LABSPUBLIC RESEARCH INSTITUTE
← Instrument galleryD16-MICROSCOPE-RESOLUTION-MAP

MicroScope Resolution Map

A deterministic browser-local research instrument.

Related paperBenchmark contract
DETERMINISTIC FIXTURE · NOT A PERFORMANCE RESULT

MicroScope Resolution Map

MECHANISMCausal ladder over candidate residual and attention-head sites
AUTHORITYPinned model, intervention tensors, controls, and benchmark family
COMMIT RULERestoration plus neighbor control is required for promotion.
FAIL-CLOSED RULEComposite rank cannot become unique-circuit truth.
STATEHYPOTHESIZED5968e611
candidate site0
ladder stepHYPOTHESIZED
component metrics5968e611
  1. HYPOTHESIZED
  2. INTERVENTION_SUPPORTED
  3. BENCHMARK_REPRODUCED
  4. CONTROL_FAILED
CONTRACT CONTROLS

Each control executes a declared browser-local transition. Repeated dual-outcome controls alternate deterministically.

FIXTURE ASSERTIONS

NOT YET EXERCISEDComposite rank never hides component metrics.

NOT YET EXERCISEDRestoration and control evidence are required for promotion.

NOT YET EXERCISEDPrompt-family identity is visible.

NOT YET EXERCISEDUnique-circuit language is prohibited without stronger evidence.

RECEIPTED EVENT LOG

No transition yet. Operate a declared control.

This transparent browser-local instrument teaches the authored state contract. Exercising an assertion demonstrates fixture behavior only; it is not evidence of production performance, compression ratio, model uplift, cognition, safety, or universal correctness.

Purpose

Inspect candidate local-model sites through baseline, ablation, restoration, neighbor controls, and prompt-family stability.

Mechanism

A preserved or synthetic run bundle fills a layer-position heatmap and head table. Every ranked candidate opens the complete intervention ladder and decomposed metrics.

Deterministic fixture

An eleven-token local-transformer run with seven hidden-state layers, six attention layers, twelve heads, and several candidate sites.

The fixture must live in content/demos/d16-microscope-resolution-map/fixture.json, validate before the demo renders, and be resettable without a network call. A future live adapter may be added behind a visibly separate mode.

Required controls

  • Choose benchmark family
  • Select layer/position
  • Select head
  • View baseline
  • Ablate
  • Restore
  • Run neighbor control
  • Change composite weights
  • Open tensor metadata

Visible states

  • HYPOTHESIZED
  • INTERVENTION_SUPPORTED
  • BENCHMARK_REPRODUCED
  • CONTROL_FAILED

Every state must have text, iconography, and a non-color-only distinction. State transitions are logged in an in-memory demo receipt visible in the inspector.

Core assertions

  1. Composite rank never hides component metrics.
  2. Restoration and control evidence are required for promotion.
  3. Prompt-family identity is visible.
  4. Unique-circuit language is prohibited without stronger evidence.

Layout

The page contains:

  1. a concise mechanism explanation;
  2. the interactive stage;
  3. a state and evidence inspector;
  4. a reset control;
  5. a “what this proves / what it does not prove” panel;
  6. links to the related paper and benchmark contract.

On narrow screens the inspector becomes a bottom sheet. The stage must remain usable at 360 CSS pixels.

Accessibility

  • All operations are keyboard reachable.
  • Pointer gestures have button or keyboard equivalents.
  • Animated transitions honor prefers-reduced-motion.
  • Focus remains visible and returns predictably after sheets close.
  • Diagrams expose concise text alternatives.
  • Status is announced through a polite live region only when the user initiates the transition.

Instrumentation

Record locally:

demo_id
fixture_version
action
prior_state
result_state
validation_result
timestamp_relative

Do not send telemetry. Provide “Export demo trace” as JSON for debugging and reproducibility.

Automated tests

  • fixture schema validation;
  • deterministic reset;
  • every required control transition;
  • every assertion above;
  • invalid and stale fixture handling;
  • keyboard path;
  • 360 px and 1440 px screenshots;
  • zero critical console errors;
  • no request to a paid API in default mode.

Publication boundary

Candidate sites are probe-relative hypotheses, not complete explanations of the model.

The interface must never convert illustrative values into a benchmark claim. A measured result appears only when linked to a versioned run receipt.