SOW traceability
Each requirement in the Statement of Work for RFQ 19AQMM26Q0468, with what this prototype already shows and what belongs to the contract period of performance. Status is stated conservatively.
| SOW | Requirement | Status | Evidence |
|---|---|---|---|
| §1, Task 3 | Training data includes Ukraine and at least one other country restricted under 19 CFR §12.104g Ukraine (CBP Dec. 24-16) plus Egypt, Peru, Greece, China, Cambodia, each checked against the eCFR §12.104g table. | Demonstrated | Ukraine (CBP Dec. 24-16) plus Egypt, Peru, Greece, China, Cambodia, each checked against the eCFR §12.104g table. View |
| §1, Task 4 | Standalone application, not connected to DoS or other USG systems or networks Static bundle. Inference, index and audit all run on the device. CSP connect-src 'self'. Works offline after first load. | Demonstrated | Static bundle. Inference, index and audit all run on the device. CSP connect-src 'self'. Works offline after first load. View |
| §1, Task 4 | Enough scope, functionality and traceability of outputs for simulated field testing Screening workflow, held-out test samples with hidden labels, hash-chained audit trail with export. | Demonstrated | Screening workflow, held-out test samples with hidden labels, hash-chained audit trail with export. View |
| Task 3 | Use CC0 training data where possible; record data sources and sharing arrangements 75% of designated-category images are CC0 or public domain; per-source inventory with licence terms. | Demonstrated | 75% of designated-category images are CC0 or public domain; per-source inventory with licence terms. View |
| Task 3 | Vet each object; standardise metadata; normalise and vectorise images Keyword and licence guards, automated image screen, manual review log, near-duplicate removal, SHA-256 fingerprints, embeddings. | Demonstrated | Keyword and licence guards, automated image screen, manual review log, near-duplicate removal, SHA-256 fingerprints, embeddings. View |
| Task 3 | No non-public DoS or USG information in the training data Every record is from a public open-access collection, with its source URL retained. | Demonstrated | Every record is from a public open-access collection, with its source URL retained. View |
| Task 4 | Technical specifications: model architecture, data inputs, performance benchmarks Model card; held-out top-1 86%, top-3 95%; model-selection study. | Demonstrated | Model card; held-out top-1 86%, top-3 95%; model-selection study. View |
| Task 4 | Governance specifications: data handling, chain of custody, trustworthy AI Data-handling spec, per-event custody fields, NIST AI RMF mapping, human disposition required. | Demonstrated | Data-handling spec, per-event custody fields, NIST AI RMF mapping, human disposition required. View |
| Task 4 | Security specifications and extensibility to FedRAMP Standalone safeguards are in place, and an NIST SP 800-53 family mapping for a hosted phase is documented. Authorisation itself is out of scope. | Partly demonstrated | Standalone safeguards are in place, and an NIST SP 800-53 family mapping for a hosted phase is documented. Authorisation itself is out of scope. View |
| Task 4 | At least 90% test coverage of contractor-developed code 100.0% line coverage on domain logic (scoring, evaluation, audit, corpus), enforced in CI. UI component tests are planned for the contract build. | Partly demonstrated | 100.0% line coverage on domain logic (scoring, evaluation, audit, corpus), enforced in CI. UI component tests are planned for the contract build. View |
| Task 4 | Free of medium/high vulnerabilities; SCA, static and dynamic scanning at each deployment SCA (npm audit) and static lint run on every build; see the scan results. A commercial SAST service and DAST are added under the QAP. | Partly demonstrated | SCA (npm audit) and static lint run on every build; see the scan results. A commercial SAST service and DAST are added under the QAP. View |
| Task 4 | Government Purpose Rights in prototype, model and data; contractor may commercialise Only permissively licensed model weights (MIT) and open-licensed data are used. A research-only model was rejected for this reason. | Demonstrated | Only permissively licensed model weights (MIT) and open-licensed data are used. A research-only model was rejected for this reason. View |
| Task 5 | Internal testing, validation and performance testing Held-out benchmark with calibration, ablations and latency. Field-image testing is a contract activity. | Partly demonstrated | Held-out benchmark with calibration, ablations and latency. Field-image testing is a contract activity. View |
| Task 1 | Kickoff, Project Management Plan with risk register, monthly reports Delivered within 15 days of award. | Contract task | Delivered within 15 days of award. |
| Task 2 | Stakeholder engagement: CHC, CHCC, law enforcement, CPEOC, cultural institutions Validates scope, user workflows and the training and testing plan. | Contract task | Validates scope, user workflows and the training and testing plan. |
| Task 6 | Simulated field testing and CPEOC concordance The audit export and test-set mode are built to capture concordance and user feedback. | Contract task | The audit export and test-set mode are built to capture concordance and user feedback. |
| Tasks 7–8 | Final Assessment Report and final briefing The Governance, Evaluation and Corpus pages are working drafts of its technical sections. | Contract task | The Governance, Evaluation and Corpus pages are working drafts of its technical sections. |