The capstone launches a complete paper research service from sealed fixtures and optional public GET data. It demonstrates architecture, provenance, validation, evaluation, event sourcing, monitoring, cost controls, and recovery. It does not connect accounts, funds, wallets, or write endpoints.
Your release bundle contains repository commit, safety manifest, dependency lock, configuration/schema versions, source policy, dataset card, raw and normalized hashes, filter waterfall, research case files, human review records, paper-event database, evaluation dashboard, cost ledger, robustness scorecard, backup/restore evidence, and operator runbooks.
Run the pipeline in order: validate safety/configuration; capture or load immutable pages; normalize and quarantine; apply frozen filters; create time-valid evidence packets; run deterministic and optional model checks; complete human review; generate synthetic fills under fixed fictional limits; append/reconcile events; evaluate walk-forward folds; publish the static report; execute failure drills; seal the manifest.
Acceptance requires full denominator reconciliation, no broken hashes, no unknown source IDs, no expired approvals, no unsafe dependencies, and a clean offline replay. The dashboard must foreground missingness, coverage, calibration, cluster concentration, cost sensitivity, and failed runs. Paper outcomes alone cannot pass the project.
Conduct a threat review: secret leakage, arbitrary host access, prompt injection, lookahead, duplicate processing, event mutation, stale report, public debug route, and alert content. Record controls and residual risks. Ask an independent reviewer to reproduce one run and trace one chart to raw evidence.
Prepare a release decision table with each required artifact, evidence path, reviewer, status, and exception. An exception cannot weaken paper-only boundaries, provenance, or denominator reconciliation. Any temporary operational exception has an expiry and a visible degraded state. Archive the signed table beside the tagged build so the demo and repository refer to the same release.
The launch is a tagged course release, not a promise of future live execution. Keep the safety gate permanent. If current Pakistani or provider guidance changes, update context notes and source policy; do not quietly reinterpret the existing evidence.
🇵🇰 Pakistan Angle
Supply an offline fixture and low-resource instructions. Display UTC and Pakistan time clearly. Include current PVARA/SECP source links and state that reachability, hosting location, or course completion does not determine legality. The artifact is engineering portfolio proof, never a signal service.
Hands-On Exercise
Build the bundle, run every gate, restore it on a clean machine/VM, and complete a signed audit checklist. Record one known limitation and one next paper-only experiment. Present a five-minute demo focused on traceability and failure handling rather than results.
Completion Rubric
- All stages, hashes, schemas, denominators, and events reconcile.
- Offline replay and backup restore pass on a clean environment.
- Threat review and failure scorecard have evidence.
- Independent reviewer traces a metric to raw source.
- No credential, transaction, currency sizing, or real-later path exists.
Sources
- NIST Secure Software Development Framework
- NIST AI Risk Management Framework
- PVARA regulations
- SECP press releases
Key takeaway: The finished paper bot is proven by traceability, safe failure, reproducibility, and honest limits—not a simulated return.