Inspiration
Extra work can get lost between a client's request, an agreement and the invoice. ScopeGuard gives freelancers a reviewable path from a proposed scope change to a verified sandbox invoice. The demonstration uses a fictional landing-page agreement and client requests; customer research and measurable business outcomes remain future work.
What it does
The working local sample assesses changes against supplied scope clauses. A colour change is included in the sample; a new FAQ page and booking work are extra-work suggestions requiring human review. The operator selects work, reviews deterministic line items and approves the exact quote. The authored example combines four and six hours at $50 per hour for $500. A model does not choose prices. Local approval is a workflow control, not authenticated client acceptance or a legal decision.
How we built it
Node.js and a JavaScript browser interface implement assessment review, quote approval and evidence export. AI adapters support an explicitly connected ChatGPT plan session or a configured backend provider. The PayPal sandbox adapter creates itemized invoice drafts and fetches their actual status and USD amount for comparison with the approved quote. A known invoice ID is retained after a failed status lookup to avoid blindly creating another invoice; uncertain creation without an ID blocks automatic recreation.
Owner-approved Sign in with ChatGPT, plan-use consent and account model discovery were verified on October 11. GPT-6-Luna was selected from the returned catalog. Two separately authorized fictional assessments across October 11–12 failed STREAM_INVALID, with zero accepted assessments. The first did not capture HTTP status or content type. The second recorded provider HTTP 200 with a missing content-type header and cancelled the response body without reading it. The underlying cause remains unknown; no returned model, completed response ID or accepted output is claimed. No assessment export or PayPal action was made by these checks.
Challenges and learning
An AI proposal should not become an invoice automatically. Clause-grounded review and deterministic pricing keep the decision visible. Remote invoice creation and later verification can fail independently, so the app preserves known invoice identity and exposes unresolved verification. Provider failures remain errors rather than silently becoming local-rule output labelled as AI. Ordinary unexpected response bodies are cancelled unread. Published diagnostics retain narrow request identifiers, timing and bounded status/content-type categories; an explicit library-only opt-in can inspect a bounded rejected-response prefix and exports fixed shape categories rather than raw content. These diagnostic paths are verified with injected responses and do not establish successful inference.
Accomplishments
The public MIT source passes 86 automated tests, covering invoice boundaries, workflow invalidation, protected backend credentials, account recovery, bounded model discovery, response completion validation, failure diagnostics and cancellation. Tests use injected provider responses and establish no live PayPal or successful AI use. Actual ChatGPT connection and model discovery are verified; successful inference and PayPal sandbox invoice verification remain pending. The existing silent development walkthrough shows the local sample and discloses the remaining integrations.
What's next
Resolve the remaining provider issue and retain one successfully completed fictional assessment with actual provider/requested-returned model provenance. Configure the sandbox, create and verify an actual reviewed PayPal invoice, and record those real actions in the final public silent YouTube video. Complete local setup instructions with necessary private testing access are an allowed judge-access route; public hosting is optional. Sending or paying a sandbox invoice is optional lifecycle evidence. No real payment, production client data, authenticated acceptance, revenue or legal protection is claimed. Codex assisted implementation, testing and documentation.
Source and evidence
https://github.com/Vcode2407/paypal-scopeguard Judge setup: https://github.com/Vcode2407/paypal-scopeguard/blob/main/docs/JUDGES.md Connection and first failure: https://github.com/Vcode2407/paypal-scopeguard/blob/main/docs/evidence/chatgpt-connection-check-2026-10-11.json Second diagnostic failure: https://github.com/Vcode2407/paypal-scopeguard/blob/main/docs/evidence/chatgpt-diagnostic-attempt-2026-10-12.json Successful AI, actual sandbox evidence and the final public YouTube video remain pending.
Log in or sign up for Devpost to join the conversation.