Skip to content
synsema why?
Measured on 2026-10-16 · Synsema 0.6.44 · same data, same model on both sides

Why Synsema?

Five problems teams are solving right now with AI and data. Each one built twice, with the same data and the same model: once the usual way in Python, TypeScript or SQL, once in Synsema. Then measured.

Case 1

MCP servers that leak by accident

5 vs 0

confidential fields the model receives after one careless tool

One careless tool leaks every confidential field in Python. Synsema refuses it at runtime.

See the evidence
Python, typical: 55Python, typicalPython, hardened: 55Python, hardened0Synsema
Confidential fields handed to the model by the careless toolSame four tools, same catalog: cost, supplier, minimum margin, maximum discount, internal note.
Case 2

Running the code an LLM wrote

58×

faster sandbox: 34 ms in Synsema, 1,987 ms in Docker

An LLM writes the analysis, Synsema runs it sandboxed in 34 ms. Docker needs 2 s.

See the evidence
Python inside DockerPython inside Docker: 1,987 ms1,987 msSynsema run_programSynsema run_program: 34 ms34 ms
Time to run the generated code safely (median of 10 questions)Both sandboxes: no network, data read-only, time limit. Unsandboxed Python and DuckDB run in 29 ms but can read your keys.
Case 3

What a dependency can do

14 vs 0

third-party packages in the same 29-line ETL

The same ETL: 14 packages and 169 MB of other people's code in Python, zero in Synsema.

See the evidence
Python ETL14Synsema ETL0
Third-party packages installed for the same ETLEach square is one package, direct or transitive. 169 MB, 2,156 .py files and 99 native binaries on the Python side.
Case 4

Which invoices can be paid without a human

0 vs 7

of 7 risky invoices caught: the LLM's own confidence vs Jev

The LLM read all 60 invoices right, and its own confidence would have paid the duplicates. Jev caught 7 of 7.

See the evidence
LLM's own confidence0 of 7Arithmetic check0 of 7Jev, calibrated (judge)7 of 7
Invoices that needed a human, caught by each approval rule7 of 60 needed review: 2 duplicates and 5 totals corrected by hand, all of them read correctly.
Case 5

Training data you cannot hide

1 of 2

data files recorded by the usual manifest. Synsema recorded 2 of 2.

Someone added training data off the books. Python's manifest missed it. Synsema's signed receipt did not.

See the evidence
1/2Python, manual manifest2/2Synsema, engine lineage
Data files recorded, out of the 2 the training actually usedSame training run, same undocumented read of extra_data.csv (80 rows).

How we measured

  • Each case was built twice by the same person on the same day, with the same input data and, where an LLM is involved, the same model (deepseek-flash, temperature 0).
  • Numbers come from the scripts in the repository, run on a Windows 11 laptop with 12 threads and 7.4 GB of RAM.
  • Where Synsema loses or ties, the page says so. Limits are listed on every case.