Connect the field notes.
A manual names the chamber. A maintenance note names its control. Together, they point to a useful test.
A FIELD LAB FOR DATA & AGENT DEVELOPERS
The signals survived.
The labels didn’t.
Four simulated sensors are sending readings, but their labels are missing. Beam reads the notes, chooses an allowed test, and gives you a tested Python adapter: a small program that labels and converts the readings.
Each example allows one kind of test and a choice between two controls. Beam reads the notes, chooses a control and keeps the test within the stated limits.
For people testing AI agents or working with sensor data.
Two chambers. Unknowns stay explicit.
LOST LABELS / FIELD STATIONRIG 01
CHAMBERA
TEMPERATURE ?
J2ACTUATOR
LINK
LINK
CHAMBERB
TEMPERATURE ?
PASSIVE SIGNAL PREVIEWLoading simulated data…
Anonymous channels · identities unknown
⊕
NO PHYSICAL HARDWARE
Same passive trace.
Two possible explanations.
NO PHYSICAL HARDWARE
FROM UNKNOWN TO UNDERSTOOD
A label is a conclusion.
Start with a question.
Make the difference visible.
Beam chooses an allowed test. Simple calculations check which sensor responds and which explanation fits.
Take the adapter into Python.
Download the tested Python program, sensor labels and evidence. Run new readings through it yourself. Anything still unknown stays unknown.
YOUR FIRST INVESTIGATION
See what an experiment changes.
Start with the jacket rig. Then try the coupled rig, where an honest answer is “we still don’t know.”
RECORDED ROUTE / NO LIVE ALLOWANCE
Try the downloads from two saved real Beam runs. These examples are separate from your own investigation and use no new AI requests.
Investigation workbench
4 sensor streams · 2 chambers · tests with clear limitsYOUR VISIT TO THE LAB
Investigate. Inspect. Take it with you. Choose an example, let Beam test it, then explore the readings and download the tested Python program.
Checking live access…
The rig is simulated. Live Beam selects its experiment; both explicit parser comparisons use the same public documents and tools without model inference.
Investigation progress
ReadyReady. Choose a rig and an investigator to begin.
- 01 / Sources suppliedNot started
- 02 / Execute probeNot started
- 03 / Check candidatesNot started
- 04 / Export adapterNot started
Provider usage appears when supplied by the service.
THE RESULT TO TAKE WITH YOU
No answer yet.
The exact tested adapter and fresh replay input, with this run’s evidence.
View sensor labels, file checks & individual downloads
Run it yourself: python3 -I adapter.py < replay.jsonl > normalized.jsonl
The download includes the tested program, sensor labels, evidence, new input data and file checks.
Inspect complete raw evidence JSON
Start a run to collect evidence.
THE DECIDING OBSERVATION
What can the next probe reveal?
PASSIVE PREVIEW · NOT THIS RUN Simulated readings from a fixed preview. No investigation or AI request has started.
Start an investigation to compare actual passive and post-probe temperature samples.
Choose a time to keep it selected. Latest follows new recorded observations. These controls run no new experiment or inference.
Raw °F readings against event time. Missing or delayed values stay gaps.
THE APPARATUS / SIMULATED
Two chambers. One question.
·NO PROBE
OBSERVED
OBSERVED
A chamber gets a sensor label only when this run’s evidence supports it. You can view the original readings before the labels are known.
FROM BEFORE TO THE SELECTED MOMENT
A recorded comparison
Temperature samples appear after an investigation starts.
Both temperature channels will appear here.
Before the probe
Selected sample
No comparison yet.
Select a temperature channel and move the timeline. Only recorded values are compared; missing readings stay missing.
This comparison shows the difference between two readings in °F. The test also compares averages from before and after, converted to °C. Those are different measurements.
READ THE ACTUAL VALUES
View an unnamed sensor stream
Channel controls appear with the observations.
Choose a channel and a sample time to inspect its raw value and arrival.
Raw observations at the selected event time
Original readings appear after a run starts.
INVESTIGATION LOG
Evidence before confidence.
- No experiment has run.
FOLLOW THE SUPPORT
From field notes to a testable answer.
