Neural Noir — a deductive mystery engine

Play a real case

The deduction, in your browser

These cases were generated by the Python engine (python demo.py 7) and committed verbatim — suspects, clues and the exact per-clue elimination sets. The buttons below run the same constraint logic the Python solver uses: a suspect survives only if every clue implicates them. Interrogate to see who each clue clears, then accuse.

This page is static (GitHub Pages) and runs no LLM — it replays committed, engine-generated cases. The real generator, constraint solver and uniqueness guarantee live in the Python package in this repo; the optional LLM hook only restyles text and never touches the deduction.

The engine

From a seed to a provably solvable case

generator

Seeded cast

Suspects share overlapping attributes so no single clue gives anyone away.

clues

Culprit-true constraints

Clues are drawn from the culprit's own attributes, so a true clue never clears the culprit.

solver

Constraint propagation

Keep a suspect only if every clue implicates them; the survivors are the answer.

guarantee

Verify, then accept

The generator runs the solver and only keeps a case when exactly one survivor remains — the culprit.

herrings

Harmless red herrings

True facts phrased to match everyone — flavour that can never break uniqueness.

optional

LLM flavour hook

If a key is set, an LLM restyles the prose. It never decides clues or the verdict.

Why it's real

The uniqueness guarantee is tested, not asserted

Exactly one solution. For hundreds of seeds and cast sizes 3–8, the solver returns one suspect and it equals the planted culprit.
Clues do real work. Removing any necessary clue makes the solver return more than one suspect — the case goes ambiguous.
Minimal clue sets. No redundant real clue survives generation; each one is load-bearing.
Red herrings stay harmless. They implicate everyone; dropping them all leaves the same unique culprit.
Reproducible. The same seed yields an identical mystery every time.

Run it yourself: pip install -r requirements.txt then python -m pytest -q (855 passing) and python demo.py.