Foundational philosophy
The Engineering AI Harness
An operating model for AI participation in engineering that must be defended: AI proposes, deterministic analysis attests, humans decide, and every conclusion regenerates from a named commit. Not a product category — the principles our products are built on.
In requirements engineering, systems engineering and certification work, output is only worth what it can be defended for — in a review, a vendor assessment, or an audit. Generative AI produces capable output with none of the properties defence requires: it varies between runs, it cannot cite the artifacts behind a claim, and it carries no accountable author.
The Engineering AI Harness is our answer to that gap. Like a safety harness or a wiring harness, it does not reduce the power it holds — it is the configuration in which that power becomes usable. It is a set of controls built into the architecture of the product, not written into a policy document, so that AI participation preserves reproducibility, provenance and human accountability by construction.
Why not AI alone
Why agentic AI alone is not enough
Agentic AI is genuinely good at engineering-shaped work: reading a codebase, drafting a trace, proposing a rewrite. The failure is not capability. It is that three properties of generative systems collide with three requirements of regulated engineering.
- Regulated engineering requires reproducibility
- The same question must return the same answer twice. Generative output varies between runs — and an answer that depends on when you asked cannot be evidence.
- Regulated engineering requires provenance
- Every claim must trace to the artifacts and rules behind it. A model can assert; it cannot cite what it computed from, because it did not compute.
- Regulated engineering requires accountability
- A named human answers for every engineering decision. Autonomy diffuses responsibility precisely where certification concentrates it.
None of this is a criticism of the models. These are properties of statistical generation — and they are exactly the properties an audit is designed to refuse. The fix is not a better model. It is an architecture around the model.
Principle → mechanism
Principles with mechanisms, not values on a wall
Each harness principle is enforced by something that exists in Sanad today — a mechanism you can run, not a commitment you must take on faith.
AI proposes, humans decide
A product principle enforced in code: inferred results are marked as candidates for human judgement, and nothing AI produces reaches evidence without deterministic verification and explicit human acceptance.
Trust requires evidence
Reports regenerate byte-identically from a named commit — the same repository state produces the same bytes, on any machine, any day.
Governance over autonomy
Human approval gates: changes enter engineering memory through reviewed pull requests and explicit acceptance steps, never through an autonomous write.
Explainability by construction
The traceability architecture: every finding cites the artifacts, the rule and the commit that produced it, so "why does it say that?" always has a mechanical answer.
Safety is contextual
Criticality bands: rigour scales with how critical an item is, and a check that stands down at a lower band is shown as stood down — never silently dropped.
Decisions must be reproducible
Evidence generation from the repository: audit artifacts are computed from a pinned repository state, not assembled by hand and defended from memory.
Built and working today
Shipped- Governance & rule packs
- Policy compiled once and enforced uniformly — rules and rigour bands resolved at load, never re-interpreted at the point of use.
- Traceability
- A typed graph over your requirements: every item knows what it reaches and what reaches it, in both directions.
- Audit-ready traceability report
- The certification deliverable — trace matrices, coverage, declared gaps and the report's own configuration identity — generated in Markdown, HTML and CSV.
- Byte-identical evidence regeneration
- Any report re-runs from a named commit and reproduces exactly, byte for byte.
- Criticality bands
- Rigour scaled to criticality, with the customer's native scheme labels shown alongside.
- Human gates
- Review and acceptance as mandatory steps in the architecture, not settings.
- EARS pattern engine
- Requirements-pattern recognition and authoring support in the vocabulary engineers already use.
- Guided setup
- A working, governed repository from the first hour — no consulting engagement required.
Under development
In developmentAI participation capabilities under development within the Engineering AI Harness architecture.
- An assistant that turns an engineer's question into governed queries over the compiled graph — facts from the graph, reasoning from the model, never the reverse.
- AI-assisted implementation mapping: ranked candidate links between requirements and code, each with a rationale — becoming facts only when an engineer confirms them.
- AI pattern assistance: suggesting the best-fitting requirement pattern and rewriting a draft into it — always as a diff the author accepts, never a silent change.
Every one of these arrives inside the controls above: proposals marked as proposals, no path into evidence, a human decision at the end. We built the governance foundation first.
Sanad
How Sanad implements it
Sanad is a requirements and engineering governance platform built on Engineering AI Harness principles. The deterministic core — compiler, typed graph, analysis engines, evidence generation — is the fact layer: it owns what is true, and it is fully functional with no AI configured at all.
AI attaches at the rim, on your terms: one gateway, your endpoint, your credentials, opt-in and removable. What AI produces is a proposal; what the engineer accepts becomes an ordinary change in Git, reviewed like any other; what the deterministic core computes is the only thing that can become evidence.
That order — foundation first, participation second — is the harness. It is why AI in Sanad can make engineers faster without making their evidence weaker.