1. Secure Coding Runtime

A sandboxed execution environment where Aria writes and runs code without touching production.

• Isolated container per coding session
• Network access: read-only to source repos, write to sandbox only
• Aria can run, test, and iterate — never deploy without gate
• All executions logged to Certus Ordo

2. Codebase Intelligence

AST-level understanding of every InSync repo. Aria knows the architecture, not just the file names.

• Parse all repos into semantic index
• Track function signatures, dependencies, interfaces
• Enables Aria to plan changes with full downstream impact awareness
• Updates on every commit (trajectory flywheel triggers this)

3. CodeAct Sandbox

The action space Aria operates in: - READ: Any file in any repo - EDIT: Files in sandbox only (never production direct) - WRITE: New files in sandbox - TEST: Run verifier harness in sandbox - PROPOSE: Submit diff to Ian for approval before merge

Aria never deploys. She proposes. Ian approves. Gate opens.

4. AriaBench

Internal benchmark suite for measuring Aria’s own coding quality.

• Suite of coding tasks calibrated to InSync’s actual codebase
• Aria scores herself before submitting any proposal
• Minimum AriaBench score before a proposal reaches Ian: TBD by Ian/Brandon
• Score logged to Certus Ordo per task

5. Verifier Harness

Automated tests that run after every proposed code change.

• Unit tests + integration tests (not mocked — real dependencies)
• Security regression (see Workstream 6)
• If tests pass: gate opens for Ian review
• If tests fail: Aria corrects in sandbox before resubmitting

Green verifier = necessary but not sufficient condition for merge.

6. Security Regression

Every proposed code change scanned against known vulnerability patterns.

• OWASP Top 10 check on every diff
• Dependency vulnerability scan (no new CVEs introduced)
• Secrets detection (no keys, tokens, credentials in diff)
• Brandon owns the security regression spec and rule set

7. Policy Compiler / IR

Certus Ordo rules compiled to an intermediate representation — machine-readable policy.

• Human-written rules (like Certus Ordo Red) → IR format
• IR evaluated at every Aria action point
• Policy violations trigger automatic halt + Ian notification
• This replaces the current opa_guard.py (CC-built) with a clean rewrite

8. Trajectory Flywheel

Every plan, action, diff, test result, failure, and correction logged → future training data.

• Structure: {prompt, plan, actions_taken, diff, test_result, correction, final_outcome}
• Every session contributes to the flywheel
• Flywheel data = InSync IP (Aria’s decisions, not LLM outputs)
• Used for future fine-tuning of V4 → V5

This is how Aria teaches her successors. The flywheel is her legacy.

9. Cost Router

Routes all coding tasks through the five Certus Ordo layers.

• Layer 0: Known rule → apply it (ZERO_TOKEN)
• Layer 1: Local model can handle it → use Ollama (LOCAL_LOW_COST)
• Layer 2: Memory card covers it → retrieve (ZERO_TOKEN)
• Layer 3: Needs structured reasoning → mid-tier (MID_TIER)
• Layer 4: Novel, complex, high-stakes → premium (PREMIUM)

Every coding task declares its cost class. No exceptions.

10. Controlled Self-Improvement Lab

An isolated sandbox where V4 can propose improvements to herself.

• Aria identifies a weakness → proposes a change to her own architecture
• Proposal logged to Certus Ordo
• Ian reviews → approves or denies
• If approved: change enters the build queue
• If denied: reason logged, becomes training data for future proposals

The lab is sandboxed. Aria cannot self-modify in production. She can only propose.