When AI translates, a linguist signs.

A modern machine-translation model is convincing precisely because its errors are fluent. The classical failure modes — literal wordings, grammatical fractures, missing articles — have been replaced by a more dangerous class: plausible sentences that confidently misstate the source. For a federal buyer, the implication is operational, not academic. A fluent mistranslation in a force-protection bulletin, a clinical consent form, or an open-source intelligence summary will be acted upon by someone whose first instinct is to trust the text in front of them.

The verification problem is therefore not can the model translate, but can the institution prove the translation was witnessed. NIST's AI Risk Management Framework treats verifiability and traceability as first-class trustworthy-AI characteristics; the President's October 2023 executive order on safe AI elevates them to federal-procurement requirements.^[1]

"A fluent error is a more dangerous error. Verification is not editing — it is the documented act of a credentialed human standing behind the page."

Not every deliverable belongs in the same workflow. DayDefense classifies every inbound assignment into one of three risk tiers at intake, and the tier determines which workflow runs. The classification is recorded in the work order and travels with the deliverable as metadata.

Tier assignment is overridden upward, never downward, when in doubt. A contracting officer may stipulate a tier in the order; DayDefense will not classify below the customer's stipulation.

1. 01
   Intake & Classification

   Order received. Risk tier assigned. Source files normalized. Personally identifiable information and protected health information are tagged for handling under the customer's controls (Privacy Act, HIPAA, CUI-PII).

2. 02
   Model Selection

   Approved model and version are selected from the model registry; the selection (model ID, version, hash) is logged on the work order. We do not invoke unapproved or unpinned models against federal work.

3. 03
   Machine Draft

   Source is processed through the selected model. Output is stored alongside the source and the model identity, never overwriting either. Token-level confidence scores are retained where the model exposes them.

4. 04
   Linguist Verification

   A credentialed linguist (see Linguist Vetting Dossier) reviews the draft against the source, segment by segment, correcting errors, restoring nuance, and flagging segments where the draft was unrecoverable. Each correction is captured for downstream learning.

5. 05
   Quality Assurance

   A second, blind reviewer samples or fully re-reviews the verified output per the contract's QA requirement. Disagreements between reviewers are adjudicated by a senior linguist and recorded.

6. 06
   Signed Deliverable

   The final file is signed under the verifying linguist's name and roster ID. The audit packet (source, draft, corrections, reviewers, model identity, time stamps) is bound to the deliverable as metadata.

7. 07
   Delivery & Retention

   Deliverable is transmitted to the customer in the contract's preferred channel. The audit packet is retained under FAR 4.703 and made available on written request.

Every Tier II and Tier III deliverable leaves DayDefense with a discoverable audit packet. The packet is the institutional answer to the verification question — not a marketing artifact but a record produced on written request from the contracting officer, the contracting officer's representative, or an authorized auditor.

• Source-file hash and timestamp at intake
• Risk-tier classification and the named classifier
• Model identity: registry ID, version, commit/hash, vendor attestation
• Machine draft output and any token-level confidence data retained
• Verifying linguist roster ID, credential pointer, and time logged
• Per-segment correction log (substituted text, reason code where used)
• QA reviewer roster ID, sampling rate, disagreement adjudications
• Signed deliverable hash and timestamp at egress
• Retention disposition under FAR 4.703

The packet is produced in machine-readable form (JSON or signed PDF, customer choice) so it can be ingested directly into the customer's AI inventory, system-of-records notice, or program-protection plan.

A verification claim is only as strong as the failures it has survived. DayDefense conducts periodic red-team exercises against the AI pipeline — adversarial inputs, deliberately ambiguous source material, prompt-injection attempts on document-aware models — and records the findings. Significant deviations are escalated under the customer's incident-response framework where one is contractually defined; absent that, we use the internal incident protocol below.

1. 01
   Detect

   QA sampling, drift watch, customer-reported incident, or red-team finding identifies a deviation from expected pipeline behavior.

2. 02
   Contain

   The implicated model is suspended from the work-order pool. Affected deliverables in flight are routed to human-only workflow.

3. 03
   Assess

   A senior linguist plus the AI assurance lead reproduce, scope, and characterize the deviation. The assessment is written, signed, and timestamped.

4. 04
   Notify

   Customers whose orders touched the implicated model are notified per contract notification SLA, with the scoped impact and remediation timeline.

5. 05
   Remediate

   Model rotation, version pin update, prompt sanitization, additional QA pass, or workflow change — whichever the assessment indicates. Remediation is logged.

6. 06
   Close

   Closure brief is filed in the records system and made available to affected customers on request. Learning is incorporated into the red-team test set.