The Ensemble

The most direct way to build AI code review is to run a single model with a single prompt over the diff. That one model is asked to be everything at once — security expert, performance engineer, architect, semantics nitpicker — and a lone generalist does each role about half as well as a focused specialist would.

Sigilix’s architecture is different. Four domain specialists run per review, in parallel — Logic & Architecture, Security, Performance, Tests — with focused prompts and role-tuned models, unified by the synthesizer that arbitrates between them. Before the LLM specialists fire, a pre-LLM layer of deterministic checks — secret scanning, AST rule packs, user-defined regex rules — extracts cheap signal from the diff.

The topology

Every review goes:

1. Deterministic checks run first — secret scanning, AST rule packs, and any user-defined deterministicChecks regex rules run over the added diff lines. Their findings are injected into the specialist prompts as authoritative facts. See Deterministic Checks.
2. Domain specialists run in parallel — logic, security, performance, and tests receive the same diff plus the deterministic findings as authoritative facts in their context. Each specialist has a different prompt and a model tuned to its role, and can’t see the other specialists’ findings. Each runs with a size-scaled budget and a cross-provider fallback that protects against same-family provider outages.
3. Findings flow into the synthesizer — the synthesizer sees all four streams plus the deterministic findings plus the diff itself.
4. The synthesizer deduplicates, calibrates, and renders — overlapping findings collapse into one; severity shifts based on agreement; review memory adjusts category-level flag-worthiness; each surviving finding earns a proof-tier receipt; the final verdict is decided.
5. One comment is posted — single GitHub review with the synthesizer summary at the top and inline findings below.

Why Sigilix runs a specialist ensemble

Splitting a review across focused specialists and a synthesizer is a deliberate design choice. Here is what each part earns its keep for, and why the whole is worth more than the same model calls run as one generalist pass.

1. Different prompts catch different things

A single-agent design with one prompt asks the model to “look for security issues, performance issues, architectural violations, and naming problems.” The model attends to roughly one of those at a time and trades off depth.

Sigilix’s specialists each have a focused prompt. The security specialist is asked only about security. Its prompt is dense with OWASP-relevant patterns, secret-leak heuristics, and authentication boundary rules. The model running that prompt finds more security issues than the same model running a generalist prompt — by a wide margin.

2. Different models suit different roles

Each specialist runs a model tuned to its role — a reasoning-heavy model for logic, faster high-volume models for security and tests, a throughput-tuned model for performance, and a calibration-strong model for synthesis. The specific model behind each role is tuned over time from telemetry; the docs describe the role, not a model ID that churns.

All specialists have cross-provider fallbacks on independent infrastructure so a same-family outage can’t silence multiple roles at once.

The right model for the job, not one model for everything. See Specialists for per-role model selection.

3. Cross-reference suppresses hallucinations

A single-agent design is prone to hallucinated findings. One model, with nothing to check it, can be confidently wrong about a function being unused, a variable being uninitialized, or a security pattern being broken — and when the file in question is actually read, the finding is fiction.

Sigilix’s synthesizer cross-references findings with the source code. If the security specialist flags a SQL injection at line 42 but the synthesizer’s structural-provenance check shows the parameter actually passes through a parameterized-query helper, the finding is suppressed before it reaches you.

The cross-reference is the difference between “AI review you tolerate” and “AI review you trust.”

4. Severity calibration uses the agreement signal

When multiple specialists flag the same code, that’s a strong signal. The synthesizer escalates the severity in those cases:

• One specialist flags + low confidence → Info
• One specialist flags + high confidence → Warning
• Two+ specialists flag → Warning or Critical (depending on category)
• Specialist + the synthesizer’s structural check confirms → Critical

A single-agent design can’t calibrate this way — it has nobody to disagree with.

5. The interface is one comment, not 40

A single-agent design that dumps every thought it has into the PR thread carries a cost you may have felt: reviewers stop reading after the third “Consider adding a docstring,” and real findings get buried.

The synthesizer deduplicates relentlessly. If two specialists both flag the same loop, you see one comment, not two. If a finding is a duplicate of one already posted on a prior SHA, you see it once.

The cost, and the curve

An ensemble makes more model calls than a single pass — that is the honest cost. What bends the curve back is memory. Every review deposits durable context into your organization’s memory index, so each subsequent review starts more org-aware: it leans on cheaper deterministic and remembered signal, suppresses findings your team has already dismissed, and spends fewer tokens re-deriving what Sigilix already knows about your codebase. The same index feeds review, triage, the CLI, and the Slack assistant — one org-aware platform that compounds with use, rather than a stack of point tools that each start cold.

Sigilix pricing is per-seat plus usage across the Free, Pro, Teams, Max, and Ultra plans, with bring-your-own-key support on the CLI and chat surfaces — see pricing for the current shape.

For most teams, the economics are clear: a single missed security bug shipped to production costs vastly more than the per-PR review cost. For teams with very high PR volume, the per-PR cost can be tuned via rate limits and path filters that scope each review.

Read next