IDD concept: Measure alignment
Intent Fidelity
Intent Fidelity measures how closely an implementation matches its original intent, across four dimensions: completeness, correctness, alignment and consistency. It is the primary control metric in Intent-Driven Development, the evidence that tells an organisation when AI systems are behaving as intended and when it is safe to grant them more autonomy.
The question it answers
Once governance controls are in place, senior leaders tend to ask two questions: how do we know this is working, and how do we know when it is safe to trust AI more? Business outcomes such as revenue, adoption or customer satisfaction help with the first question only slowly, because they are lagging indicators; by the time an outcome reveals that work was misaligned, the misalignment has been running for some time.
Intent Fidelity is a leading indicator. It is measured as each piece of work is completed, against the Intent Specification it was built from, so it reveals divergence early, while it is still cheap to correct. That makes it the right basis for decisions about Risk Dials.
The four dimensions
- Completeness
- Did the implementation address everything that was specified? Missing features, ignored constraints and skipped ethical considerations all reduce completeness.
- Correctness
- Does it work as specified? An implementation can address every requirement and still fail a performance threshold, in which case it is complete but not correct.
- Alignment
- Does it reflect the real intent, the purpose behind the requirements as well as their letter? This is the most important dimension and the least automatable, because assessing it depends on people who understand why the work exists.
- Consistency
- Does it fit coherently within the existing system? Code that works but uses the wrong architectural patterns, or contradicts the established domain model, is inconsistent.
The same standard for people and machines
Intent Fidelity applies equally to human and AI implementations. People are not exempt from measurement, and AI is not held to a harsher standard than the colleagues it works alongside. This baseline principle has a useful diagnostic consequence: when human and AI implementations of the same specification both score poorly, the problem is usually the quality of the intent rather than the implementer, and the right response is better Intent Shaping.
IDD does not ask leaders to trust AI more than humans. It asks them to trust measurement more than intuition.
Using fidelity to adjust governance
Fidelity scores, sustained over a meaningful number of implementations, determine where a Risk Dial can safely sit.
| Measured fidelity | Evidence required | Appropriate control |
|---|---|---|
| Below 85% | Any, including early implementations | Red: full review |
| 85 to 95% | Sustained across twenty or more implementations | Amber: spot-checking |
| Above 95% | Sustained across fifty or more implementations | Green: monitoring only |
These thresholds apply per workflow step and, in multi-agent systems, per specialist agent, since different agents earn trust at different rates. Some areas, such as security, payments and core infrastructure, remain under full review whatever their scores.
From measurement to business language
Intent Fidelity translates directly into terms a board recognises. Rising fidelity alongside reduced review effort is a velocity improvement; early detection of misalignment is risk reduction; and a sustained record of high fidelity is the evidence an organisation needs to move work from pilot into production with confidence. It also changes the nature of quality assurance, which shifts from asking whether the code works to asking whether it faithfully represents the governing intent.
Where it comes from
Intent Fidelity was set out in Measuring Intent Fidelity, published on 8 February 2026, and forms Chapter 6 of the ebook.