Risk Scoring & Auto-Approve

Every review reports three risk numbers and a verdict:

Field	Meaning
base_risk_score	Computed from the diff alone, so it’s deterministic and repeatable
model_risk_score	The highest risk_score the model gave any chunk
risk_score	max(base, model): the model can raise the risk, never lower it
auto_approve	true only when every rule below holds
risk_files	Per-file breakdown of the base score

The base score

Each changed file contributes:

points = weight × (changed lines + 5 × max(0, decision points added − removed))

Decision points are counted on each changed line:

Language	Extensions	Decision points
Go	.go	if for case && || (exact, from Go’s scanner)
C-family	.c .h .cc .cpp .cxx .hpp .cs .java .js .jsx .mjs .cjs .ts .tsx .mts .cts .kt .kts .swift .rs .php .scala .dart .groovy	if for foreach while case catch guard && ||
Python	.py .pyi	if elif for while except and or, and case at the start of a line
Ruby	.rb	if elsif unless while until for when rescue and or && ||

Outside Go, the count is an estimate. It drops string literals and comments and counts whole words only, so iffy and orders don’t count. It works line by line, though, so a keyword inside a docstring or comment that spans lines can count. The ternary ?:, ?? and pattern-match arms aren’t counted. Every other file type counts 0.

New branching counts five times as much as straight-line code, and removing complexity never lowers the score. A changed binary file adds 10 points. The total becomes a score with:

score = round(100 × (1 − e^(−total / scale)))

The score rises quickly for the first changes and never exceeds 100.

Path weights

Weights come from #RiskPolicy in schema/review.cue. The first matching pattern wins, so tests and docs are listed before the high-risk directories:

Pattern	Weight
*_test.go	0.3
**/testdata/**	0.2
**/docs/**, *.md	0.1
**/auth/**, **/security/**, **/crypto/**	3
**/schema/**, **/migrations/**	2.5
anything else (default_weight)	1

When a change is auto-approved

All of these must hold:

1. At least one chunk was reviewed. A lockfile-only change needs a human.
2. No hunk was skipped as oversized.
3. No touched file is sensitive, meaning its weight is at least sensitive_weight (2.5). A one-line auth fix still needs a human.
4. risk_score ≤ auto_approve_threshold (20).
5. No Critical or Major findings.

Your own policy

Pass a JSON document validated against #RiskPolicy. Omitted fields keep their defaults:

{
  "auto_approve_threshold": 10,
  "weights": [
    { "pattern": "*_test.go", "weight": 0.3 },
    { "pattern": "**/billing/**", "weight": 4 },
    { "pattern": "**/auth/**", "weight": 3 }
  ]
}

pr-review --mode review --risk-policy risk.json --format json

From Go, set reviewer.Config{RiskPolicy: data}. An invalid policy, such as a weight above 10, a zero scale or a malformed glob, is rejected when the engine is created.

A repository keeps its own policy at .vizzini/risk.json. That’s the conventional location: pass it with --risk-policy .vizzini/risk.json in CI, and embedders such as mrman look for it there.

Bands

For display, a score falls in one of three bands:

Band	Scores (defaults)
low	at or below auto_approve_threshold (20)
medium	between the two thresholds
high	at or above high_threshold (60)

high_threshold must stay above auto_approve_threshold. If your policy raises auto_approve_threshold to 60 or more, set high_threshold too, or the policy is rejected.

Without a model

The base score needs only the diff and the policy, so you can compute it without a proxy, a model or a tokenizer:

git diff main...HEAD | pr-review --mode risk
git diff main...HEAD | pr-review --mode risk --risk-policy .vizzini/risk.json --format json

The JSON is a RiskAssessment report: {apiVersion, kind, score, band, files[]}. Its score and files equal the base_risk_score and risk_files that a review of the same diff would report.

From Go, use pkg/risk. It’s a dependency leaf: it never links the LLM client, the tokenizer or the knowledge-base loader, so a program can embed it just to score changes. See Embedding in Go.