Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.sigilix.ai/llms.txt

Use this file to discover all available pages before exploring further.

Every Sigilix finding goes through three rounds of confidence scoring before it reaches a PR. This page explains the math.

​
Per-specialist confidence (1-5)

Each specialist scores its own findings on a 1-5 scale:
ScoreMeaning
5Critical — broken correctness, security, or data integrity
4Important but not catastrophic
3Material maintainability concern
2Minor improvement (taste, naming, small refactor)
1Nit — surfaced for awareness
Scores 1-2 are advisory. Scores 3-5 may block the merge depending on category.

​
Cross-reference adjustment

When Core synthesizes, each finding’s confidence is adjusted by structural-provenance checks:
CheckAdjustment
Line valid in diffNo change
Line invalid (hallucinated)Drop entirely
Symbol exists in fileNo change
Symbol doesn’t existDrop entirely
Pattern matches claimed unsafe codeNo change
Pattern doesn’t matchDown-grade by 1 or drop
A score-5 finding can become score-3 or be dropped if its structural provenance is weak.

​
Agreement-based escalation

Multiple specialists flagging the same code is a strong signal. Core escalates:
InputsResulting score
1 specialist, score 3Score 3 (no change)
2 specialists agree, scores 3 + 3Score 4
3 specialists agreeScore 5
1 specialist score 4 + Core’s structural check confirmsScore 5
A finding posted at score 5 blocks the merge with a Critical badge. Score 4 is Warning. Score 3 is Info.

​
Score-1 advisory cap

To prevent comment dilution, Core caps the number of score-1 (nit-level) findings posted inline:
  • The top 5 score-1 findings (by alphabetical headline) are aggregated into a single “Advisory nits” line in the synthesizer summary
  • The remaining score-1 findings are recorded in telemetry but not surfaced
You can change the cap via sigilix.yaml.

​
Severity vs. score

Severity is a categorical label shown to the reader. Score is an internal numeric used by Core. The mapping:
  • Score 5 → Critical
  • Score 4 → Warning
  • Score 3 → Warning (or down-graded to Info if cross-reference is weak)
  • Score 2 → Info
  • Score 1 → Info (aggregated into summary, not posted inline)
The verdict (Approve vs. Request changes) is decided as:
  • Any Critical → Request changes
  • Otherwise → Approve

​
Tuning thresholds

You can tune what counts as Critical / Warning / Info via sigilix.yaml: 
thresholds:
  critical_min_score: 5  # default — only top-tier blocks
  warning_min_score: 4
  info_min_score: 2
  score1_cap: 5          # how many score-1 nits to aggregate
See the sigilix.yaml reference for the full schema.

​
Telemetry

Every finding’s lifecycle is recorded in telemetry, including:
  • Original specialist score
  • Cross-reference adjustments (and reasons)
  • Agreement escalations
  • Final severity
  • Whether it was posted, suppressed, or aggregated
Telemetry is internal — used for monitoring and prompt-engineering. Customer-visible analytics (per-repo finding counts, severity distribution over time) are part of the upcoming Pro/Max insights dashboard.

sigilix.yaml

Tune thresholds, disable specialists, configure ignore patterns.

Common Errors

What to do when reviews don’t appear, or appear with footnotes.