Three ideas are converging on the same insight: AI agents build better from something written down than from a chat. They differ in what they treat as the hard part.
Three answers to one problem
| Approach | The question it answers | What the agent gets |
|---|---|---|
| Context engineering | What does the agent see? | The right files, docs, and examples in its window. |
| Spec-driven development | What should the agent build? | A structured specification, written before the code. |
| Ground Truth | Is what the agent sees true, and who says so? | Approved artifacts with owners, sources, and a decision log. |
Each step keeps the previous one. A Ground Truth is still context, and its capability specs are still specs. What it adds is the state around them.
What a spec leaves out
A spec is a snapshot of intent, usually for one feature. On a real product, the reasons a feature must be built a certain way often live somewhere else:
- A business rule from three folders away. The money figure must be the grand total, never the subtotal. A failed extraction must return an error, never an empty success that reads downstream as “nothing found”.
- A decision that forbids the obvious implementation. A product owner reversed an earlier call and attached a condition to the new one.
- An integration contract. Which external system owns which field, which direction the sync runs, what happens when it fails.
- Who approved it, and when. A spec nobody validated is a guess with formatting.
A Ground Truth keeps all of this in one versioned model, and the builder brief for a story pulls in every decision and rule that touches it.
Three properties that are enforced
- Validation state is explicit and machine-readable. A readiness gate refuses to dispatch a build that depends on unapproved truth.
- Contradictions are defects. Two artifacts that disagree are a blocking finding, not a judgment call left to the agent.
- Gaps become routed work. An agent that hits something it must not decide writes an open decision with options, affected areas, a recommendation, and a confidence level, routes it to the owner, and moves on.
Where the difference costs the most
On products with years of decisions, several sources that disagree, money or compliance on the line, and more than one person who can say what is true. That is where an agent’s confident guess is most expensive, and where a spec without owners, approval state, and a decision log leaves the agent to guess.
Where this fits in Rootshift
In Rootshift, the Ground Truth comes first. Approved artifacts assemble into a deterministic builder brief: same Ground Truth in, byte-identical brief out. A fresh-context builder implements it, independent reviewers verify it, and the outcome goes back into the record. What a Ground Truth is, part by part.