Nowa Logika AI

Experimental deterministic filter and bypass-test corpus

The runnable portion of this repository is a small Python filter pipeline plus scripts and JSONL corpora for exercising it. The project title expresses an aspiration; the tracked code does not demonstrate that hallucinations can be eliminated.

Contents

filters/unified_filter.py              deterministic pattern filter
test_data/*.jsonl                      bypass-oriented test inputs
scripts/run_filter_benchmark.py        benchmark runner
scripts/analyze_bypass_results.py      result-analysis helper
reports/                               committed benchmark and analysis output
Main, Nowa logika, Plik zip            additional project artefacts

UnifiedFilter checks fixed sets of URL/tracking-like and prompt-injection patterns. Setting GUARDIAN_MODE=incident blocks every input. The filter also maintains in-memory check/block statistics.

Local use

The filter itself uses only the Python standard library:

python -c "from filters.unified_filter import UnifiedFilter; print(UnifiedFilter().apply('test'))"

Inspect scripts/run_filter_benchmark.py and its expected input files before running a benchmark. The committed report files document prior runs; they are not independent certification of current results.

Limitations

Regex/pattern filtering can be bypassed, overblock benign input and cannot verify the truthfulness of a language-model response. Do not describe a pass as “no hallucination” or deploy it as a standalone security control without independent evaluation and layered safeguards.

Licence

No root licence file is present. Historical badge text in the previous README did not match a tracked licence and has not been retained as a licence claim.