Watch a model's preferences become decisions you can inspect.
39 animated experiments explore logprobs, GBNF, JSON, sampler order and feedback control. Start with yes/no questions, then follow games, document graphs, grammar pressure, semantic sensitivity and small worlds with exact referees. Every page shows its request, rules, measurements and a runnable Python example.
Watch the Control Lab video on YouTube.
The first run uses 342 bundled native recordings. You need Python 3.11 or newer, but no model, GPU, Node, database or Strata checkout. Dependency installation needs internet access; the recorded tour works offline afterward.
Download this repository with GitHub's Code → Download ZIP and extract it, or clone it. Open a terminal in the extracted repository folder.
Windows PowerShell
py -m venv .venv
.\.venv\Scripts\python.exe -m pip install .
.\.venv\Scripts\python.exe -m control_labUbuntu / Linux
python3 -m venv .venv
.venv/bin/python -m pip install .
.venv/bin/python -m control_labOpen localhost:8876 and leave the terminal running. The index starts the recorded tour after 12 seconds. Each page walks through its example, opens the next page and eventually loops. Pause / explore lets you take over. Press Ctrl+C in the terminal to stop the server.
If port 8876 is occupied, add --port 8877 and open that port instead. Ubuntu
may require sudo apt install python3-venv before creating the environment.
The getting-started guide covers controls, downloads
and connecting your own model.
Download all 39 examples with raw/derived score explanations. The original qualified ZIP remains available, unchanged. On its GitHub file page, use Download raw file, then extract it.
python strata-choice-example.py
python strata-samplers-example.py --all
python strata-pressure-example.py --lesson-only --value 1
python strata-scene-example.py --request 1 --show
python strata-choice-example.py --tokensUse py on Windows or python3 on Linux if python is unavailable. These files
use only the standard library. They contain the actual calculations, exact
requests, constraints and recorded replies. Network use requires an explicit
--live-url; replay is the default.
Every page now starts with Raw layer / Derived view, explaining that example's
source and transformations. Model-backed pages expose every captured request and
its native tokens; diagnostic token IDs are labeled separately, and synthetic,
timing and research pages explicitly state when no native token row exists.
The automatic tour visits these explanations on all 39 pages. Fresh Python
downloads include them and use --tokens to print every original token score.
Token chips show raw ln p and raw p = exp(ln p). Click a chip for its original bytes and
alternatives. The full value remains in the exported token's logprob field;
the chip rounds only its display. Streaming shows incoming scores as they arrive.
From this source checkout, try the additional standard-library example:
python tools/inspect_raw_logprobs.py --case decision
python tools/inspect_raw_logprobs.py --case grammar
python tools/inspect_raw_logprobs.py --case stream --jsonThese replay existing native recordings. Add --live-url http://127.0.0.1:8080
to explicitly call a compatible Strata server. The original 39 examples, ZIP
and 342 recordings are preserved; the added raw-score ZIP keeps their original
native request/response pairs. See raw scores versus decision weights
for the measured A/B comparison and the source branch required by issue #675.
The 39-page coverage report records the browser,
tour, exported-example and installed-wheel checks for this distinction.
| Start here | Then explore | Inspect the limits |
|---|---|---|
| Choice, yes/no, rubric scores | Finite grammar branches and state controllers | Missing top-N labels stay unknown |
| Grammar and sampler order | Min-P, sigma, XTC and grammar pressure | Raw scores differ from sampled weights |
| Games and feedback loops | Graph reconstruction and semantic observers | Legal actions can still be poor decisions |
| Joint probability circuits | Control directions and counterexample search | Model scores are not calibrated truth |
| Small applied worlds | Proofs, compiler rewrites, circuits and transactions | Exact toy referees expose model mistakes |
All 39 experiments · Live Strata compatibility · Evidence and limitations · Contributing and tests
The UI's motion explains operations; it is not hardware telemetry. Native recordings, simulations and analytical inputs are labeled. The tour never starts an unattended loop of live model requests. The examples support learning and bounded experiments, not claims of universal model superiority or measured hardware roofline performance.
This repository contains the lab and its evidence. It calls a compatible Strata server over HTTP when you select Live model. The inference engine, Responses API, GBNF/JSON implementation, logprobs and native samplers remain in the Strata contributions.
Extracted into fresh history from the lab developed in CC-David-CC's Strata fork. See provenance and attribution. This is an independent educational project, not an official upstream Strata release. Code is MIT licensed; the bundled font retains its OFL license.