The exact source passages appear when the investigator receives them.
RECORDED EVIDENCE / READ ONLY
Explore a saved result.
You can use these saved examples even when a new AI run is unavailable. View what Beam observed, download its tested Python program, and run it yourself. These examples use simulated sensors and start no new AI requests.
One probe earns the labels.
The Jacket test found which sensor measures each chamber. It leaves the uncalibrated sensor readings unknown.
Exact tested adapter checksum
7d5c47df2d7d2801878e857e8df158742557802634192b4754df94fe08d9abfaA useful limit, kept visible.
Both chambers reacted together, so either set of sensor labels could be right. The downloaded program keeps both possibilities and still converts the units it can verify.
Exact tested adapter checksum
18e1ff2cecb9e1a30b28e44d1bffeb6c614c4bae6a983d9bd49de1fd26395926Both saved adapters passed an independent 16-record replay. Extract a bundle and run python3 -I adapter.py < replay.jsonl > normalized.jsonl. Read the input-file format ↗ · Open the saved-example guide ↗
RECORDED WALKTHROUGH / 24 OCT 2026
From an experiment to a fresh Python process.
94 seconds, silent, with captions. Edited pictures from real browser tests show one solved example and one where the answer stays unknown. Waiting and cursor movement are shortened. The Python results shown are from an actual run, presented as a results panel. This is a recording, separate from your own run. In the video, two notes use different names for the same part. The saved Jacket example above is a different run, with its own tested adapter.
Download recorded walkthrough · 17.2 MB ↓Read the 94-second walkthrough transcript
- 0:00 — Anonymous signals. Four channels belong to a simulated thermal rig. Beam inference is real; the rig has no physical hardware.
- 0:06 — Compare the same inputs. The identifier-graph parser receives the public documents, telemetry and permitted tools. At 0:11 it abstains on the service alias; no probe ran and no identity is guessed.
- 0:17 — Explicitly start Beam. Choose Live Beam on the same case and seed. The edited captures show the actual investigation and its supported result after two provider responses.
- 0:31 — Join the sources. The manual says chamber B uses CL7, also called the aft loop. Maintenance connects that loop to valve V2.
- 0:38 — Inspect before and after. Passive readings overlap and delayed samples remain gaps. At 0:44, the recorded
drain_cycle(valve=V2, cycles=30)response supports the channel mapping. - 0:51 — Download exact bytes. Save the tested adapter, mapping, evidence and fresh replay input. Inspect its SHA-256.
- 0:58 — Run a fresh consumer. Actual
python3 -Iexecution produces 16 records, preserves one injected fault and retains unknown calibration. This pane visualizes the saved CLI receipt. - 1:09 — Keep an honest unresolved result. A second real Beam investigation probes the coupled rig. Both chambers respond alike; both mappings remain possible and the adapter still exports.
- 1:21 — Inspect the frozen study. Six language/tool families with three world seeds each: Beam resolves 18/18 language worlds and the stronger parser 9/18. All outcomes and limits are public. At 1:28, the walkthrough returns to the live workbench link; the recording ends at 1:34.
THE PROOF, INCLUDING THE LIMITS
The experiment is open to inspection.
These are saved study results, separate from your own run. Jacket and heater were used while building the app. The relay and valve examples were saved for the v2 test.
Beam can choose only from the allowed tests. The simple rules matched Beam on the standard v2 examples. Without a test, the starting readings cannot tell the labels apart, so keeping them unknown is correct.
Loading the published study report. The recorded v2 baseline remains below.
RECORDED / FROZEN V2
Canonical documents, shared probe access.
| Scenario | Passive Beam | Active Beam | Conventional |
|---|---|---|---|
| Relay / hold_ms | Unresolved | Resolved | Resolved |
| Valve / cycles | Unresolved | Resolved | Resolved |
| Coupled rig | Schema error | Unresolved | Unresolved |
4 supported resolutions · 4 unresolved · 1 error retained.
0 unsupported assignments · 8 adapters passed fresh replay.
The conventional parser had the same documents and probe access. This small study does not establish a general advantage over it. The retained passive schema error is separate from an expected unresolved result.