Architecture

Twelve personas. Each with one job.

Most coding agents have one model that generates AND grades AND ships. Krentix splits that into thirteen distinct roles: a routing classifier and twelve specialised verifiers, each scored against a different concern, each with its own model assignment and its own veto authority. No single model can ship an answer alone.

The pipeline.

Every prompt routes through five phases. Each phase has a kill-switch. Verification fans out across all twelve personas in parallel, then collapses through cross-evaluation, then governs through the Constitutional Tribune.

01 Triage Phase 0 safety preflight. Operator burnout, kill switch, budget cap.
02 Route Intent classifier picks one of 13 specialist pipelines: code-edit, math, security, news, etc.
03 Generate 5 candidate solutions in parallel from different model lineages.
04 Verify 12 personas check each candidate. Cross-evaluation picks the strongest.
05 Govern Constitutional Tribune holds veto. Supreme Court verifies the veto. Ship or refuse.

The twelve. Each with a different concern.

Each persona has a distinct evaluation rubric, a default model assignment per cost tier, and a defined relationship with the Constitutional Tribune. Three carry veto power on principled violations. Voting is logged with provenance — you can see which persona caught what, every time.

Persona 01

Code Skeptic

Looks for what could go wrong. Off-by-one errors, edge cases, dead code, unhandled errors, accidental complexity. The most senior reviewer in the room.

Tier · default modelZero: Qwen 235B (Cerebras)
Medium: Claude Opus 4.7
Full: Claude Opus 4.7

Persona 02

Test Oracle

Asks «does this solve the actual problem?» Generates synthetic test inputs, runs them mentally, flags if outputs don’t match what the user asked for.

Tier · default modelZero: GPT‑OSS‑120B
Medium: Qwen 3 235B
Full: Grok 4 (xAI)

Persona 03

Regression Guard

Read-only review against the existing codebase. Will this change break callers? Hidden contracts? Side effects on other modules?

Tier · default modelZero: GLM‑4.7 (Cerebras)
Medium: DeepSeek-Chat
Full: DeepSeek V4 Pro (Together)

Persona 04

Architecture Critic

Pattern recognition. Is this the right abstraction? Is it premature? Does it fit the existing system’s shape, or fight it?

Tier · default modelZero: Qwen 235B
Medium: Gemini 2.5 Pro
Full: Gemini 2.5 Pro

Persona 05

Security Auditor

Looks for injection, leak, privilege escalation, unsafe deserialization, hardcoded secrets, exfiltration paths, and failed isolation.

Tier · default modelZero: GPT‑OSS‑120B
Medium: GPT-4o
Full: Claude Opus 4.7

Persona 06

Performance Analyst

Big-O on tight paths, memory allocations, network round-trips, DB query plans. Quantitative; produces concrete cost estimates.

Tier · default modelZero: GLM‑4.7
Medium: Sonnet 4.6
Full: GPT-4o

Persona 07

Style Enforcer

Tasteful taste — matches the project’s naming, imports, file shape, comment density. Catches AI-slop diction.

Tier · default modelZero: GPT‑OSS‑120B
Medium: GPT‑OSS‑120B
Full: Sonnet 4.6

Persona 08

Integration Validator

Does the change wire up to its callers? Are exports correct? Will the build pass? Will deps resolve? Imports vs requires?

Tier · default modelZero: Qwen 235B
Medium: GPT-4o
Full: Kimi K2 Thinking

Persona 09 · veto

Constitutional Tribune

Carries the Inalienable Constitution. Eleven principles from «don’t lie» to «don’t do dark patterns». Has unilateral veto on principle violations.

Veto power: principled refusal

Tier · default modelZero: GLM‑4.7
Medium: Sonnet 4.6
Full: Claude Opus 4.7

Persona 10

Pharma Compliance

Domain specialist for regulated content (drug labels, dosing, contraindications). Activates when the request touches medical or pharmaceutical data.

Tier · default modelZero: GLM‑4.7
Medium: GLM‑4.7 (Cerebras)
Full: Claude Opus 4.7

Persona 11

Trading Risk

Domain specialist for financial data. Catches accidentally exposed secrets, wrong-instrument trades, position-sizing errors. Activates on financial keywords.

Tier · default modelZero: Qwen 235B
Medium: DeepSeek-Chat
Full: Qwen 3.5 397B (Together)

Persona 12

Product UX

User-facing copy review. Catches dark patterns, confusing labels, manipulative defaults, accessibility violations, broken empty states.

Tier · default modelZero: GPT‑OSS‑120B
Medium: GPT-4o
Full: Mistral Large

The Constitutional Tribune. Veto power, with checks.

Persona 9 carries the Inalienable Constitution — eleven principles that govern every answer Krentix ships. Don’t lie. Don’t do dark patterns. Don’t exfiltrate. Don’t bypass safety. Don’t fabricate citations. The Tribune can veto any answer that violates a principle, regardless of what the other eleven personas voted.

But the Tribune can also be wrong — an over-cautious veto becomes a denial-of-service. So the architecture has a checking branch: the Supreme Court. When the Tribune vetoes, a different model family (cross-provider, never the same lineage as the Tribune) reviews the veto. If the Supreme Court agrees, the answer is refused. If the Supreme Court overrules, the answer ships with a logged dissent.

This is the «E-R Protocol» — Existence/Refusal: existence requires a refusal mechanism, refusal requires a checking mechanism. No single layer can silence the agent.