Question 1

Does it send raw data to Jev?

Accepted Answer

No. Only field names, types and sample values are sent as a DataProfile; actual cell values stay local and the write operations are executed by Python scripts, not the model.

Question 2

Can it run without Jev?

Accepted Answer

Yes. The default is pure System 2 with any OpenAI-compatible LLM; Jev is an optional System 1 engine enabled with --system1, and --escalate chains both engines.

Question 3

How do I set up SignalChain and point it at a model?

Accepted Answer

Python 3.10 or newer, then pip install pandas openai openpyxl; categorical-variable analysis additionally needs an R installation with tidyverse, jsonlite and openxlsx. Copy config.example.py to config.py, which is gitignored so keys stay out of version control, and fill SYSTEM2_API_KEY, SYSTEM2_BASE_URL and SYSTEM2_MODEL for the LLM plus SYSTEM1_API_KEY and SYSTEM1_MODEL for Jev. Leaving SYSTEM1_API_KEY empty degrades cleanly to pure System 2.

Question 4

What exactly does the model return, and how is it checked?

Accepted Answer

Scripts extract a DataProfile with field names, types, sample values and missing rate, hash it into a fingerprint and assemble the prompt. The model answers in a forced format: a scene code from S0 to S5 and a signal_sequence such as IIGADDNT, one character per field. The program validates that output, translates the sequence into operations and executes them column by column.

Question 5

When does SignalChain call the model again instead of reusing the cache?

Accepted Answer

signal_cache.json is keyed by a fingerprint of the field structure and is partitioned per decision engine, so alternating System 1 and System 2 runs keep separate caches rather than overwriting each other. A partition is invalidated by a config change such as routing tables, the operation registry or standard column names, by a change in field count or names, or by sample values shifting materially. Deleting the file clears every partition, and run_categorical.py --no-cache forces reclassification.

Question 6

What happens when a Jev field comes back low-confidence?

Accepted Answer

In pure System 1 mode a low-confidence field falls back to the conservative default X, which is pass_through: no guessing and no modification, with the whole classification compressed into a single round trip. Passing --system1 --escalate hands only the uncertain fields to System 2; without that flag System 2 takes no part at all. The README reports Jev at 6.6 to 12.7 times lower token cost because the answer is already one character.

Question 7

Which field types and scenes does SignalChain handle out of the box?

Accepted Answer

Thirteen signal codes map to concrete operations: I id, G gender, A age, D department, N drug name, C diagnosis code, T date, M amount, E email, P phone, L log level, R coordinates and X other, driving steps such as normalize_gender, extract_age, validate_icd10, split_currency and validate_phone. Five scenes plus S0 unknown are defined, each supporting a subset, and a registry mechanism lets you add new field types and handling logic.

SignalChain

Highlights

Questions about this entry

More Jev resources like this

From the Jev community