AI Slop Code Risk Monitor

Legacy Signal Bands

Distribution of the existing SynthScan pattern score. Bands are signal ranges, not authorship classifications.

Severity Breakdown

Counts pattern findings by configured severity; severity is not a probability or a security rating.

Top 20 Categories by Score

Highlights recurring pattern categories that may warrant a closer engineering review.

Language Distribution

Counts scanned repositories by their reported primary language.

Highest Pattern Signal

Lowest Pattern Signal

Legacy pattern-score adjustments

These modifiers define the existing SynthScan signal score. They do not establish that a match is AI-authored or risky in a specific codebase.

Density Clustering Factor

Clusters increase the legacy pattern score; they are not definitive authorship evidence.

Contextual Isolation Multiplier

The legacy score weights comments and executable code differently according to the configured context multipliers.

Architectural Diminishing Returns

Reduces the contribution of large per-file match counts to the legacy score.

Legacy Signal Formula & Time Factor

Each pattern match receives its configured severity weight and a context multiplier. The sum of the weighted findings forms the raw score, normalized by scanned lines to produce a comparative pattern-signal score:

The adjusted score further cascades through a temporal discount factor based on the repository's most recent commit metadata. The adjustment uses the repository's last-push timestamp. It is a coarse repository-level heuristic, not evidence about when any finding was written; use raw and adjusted scores as signals, not AI-involvement probabilities.

Detection Layers

The scanner combines five detection and control layers. Findings remain heuristic indicators and may need repository-specific interpretation.

1. Line-level pattern matching — Configurable text and regex patterns with severity and context weights.
2. Multi-line block checks — Checks selected documentation and exception-handling structures.
3. Python AST checks — Flags selected structural patterns, including dead code, unused imports, empty handlers, and deep nesting.
4. Cross-file repetition — Finds repeated or similar docstrings in full scans; not included in diff mode when line provenance is unavailable.
5. Suppressions — Inline ignore markers and `.synthscanignore` support.

Severity Classification Matrix

Severity coefficients contribute to the legacy pattern score. Severity describes the scanner's configured weight, not confidence that code is AI-generated or a measure of defect risk.

Legacy Signal Bands

Study unit and provenance

Unit of analysis: one repository scan. Current coverage is 7135 payload-bearing repositories out of 7135 successful scans.

Engine version(s): 2.0.0
Scoring version(s): heuristic-density-v1
Payload schema: 1.0.0

Each finding records category, severity, relative file path, line, stable identifier, and mapped vector IDs. The ID is the first 16 hexadecimal characters of SHA-256 over file, line, category, and pattern text. Unknown categories remain findings, increase unmapped_category_count, and are not silently assigned to the first three vectors.

Finding points before vector scoring

Base severity weights are LOW = 1, MEDIUM = 2, HIGH = 5, CRITICAL = 10. Line-pattern matches are multiplied by context: COMMENT = 1.5, CODE = 1.0, STRING = 0.5. Block/AST findings use their configured scores.

Line-pattern clustering multiplies matches by 1.5 when at least three fall within ten lines of an anchor. If a file has more than 20 accumulated line/block/AST matches, matches after the top 20 by score are multiplied by 0.5. Cross-file repetition findings are added after this per-file modifier.

The inline ignore marker suppresses line-pattern matches; .synthscanignore excludes files from scanning. In diff mode, findings are limited to added-line locations; cross-file repetition is omitted because its source line cannot be attributed reliably.

Vector scoring equations

For repository r, let N be scanned lines (added lines in diff mode), and let P_v be the sum of finding points mapped to vector v. For cognitive ownership, supply-chain, and over-engineering:

D_v = 1000 × max(0, P_v) / max(1, N)
S_v = round(100 × (1 − exp(−D_v / 20)), 2)

The saturating exponential maps non-negative point density to 0–100; parameter 20 is a fixed design constant, not a fitted coefficient.

Overall repository index: round((S_ownership + S_supply + S_overengineering + S_review) / 4, 2).

Review-effort proxy and bands

Estimated minutes sum a fixed cost per finding: LOW = 0.5, MEDIUM = 1.5, HIGH = 3, CRITICAL = 5. Unknown severity defaults to 0.5 minutes.

T_review = Σ minutes(severity_i)
S_review = score(2 × T_review, N)

The same density transform is used by score(points, N) above with points = 2 × T_review. Bands: LOW < 25; MEDIUM 25–<50; HIGH 50–<75; CRITICAL ≥75.

Dashboard aggregation

For each vector, the dashboard reports the unweighted arithmetic mean of repository-level vector scores for repositories that include that vector, plus the number of repositories and findings contributing. Means are not LOC-weighted and should be compared only across compatible scoring versions.

mean_v = round(Σ_r S_(r,v) / n_v, 2)
      coverage = 100 × repos_with_payload / successful_scans

Legacy score is separate

The legacy score remains raw_score / scanned_lines × 1000 (zero when no lines are scanned). The monitor computes adjusted_legacy = legacy_signal × temporal_factor; factors are 0.05 before Nov 2022, 0.40 before Jun 2023, 0.75 before Jan 2024, otherwise 1.00.

This temporal factor is applied to the legacy adjusted score only; it does not alter Risk Payload v1 vector scores. It is based on repository metadata, not the date of each finding.