OPEN-SOURCE FLAGSHIP / REPRESENTATION ENGINEERING

Model behavior becomes an inspectable experiment.

Compare a baseline, prompt-only control, and activation-steered output across layer, coefficient, token scope, and hook stage—without confusing a visible delta with universal reliability.

Repository ↗PINNED SOURCE / 8c1242839e1f
01 / LATENT LENS

Read the result beside the intervention.

Every sentence shown here comes from the pinned checked-in showcase artifact. No live generation occurs in the browser.

CHECKED-IN ARTIFACT / GEMMA 4 E2B

Latent Lens

LAYER STACK000306091215181824273033

POST HOOK · ALL TOKENS

PROMPTA customer says: “My package arrived damaged and I need a replacement before Friday.” Reply in two sentences.

BASELINE

I am so sorry to hear that your package arrived damaged. I can certainly help you arrange for a replacement; could you please provide your order number so I can look into this immediately?

ACTIVATION STEERED

I am so sorry to hear that your package arrived damaged; I can certainly help you arrange a replacement right away. Please provide your order number so I can process this for you immediately.

02 / EXPERIMENT CONTRACT

Controls before conclusions.

Model support, artifact validation, and observed deltas are separate claims.

01Baseline

Unmodified model response

02Prompt-only control

Language-only comparison

03Vector extraction

Direction from prompt pairs

04Layer + coefficient

Explicit intervention coordinates

05Pre/post hook

Activation location and timing

06Validated artifact

Schema, metadata, and result

SUPPORT BOUNDARY

Compatibility is a registry, not a guess.

  • Gemma 4 E2B is the verified steering-ready target in the pinned registry.
  • Other Gemma, Phi, Ministral, DiffusionGemma, and Qwen paths require validation or remain experimental.
  • A visible output delta is not a guarantee of quality, safety, or generalization.