VINDEX3

III · THE EVIDENCE

THE EXECUTION IS A RECORD.

A computation should survive the computation. A canonical recording preserves what the execution wrote, where it wrote it and which declared reader inspected it. No model needs to be running to inspect those values again.

ONE EXECUTION, LEFT AS EVIDENCE · GEMMA 3 4B

“The capital of France is”

Follow the recorded readout for token #9079 — “ Paris” in the capture study.

L0RECORDED PROBABILITY · 0–100%L33
Token rank
1
Lens probability
69.74%
Carrier norm · L2
40683.05
Applied-write norm · L2
3227.98

Rank 1 at layer 24. Almost 100% at layer 26. Back to 80% at the final layer. A readout can strengthen, then weaken. The record lets you return to each write.

READ THE PROVENANCE & SCOPE

Recorded 2026-09-20 · production CPU · final prompt position 5 · head-v1 lens. 408 carrier writes, 204 readouts, 1,446 events. Hardware model is not recorded.

Stored BF16; the production image also pins Q8 requantisation for 103 operands. A representation label alone does not describe the arithmetic.

Prompt text, model name and token spelling come from the capture study. The JSONL stores token IDs and model identity. Probability is exp(recorded log p); this is a lens readout, not generated output or a causal claim.

Run: gemma3-4b-france-lens
File SHA-256: 12cdf4d9d06f5675233d71556c2158702cff88dd89efe42af72ab9dfb345dbd4

In Observatory, choose “Open Paris recording”, or open the downloaded JSONL. This exhibit reads a checked excerpt; Observatory validates and replays the original record. Neither runs inference.

OBSERVE → RECORD → REPLAY

Every write has an address.

THE SAME EXECUTION, MADE INSPECTABLE
  1. 01 · EXECUTEThe canonical traversal

    Observation subscribes to the executor’s carrier writes. It does not reconstruct a second forward pass for display.

  2. 02 · RECORDIdentity, values, receipt

    Run provenance, basis identity, ordered events and receipt bindings travel in the native JSONL.

  3. 03 · REPLAYObservatory

    Validate the file, revisit a position and depth, and inspect recorded lens values. No inference is required.

The receipt binds the event log and declares completeness. It does not independently authenticate the producer, verify the model’s identity or prove agreement with another implementation.

Read the native recording contract →

If a record identifies a mechanism, you can ask what changes when you change it.

INTERVENE · A SEPARATE EXPERIMENT

Change a head. Resume the computation.

A declared head intervention changes the mixed head value inside attention, before the gate, output projection and post-attention normalization. Downstream execution resumes from that change. Subtracting a displayed contribution after the fact is a different operation.

COUNTERFACTUAL PROTOCOL · NOT A RESULT FROM THE PARIS RECORDING
  1. BASELINEFreeze the comparison

    Name the model, representation, prompt, addresses, reader and decision rule.

  2. MANIPULATIONZero · scale · replace

    Declare the layer, query head and positions. Record the declaration, donor provenance when needed, and actual firings.

  3. CONSEQUENCECompare the continuation

    Measure the change against the baseline and controls. A firing receipt alone establishes no causal effect.

The interface is observe --intervene. Head intervention currently covers CPU softmax decode; MLA, conv-QKV, batch prefill and Metal head intervention remain outside that scope. The Paris recording above contains no intervention.

The head-intervention contract →

THE CLAIM CAN BE TESTED.

WHAT KIND OF THING DO WE KNOW?

  1. DECLARED

    The graph says this operator exists.

    A description of the program, before it runs.

  2. OBSERVED

    This carrier was written.

    A value captured at a named execution boundary.

  3. ATTRIBUTED

    This head contributes under this readout.

    Descriptive support depends on the reader and normalization contract.

  4. INTERVENED

    Changing this head changed the continuation.

    Requires a recorded manipulation and a controlled comparison.

  5. TESTED

    This effect held within the frozen experiment.

    A scoped result. Its controls, subject and decision rule travel with the claim.

These are different claims, each needing its own evidence. Observation does not establish attribution; attribution does not establish a counterfactual. Even an intervention needs controls before it supports a causal conclusion.

THE EVIDENCE CONTRACT →

The recording preserves an execution. The evidence ledger preserves the scope of what was learned from it.