Detection Algorithm

A technical specification of the Behavioral Signal Aggregator (BSA) powering asshole-gc. Version 1.0. Last updated 2026-03-23.

Overview

asshole-gc uses a multi-pass Behavioral Signal Aggregator (BSA) to evaluate subjects across 14 weighted dimensions. The pipeline has four stages:

1. Signal collection — passive observation, no active probing
2. Dimension scoring — 14 weighted behavioral axes, rolling 72-hour window
3. Consecutive day counting — the core threshold mechanism
4. False positive suppression — context normalization before flagging

A subject is only flagged when all four conditions are met: threshold exceeded, confidence above minimum, false-positive score below ceiling, and at least one manual review pass (the --dry-run step).

Signal Collection

Signals are collected passively across the subject's observable behavioral surface. The system does not intercept private communications or require instrumentation. Observable surface includes:

• Communication tone and word choice in shared channels
• Response patterns: frequency, latency, selectivity
• Credit attribution (or lack thereof)
• Reaction to criticism vs. reaction to praise
• Behavior under low-stakes vs. high-stakes conditions
• Treatment of people with less power than them
• Follow-through on commitments

Signals are timestamped and windowed. The rolling window is 72 hours by default, configurable via --window. Each signal has a weight, a context tag, and a decay function.

struct Signal {
    timestamp:   DateTime,
    dimension:   Dimension,   // which of the 14 axes
    weight:      f32,         // 0.0 – 1.0
    context:     Context,     // WorkStress | Personal | Deadline | Normal
    raw_text:    Option<String>,
}

14 Behavioral Dimensions

Each dimension is scored 0.0 (clean) to 1.0 (severe). Weights reflect how reliably each dimension predicts chronic vs. situational behavior.

BSA Scoring

The BSA computes a weighted aggregate across all 14 dimensions, normalized by the context weight of each signal. Context weights reduce scores during known stress periods (deadlines, personal crises) to avoid false positives from situational behavior.

fn bsa_score(signals: &[Signal]) -> f32 {
    let windowed = signals.filter(|s| s.within_window(72.hours()));

    let raw_score: f32 = windowed
        .group_by(|s| s.dimension)
        .map(|(dim, sigs)| {
            let dim_score = sigs.iter()
                .map(|s| s.weight * context_factor(s.context))
                .sum::<f32>()
                .min(1.0);
            dim_score * dim.weight
        })
        .sum();

    // Normalize to [0, 1]
    raw_score / DIMENSION_WEIGHT_SUM
}

Context factors:

Consecutive Day Counter

The consecutive day counter is the core threshold mechanism. A "bad day" is defined as any 24-hour period in which the subject's BSA score exceeds 0.55 (configurable via --day-threshold).

The counter increments on bad days and resets on genuinely good days. A good day requires a score below 0.30 — not just below threshold. This prevents gaming via minimal compliance.

Default threshold: 3 consecutive bad days. This was chosen because:

• 1 bad day — noise. Everyone has them.
• 2 bad days — pattern forming. Worth noting.
• 3 bad days — this is who they are right now. Time to act.

False Positive Suppression

The FP suppression layer runs after BSA scoring and before flagging. It applies three filters:

1. Observer bias correction — if the observer reporting signals has a known conflict with the subject, dimension weights are reduced by 20%.
2. Mirroring check — if the subject's flagged behaviors closely mirror behaviors exhibited by the reporting observer, confidence is reduced. Some people bring out the worst in each other.
3. Single-source guard — signals from only one observer require a higher threshold (4 consecutive days instead of 3) before flagging.

The FP rate with --dry-run review is estimated at <2%. Without review, it is estimated at ~8%. Always use --dry-run.

Grace Period Logic

When --grace-period is passed, the subject receives a single clear notification before collection is executed. The grace period is 7 days by default (--grace-days N).

During the grace period, the counter continues running. If the subject demonstrates genuine improvement (score drops below reset threshold for 3+ consecutive days), the flag is cleared and they are moved to Status::Monitoring instead.

Worked Examples

Example A: Dave (Engineering)

Example B: Casey (Slack)