The Cognitive Delivery Index
How do you know whether your governance is real or just well documented?
The CDI scores a team across six dimensions, each on a five-level ladder. The point is not the score. The point is the distance between level four and level five.
The six dimensions
Provenance Integrity
Can you reconstruct what agents produced and why?
Supervisory Clarity
Is it clear who is accountable for an agent's output?
Scope Definition
Does the agent know the boundary of the work at runtime?
Review Substance
Does review change the work, or only sign it off?
Constraint Fidelity
Do stated constraints actually hold when the agent runs?
Capability Calibration
Is trust in the agent matched to its demonstrated reliability?
Absent to Substantive
Level four is fully documented. Every agent contribution tagged, every prompt retained, the trail complete on paper. Level five is when that control demonstrably changes behaviour. The gap between them is the central contribution of the CDI, and most organisations sit at four believing they are at five.
How real is your governance? Take the ten-minute CDI self-assessment.
Six dimensions, one honest read, nothing leaves your browser.
How a CDI score is read
The Cognitive Delivery Index is a measurement instrument for AI governance in software delivery, not a maturity model: most maturity models stop at level four, which is where its central question begins. Each of the six dimensions is scored on the five-level ladder, from level one, Absent, to level five, Substantive. Score the team as it is, not as it should be.
Six scores add up to a total between 6 and 30, and the total is the least useful thing the instrument produces. What matters is the shape of the profile. The self-assessment reads it in three ways:
- The gap. Any dimension at level four with none at level five is flagged as the governance theatre gap: complete on paper, not yet changing behaviour.
- The pattern. Fours and threes with no five is the audit-ready profile. Scores within one level of each other make a flat profile. A spread of three levels or more is a split profile.
- The nearest team. The profile is matched to the closest of three example teams, lagging, theatrical or balanced, with the first move that team would make.
The CDI companion guide sets out the full scoring guidance for each dimension and level.
What level five looks like, and what it does not
Take provenance integrity. The documentation can be identical at level four and level five. The behaviour is not.
- You can answer "show me the prompt that produced this code" in minutes, not days
- Provenance data has been used in a post-incident review
- Supervisors reference provenance in their review conversations
- The provenance database exists but no one has logged in for weeks
- Prompts are stored, but in a format that is not searchable
- Tagging is done at the end of the sprint, retrospectively, by memory
The orange cell is always level five, filled or empty. An empty orange cell is the governance theatre gap made visible.
Take one of your teams. Score them against the six dimensions. Be honest about where they actually sit. Pick the two with the lowest scores. Decide what one improvement would look like for each in the next quarter.
The CDI is deposited with a permanent identifier. The DOI below always resolves to the current version.
Cruz Falzon, J. (2026). The Cognitive Delivery Index: a self-assessment instrument for governed AI participation in software delivery. Cognitive Delivery Press. https://doi.org/10.5281/zenodo.22286380
Licensed CC BY 4.0. Use it, adapt it, build on it, including commercially.
From the library
What accounting did when volume broke the check
An audit once meant verifying every transaction. Volume ended that, and the profession's real answer was not sampling but withdrawing a claim it could…
Your board does not want an AI governance update
Provision 29 took effect for financial years beginning on or after 1 January 2026, and December year ends are inside the first declaration cycle…
Reporting AI to the board
A 27-page working guide to reporting AI upward: what a board actually asks for, the three tiers of evidence behind any claim, a specimen…
The CDF DORA supplement
DORA tells you whether a team ships well. It does not tell you whether a control demonstrably changes what an agent does.
The CDI cheat sheet
Version 0.1 of the one-page CDI reference, superseded by the CDI Companion Guide. Kept because the URL and the PDF are already in circulation.
Clean data is not a control
Data quality answers the easy half of the trust question, and leaves the part that actually governs the output untouched.