Skip to main content
Four parallel domain specialists produce a lot of overlap. Without a synthesizer, that overlap shows up as 40 redundant comments on the PR. Harmonia is the synthesizer that turns four streams into one signal — and the place where the believability pipeline does most of its work.

​
Where Harmonia sits in the believability pipeline

Believability in Sigilix is architectural, not a prompt. A finding cannot post unless it earns its way through the 5-stage believability pipeline: evidence → provenance contracts → refute/execute → proof-tier receipts → memory. Harmonia is the stage where evidence is checked against provenance contracts, refuted-or-confirmed, and stamped with a proof-tier receipt (VERIFIED / GROUNDED / MODEL) before anything reaches the PR. The four stages below are how Harmonia does that.

​
The four stages

​
1. Collect

Specialists submit findings in a structured shape: 
{
  "specialist": "security",
  "path": "src/api/checkout.ts",
  "line": 142,
  "category": "security",
  "severity": "critical",
  "headline": "missing CSRF verification on POST /checkout",
  "body": "...",
  "citedCode": "app.post('/checkout', handler)"
}
The specialist field is the role name (logic, security, performance, tests). The public-facing brand names — Metis, Argus, Iris, Eunomia — are rendered into the final review comment by Harmonia. Each finding must carry the evidence it stands on (the cited code it quotes). Findings missing required fields are dropped at this stage.

​
2. Cross-reference

Harmonia performs structural-provenance checks against the source code to suppress hallucinations. This is the grounding gate: a finding that cannot be anchored to cited code does not earn a GROUNDED receipt and is suppressed or demoted.
  • Line-validity check. Does the finding’s path:line actually exist in the diff? Hallucinated line numbers are dropped.
  • Symbol-resolution check. Does the function/variable referenced in the finding actually exist in the file? Hallucinated identifiers are dropped.
  • Pattern-match check. For security findings, does the claimed unsafe pattern (e.g., “passes user input to SQL template”) actually match the code? Pattern-mismatches are dropped or down-graded.
Findings that pass all three checks proceed. Findings that fail are silently suppressed — they never reach the PR.

​
3. Calibrate

Harmonia then performs deduplication and severity calibration: Deduplication. Overlapping findings (same path:line from multiple specialists) are merged into one. The merged finding’s body draws from each specialist’s contribution; the severity is the maximum of the inputs. Severity calibration. Each finding’s severity is recalculated:
InputsFinal severity
1 specialist, low confidenceInfo
1 specialist, high confidenceWarning
2+ specialists, agreementWarning or Critical
Critical-tagged + structural check passedCritical
Critical-tagged + structural check skepticalWarning (down-graded)
Advisory cap. If a flood of low-severity (Info) findings exists, Harmonia surfaces the top few inline and aggregates the rest into a single “advisory nits” line in the summary. This prevents review-comment dilution. Proof-tier receipts. Each surviving finding is stamped with a tier pill — VERIFIED, GROUNDED, or MODEL — recording how strongly it was witnessed. See Confidence Scoring.

​
4. Render

Harmonia writes the final comment. The shape:
  1. Harmonia summary — what was reviewed, how many findings survived, what verdict
  2. Inline findings — anchored to specific path:line, tagged by specialist + severity + proof-tier receipt
  3. Suggested patches — included where Harmonia’s structural check confirms a clean fix is in scope
Sigilix is inline-first: anchorable findings post on the line they quote, not buried in a review body.

​
Failure modes

​
Specialist provider failures

If one specialist’s model is overloaded (503) or exceeds its size-scaled budget, Sigilix retries once on the primary, then fires the cross-provider fallback. Primary + fallback run on independent infrastructure so a same-family provider outage can’t silence the same role twice. If all attempts for a specialist fail, the specialist’s findings are skipped — but Harmonia still synthesizes from the remaining specialists. The verdict is still posted with a footnote: _3 of 4 specialists succeeded._

​
Stale-head guards

If the user pushes a new commit while Sigilix is mid-review, the old review would be stale. Sigilix has two stale-head guards:
  • Before fan-out. If the PR’s head SHA changed since the webhook fired, abort.
  • Before posting. If the head SHA changed during specialist execution, abort and let the new webhook fire its own review.
Both guards prevent stale reviews from racing to post.

​
Submit failures

If GitHub rejects the inline-anchor positions in the review payload (typical 422 for a bad line number), Sigilix falls back to an anchorless review with all findings rolled into the body. The user sees one coherent review, just without inline anchors. This recovers verdicts that would otherwise be silently lost.

​
Why this beats single-agent review

A single-agent reviewer has no synthesis stage. It produces raw output and posts it. There’s no deduplication, no cross-reference, no calibration, no proof-tier receipt. Every false positive ships. Every redundant comment ships. Every hallucinated line number ships. Harmonia is the difference. The four-stage pipeline is what makes Sigilix’s reviews believable and readable.

Believability Pipeline

The five gates a finding earns before it can post.

Confidence Scoring

Proof-tier receipts (VERIFIED / GROUNDED / MODEL) and the grounding gate.