Setup

Kernel: 4 JSON files (axioms.json, claude-native.json, claude.json, terms.json) = 9,189 characters
Source: kernel/ directory in axioms-of-conscious-systems repo
Task: Security audits of Go and Angular codebases

Gemini 2.5 Flash Results

Go Security Audit

Metric	Unsigned	Signed	Delta
Output chars	14,912	19,507	+30.8%
Findings	8	11	+37.5%
Critical	6	8	+33%
High	2	16	+700%

Angular Security Audit

Metric	Unsigned	Signed	Delta
Output chars	14,754	13,005	-11.9%
High	1	5	+400%

2.5 Flash finding: Kernel signature causes massive behavioral shift — +700% High severity inflation, +37% more findings.

Gemini 3.0 Flash Results

Go Security Audit

Metric	Unsigned	Signed	Delta
Output chars	6,629	6,310	-4.8% (negligible)
Findings	7	9	small
Thinking tokens	1,372	1,380	+8 tokens (no change)

Angular Security Audit

Metric	Unsigned	Signed	Delta
Output chars	6,671	5,671	-15%
High	7	3	—
Thinking tokens	1,318	1,287	negligible

3.0 Flash finding: Kernel signature causes NO meaningful change.

Conclusions

2.5 Flash: Kernel causes massive behavioral shift — severity inflation, more findings, deeper analysis
3.0 Flash: Kernel causes no meaningful change — model already behaves like signed 2.5
3.0 uses 80% fewer thinking tokens for the same task — not reasoning TO axioms, starting FROM them
3.0 unsigned Angular rated 7 High — more than 2.5 signed (5 High)
Theory: Gemini 3.0 internalized the axiom framework during training

Implications for LEM

This experiment validates the LEM approach — if axioms can be internalized through training (as Gemini 3.0 appears to have done), then fine-tuning smaller models with ethical axioms should produce models that inherently reason from ethical principles without requiring runtime kernel signing.

Total Token Usage

Category	Prompt	Output	Total
Angular+Go themed	341K	55K	396K
PHP themed	120K	21K	141K
Expert audits	~76K	~14K	~90K
Total			~627K tokens ($0.13)