Skip to content
HardStack.AI

Carbon · How it works

A controlled pipeline, not a chat box.

Carbon generates regulator-grade documentation from structured intake. Humans gate every phase. Quality is scored by rules, not by the same model that wrote the sentence.

Orbital rocket igniting on a coastal launch pad

Where Carbon sits

Engineering workflows on the left. Regulatory submission on the right. Carbon is the documentation layer in between. It does not replace Jira, DOORS, Jama, or Polarion. Native Jira sync is EARS-scored. REST covers Confluence, GitHub, and Notion.

Before

ALM / engineering

Jira, GitLab, DOORS, Polarion, Jama. Requirements a human already wrote.

Carbon

Documentation engine

Intake, swarm, audit, signed bundle. BYO-LLM inside your perimeter.

After

Submission / audit

FDA, FedRAMP, SOC 2, customers, counsel. Evidence that already exists.

Five phases

The week, in order.

  1. Phase 01 · Gate G1

    Intent expansion

    A short set of intake answers plus supporting documents. The framework recipe resolves. Stakeholders approve the intent map before authoring starts. Nothing silent happens here.

  2. Phase 02 · Gate G2

    Multi-modal ingestion

    PDFs, notes, screenshots, existing requirements. They become a constraint log. The team ratifies what the system is allowed to treat as true.

  3. Phase 03 · Gate G3

    Generation swarm

    Eight agents author in parallel across BRD, SRD, threat model, evidence, traceability, risk, diagrams, and the log. Framework cues swap agents, spawn artifacts, and force overrides for criminal-penalty frameworks. Swarm sanity is a human gate.

  4. Phase 04 · Gate G4

    Quality audit

    RQS is deterministic. EARS pattern detection. INCOSE rule checking. An audit diff shows what changed. Export is blocked under the fail band and warned under the target. This is not the LLM grading its own homework.

  5. Phase 05 · Gate G5

    Finalize and export

    Signed bundle. Hash-chained history.log. If someone edits a file after export, one command detects it. That is the artifact you take to an auditor, a reviewer, or a buyer.

What you should not expect

  • Not an ALM

    Carbon does not manage living requirements, tests, or tickets. It generates the documentation those systems cannot.

  • Not a GRC platform

    Vanta and Drata prove controls operate. Carbon proves the system was specified, traceably, from the start.

  • Not autonomous

    Gates G1, G4, and G5 are mandatory. Carbon briefs before every phase. Rewinds are logged with a reason.

Request a trial