Informal-fallacy detection

Status: core (Apache-2.0), heuristic, on by default in reasoning checks. A match means "look here", not "this is fallacious" — the detector flags the surface markers of fallacy families and reports candidates for review.

The step-logic verifier catches structural faults — a conclusion that does not follow, a circular step. This catches the named informal fallacies a fluent answer slips in: attacking the person, appealing to the crowd, forcing a false choice. It flags them from their characteristic phrasing.

Families detected

Family	Example marker
ad_hominem	"you're just biased", "he is incompetent"
appeal_to_authority	"because experts say so"
bandwagon	"everyone knows", "nobody believes"
false_dichotomy	"either with us or against", "only two options"
hasty_generalization	"this proves that all…"
slippery_slope	"will inevitably lead to", "next thing you know"
appeal_to_emotion	"think of the children", "you should be ashamed"
post_hoc	"rose after … therefore", "correlation … causes"

Circular reasoning is deliberately omitted — the reasoning-chain verifier already detects it by step overlap.

Usage

from director_ai.core import detect_fallacies, verify_reasoning_chain

detect_fallacies("Everyone knows this is best.").types     # -> ["bandwagon"]

result = verify_reasoning_chain("Step 1: … Step 2: everyone knows it works. …")
result.fallacies        # heuristic markers (FallacyMatch list)

Fallacies are reported on result.fallacies but, being lower-precision than the structural checks, do not affect result.chain_valid — they are an advisory signal alongside the high-precision step-logic and arithmetic verdicts. Disable the pass with verify_reasoning_chain(text, check_fallacies=False).

Polyglot backend

The marker scan runs through the Rust rust_detect_fallacies kernel (backfire_core::compute::detect_fallacies) with an identical pure-Python regex pass. The patterns are deliberately lookaround- and backreference-free so the Python re and Rust regex engines return the same matches in the same order — verified bit-for-bit in the benchmark and tests. The dispatch is purely a speed choice. See Rust Acceleration.

Measured

python -m benchmarks.fallacy_detection:

Metric	Value
Detection accuracy (labelled markers + clean controls, n=20)	1.00
Control false-positive rate	0.00
Rust ↔ Python parity	exact
Rust speed-up over Python	~5.6×

These numbers are on a controlled marker set — real prose is noisier, and the detector is a triage aid, not a proof. Genuine appeals to authority (a cited study) read like the fallacious kind; a match means "review", not "wrong". Numbers come from the committed benchmark and benchmarks/results/fallacy_detection.json.