The binder in the banker's box
An accreditation package arrives at a program office. Nine hundred pages of system security plan, scan results, and control narratives, assembled by a compliance contractor over four months and finalized in the spring. The security team starts the review in the fall. By then the vendor has shipped three more releases. The scan results describe binaries that no longer exist. The architecture diagram shows a service that got replaced in June. Everyone in the room knows the document and the software have drifted apart. They review the document anyway, because the document is what the process accepts.
This is the graveyard of good software. Programs do not die because the code fails. They die in the gap between the code and the paperwork about the code. The paperwork takes months to produce, so it is stale on arrival, so the review takes longer, so the paperwork drifts further. Teams that ship fast get punished for it, because every release resets the drift clock. The rational move is to ship slowly, and plenty of vendors have made exactly that adjustment.
The slowest part of fielding defense software is not writing it. It is proving it. And the proof has always been a separate product, built by different people on a different schedule from the software it describes. That separation is the whole problem.
One hundred and ten controls, by hand
The scale of the document drill is easy to underestimate. CMMC 2.0 asks over 200,000 defense industrial base companies to demonstrate 110 security controls. For most of them, demonstration means screenshots pasted into spreadsheets, policy memos written to match assessor expectations, and evidence folders assembled in the weeks before an audit and abandoned the week after. The evidence describes a moment in time. The software keeps moving.
The Pentagon knows this. The continuous ATO push exists because the department concluded that point-in-time accreditation cannot keep pace with modern release cadence, and the FY 2026 NDAA's acquisition reforms push commercial-first buying and faster software pathways in the same direction. But cATO guidance mostly tells programs to automate their evidence collection. It does not tell vendors how. A vendor whose compliance evidence is written by humans after the fact cannot participate in continuous accreditation no matter how the program office structures the process. The evidence pipeline is the vendor's problem to solve.
The proof is a build step
EdgeLance solved it by moving the proof into the build. Every release ships with its own evidence package, generated by the same pipeline that produces the software. Four artifacts come off every build. A CycloneDX SBOM enumerating every component in the release. A NIST 800-53 control matrix generated from the platform's signed audit ledger. A model provenance attestation covering every AI model the release can run. And a STIG posture report for the configurations the release enforces.
The package is not a description of the release. It is an output of the release, hash-bound to the exact binaries being evaluated. Change the code and the next build emits a new package that reflects the change. Nothing has to be updated by a compliance team, because there is nothing to update. The evidence regenerates the way a test suite reruns.
This works because of a design decision made long before compliance was the goal. EdgeLance records every consequential action in a signed, append-only audit ledger. Device enrollment, model approvals, policy decisions, mission burns, all of it lands in the ledger as typed entries with signatures and timestamps. The ledger exists so operators can prove what their AI did and when. It turns out the same ledger proves what the platform did, which is what an accreditor is asking.
The ledger is the evidence, not a summary of it
The 800-53 matrix is the clearest example. Most control matrices are prose, a narrative asserting that the system satisfies AC-2 because a paragraph says so. The EdgeLance matrix is generated by mapping ledger entry types to controls. Enrollment entries demonstrate the account management and device identification controls. Model approval entries demonstrate configuration change control, because no model runs on the fleet without a recorded human approval. Policy decision entries demonstrate access enforcement, with the actual denials and grants attached. Burn entries demonstrate media sanitization, with the cryptographic destruction event signed into the record. The matrix does not claim the controls are met. It cites the entries that met them.
Model provenance gets the same treatment. Every model artifact on the platform carries a SHA-256 manifest chain covering the base weights, each quantization step, and every adapter applied on top. The attestation in the evidence package walks that chain from the approved source artifact to the exact files deployed on the fleet. An accreditor asking where the AI came from gets a verifiable chain of hashes, not a vendor's assurance. This matters more every year, because model supply chain questions are arriving in accreditation reviews faster than guidance for answering them.
What the accreditor opens instead
Go back to the program office. Same review, different vendor. The security team receives an evidence package generated from the exact release candidate sitting in front of them. The SBOM hashes match the binaries. The control matrix cites ledger entries from the platform's own operation, not narratives written to satisfy a template. The provenance chain resolves. When the team asks about a control, the answer is a query against the ledger, not a follow-up meeting three weeks out. And when the vendor ships the next release, the next package arrives with it, current by construction.
That is what continuous ATO actually requires from a vendor. Evidence that cannot go stale, because it is the code's own record of itself. The 110-control spreadsheet economy exists because compliance and software were built by different people at different times. Build them as one artifact and the drift disappears, along with the months spent reconciling it.
The banker's box stays in the truck. The evidence for the release under review came off the build that produced it, this morning.