Three-Lane Probe

Probe

Question

Planned provider calls this run: 1. Handshake doubles each lane. Timeout is 30s per lane, covering handshake plus task.

Provider

Who answers

Mock writes a synthetic receipt. That is not a Grok answer and not an injector test.

How the pack is attached

Pack rides in the system slot. One provider call per lane.

Tamper DEMO bytes (CRLF)

Tamper fails hash and length. Packed and matched-length lanes block; the plain lane can still run. Bytes are never repaired.

Sealed artifact

DEMO fixture integrity

A label is not enough. The accepted workbench checks the manifest, full hash, byte length, encoding, and line-ending policy. Any UTF-8 file renamed as QOFT is still invalid here if those seals fail.

QOFT injector

Not shipped in this public viewport. The local workbench also requires scoped hash 83b31740773a…,80,140 bytes, LF-only, and 14,224 o200k tokens. Those bytes stay in the private harness.

full 6ee2dc6f39bbc24aee92656aa74f71020caf7d4ca4c0897a3b14c0b245242db6

How to read this

Same question. Three setups.

Think of three render passes of one shot: beauty, a LUT, and a blank LUT of the same file size. If the packed pass looks like the beauty pass, the pack did not change the picture. If packed looks like the stone pass, the difference may only be extra length.

01 · ClearPlain
The question only. No extra instruction pack.
BASELINE
02 · FrontPacked
A sealed instruction pack, then the same question.
FULL_INJECTOR
03 · FogMatched length
The same token count as the pack, filled with the word “stone,” then the question.
NEUTRAL_LENGTH_CONTROL

Clear

Plain

BASELINE

queued

The question only. No extra instruction pack.

2×12 prompt fingerprint—

Stored hash: —
Wire hash: —
Seal: PENDING
Context tokens: 0

Waiting for this lane.

Front

Packed

FULL_INJECTOR

queued

A sealed instruction pack, then the same question.

2×12 prompt fingerprint—

Stored hash: —
Wire hash: —
Seal: PENDING
Context tokens: —

Waiting for this lane.

Fog

Matched length

NEUTRAL_LENGTH_CONTROL

queued

The same token count as the pack, filled with the word “stone,” then the question.

2×12 prompt fingerprint—

Stored hash: —
Wire hash: —
Seal: PENDING
Context tokens: —

Waiting for this lane.

Committed artifacts

Event rail

0 recorded0 blocked0 hash matches

Run the probe to write hashed call records. Errors store the exception type only — never the provider message.

Classification

DEVELOP — public Grok viewport inspired by Prompt Engineering Workbench v0.2.0. Not protocol-conformant with the local Python harness. DEMO fixture only. QOFT injector is not shipped. Evaluation remains null.

Evidence chain

What this can claim

Messages reaching the provider are hashed with messages-json-v1 immediately before the request. A mismatch refuses the send.
Tokenizer is JS o200k_base via gpt-tokenizer — same encoding family as the workbench, not the bundled hash-verified file.
Failed calls record exception type and an end timestamp. They do not store xAI error text.
One timeout covers the whole lane attempt. Evaluation is null. This is not a QOFT activation test.

Preview hashes

Assembled locally

BASELINE: blocked or assembling
FULL_INJECTOR: blocked or assembling
NEUTRAL_LENGTH_CONTROL: blocked or assembling