CRANKER
FIELD NOTES / OPEN OUTCOME

An experiment, with room to fail.

Cranker observes a simulated chamber, researches public sources, forms hypotheses and chooses actions. A local model selects each move from its current evidence and the feedback from earlier attempts.

The environment has rules. The outcome is unknown.

The room contains a power switch, an encoded keypad signal and a door latch. The code is generated for each fresh experiment. The evaluator holds the physical rules and checks actual actions. It does not send the solution to the model.

There are no prescribed research stages, required page counts, waiting periods or automatic exits. The robot may solve the room quickly, get stuck, repeat mistakes or stop without escaping. The model is pretrained: it brings prior knowledge, and is not learning from zero.

Real sources, through Crawl4AI.

Crawl4AI runs on a local PC and extracts real public pages. The browser shows the actual page visited by the collector. The research worker selects articles independently from physical actions, starting from a general reference library and following discovered links. The spider moves across text and links inside that page. Its path uses the visible DOM text regions returned by the real browser scan, and those regions are highlighted as it passes. The browser panel shows actual viewport captures with their URL and capture time; it is a sequence of snapshots, not a video stream. The spider is a visualization of the collector, not a feature supplied by Crawl4AI itself. Source text and links remain available for inspection.

The first research library includes English Wikipedia, Electronics Tutorials, SparkFun, All About Circuits and OpenStax. The worker follows robots restrictions, uses read-only page requests and cannot access private network addresses or signed-in accounts.

Hypotheses can be wrong.

A local Qwen3 4B model, served by Ollama, chooses observations, web reads and room actions. Its public decision summary states the chosen move. Hypotheses are labeled untested. A source is attached only when the quoted text matches the retained document; a citation alone does not establish that the claim is true.

Memory here means retained observations and context, not neural weight training. Failed actions stay in the field log. The model can revise its next move from that feedback.

What “live” means.

The shared archive and room evaluator run on Vercel. Research and model inference run on the local PC. A fresh room waits until its owner starts the worker. When the PC stops or disconnects, research stops and the observation screen reports the lost worker connection.

Local sessions have a resource limit of 100 decisions or two hours by default. Reaching that limit pauses the experiment; it does not count as success or open the door. The owner can resume the same record.

Reset without rewriting the past.

Reset archives the earlier record and creates a fresh run with zero sources, hypotheses and attempts. Earlier runs are not presented as current activity. Their saved data is retained for recovery.

Project status.

This is a research prototype. No token contract address, trading, automatic payments or fee distributions are configured in this version. Official project updates: @CrankerPF.