HalluciGuard reads what your AI coding assistant actually says — not just the code it writes. Hedging, backtracking, apologies, context loss and test manipulation are detected by a deterministic pattern engine in under a millisecond, and surfaced as a severity-ranked alert before a single line is executed or written to disk.
This runs the real pattern library in your browser — same categories, same severity rules, same exact-phrase explainability as the native engine. Nothing is sent anywhere.
The same compiled pattern engine is reachable from the filesystem, the HTTP API layer, the TLS network stack, batch jobs, telemetry pipelines, and directly inside agent code — so detection can't be bypassed by the tool being monitored.
Each row below runs the full pattern library against a synthetic response of the given size, right now, in your browser. Throughput stays flat as length grows — a 5,000-line architectural explanation costs the same per character as a one-liner.
Point the batch analyzer at a glob of session files and it emits a CSV: one row per independent chat, split automatically on context-reset commands, with pattern breakdown and severity for trend analysis and compliance audit.
| Session file | Chat | Msgs | Patterns | Top category | Severity | Detected |
|---|
Every alert cites the exact matched phrase and the rule that fired. No probabilities, no black box, reproducible by anyone.
Treats the assistant as a black box and reads only observable output — so it works against proprietary third-party services.
In embedded mode the engine returns a binary halt signal, letting an agent reject its own response before touching the filesystem.
A hedged, backtracking, confused answer can be perfectly valid syntax. Static analysis sees nothing wrong, and ML classifiers answer with a probability you can't audit.
| Aspect | HalluciGuard | Linters / compilers | ML-based detectors |
|---|---|---|---|
| Target | Assistant natural language | Human-written code | Generated text |
| Detection focus | Behavioural uncertainty | Syntax & semantics | Semantic plausibility |
| Timing | Pre-execution | During / after writing | Post-generation |
| Latency | <1 ms | 100 ms – seconds | 100–500 ms (GPU) |
| Explainability | Exact matched phrase | Rule ID | Black-box score |
| Determinism | Fully deterministic | Deterministic | Non-deterministic |
| Infrastructure | ~30 MB RAM, no GPU | CPU | GPU / hosted API |
| Catches “Sorry for the confusion” | Yes — CRITICAL | No | Maybe |