Origin, access model, eligibility, and substitution are reviewer diagnostics, not judging criteria. Authentication requirements are not penalized.
02 The two blind reviews
Two independent reviewers scored this project blind, from a sanitized evidence packet.
Scores are shown separately so the reasoning stays inspectable. A withheld score means the reviewers differed by more than two points.
About text describes untrusted supplier instruction rejection and exact stored mandate evaluation.
Repeated requests resolve to the original ticket rather than duplicate purchase.
Execution
7 /10
The single transaction story is coherent and covers adversarial content, denial, eligible purchase, refresh, and duplicate handling, though no video is supplied.
Evidence cited
Description gives a complete positive and negative transaction path.
Facts report demo alive and submitted frame sheets, but no video.
Impact
9 /10
Travelers have a clear desire to delegate tedious booking while retaining exact financial and policy authority; the safeguards directly address consequential failure modes.
Evidence cited
Traveler-set purchase cap, refundability, arrival, and connection rules.
Independent approval and revocation preserve human authority.
Creativity
8 /10
Combining agent purchasing with explicit authority boundaries, hostile supplier content, and idempotency is a strong original interaction design.
Evidence cited
Supplier prompt-injection example is handled as untrusted evidence.
Policy evidence and duplicate-safe ticket issuance are central mechanics.
Reviewer B (round 2)
confidence 87%
Leverage
8 /10
Structured mandate, trust annotations, policy evaluation, idempotency, revocation, and human-authorized consequential action address reliability and safety gaps that UI-driving agents handle poorly.
Supplier prompt injection is treated as untrusted and cheaper ineligible quote is denied.
Repeated purchase returns original ticket; authority can be revoked without deleting evidence.
Execution
8 /10
The packet describes a complete transaction flow with denied and authorized states, and frames show flight search, evidence/activity, purchase denial, and later authorization; no real purchase occurs by design.
Evidence cited
Frames show LAX-to-JFK offers, eligibility/policy states, Activity and Evidence, and purchase denied/authorized banners.
Description gives a specific adversarial supplier and idempotent ticket scenario.
Impact
9 /10
Travelers have a concrete desire to delegate booking while retaining exact constraints; safe sandbox execution and auditable denial/approval directly address trust and financial risk.
Evidence cited
Specific traveler mandate and booking flow.
Human can revoke future authority and inspect evidence.
Sandbox prevents card charge or airline order while demonstrating consequential-workflow controls.
Creativity
8 /10
The combination of browser-agent booking, policy authority, adversarial supplier content, idempotency, and evidence is an ambitious and memorable safety-oriented interaction model.
Evidence cited
Agent cannot authorize itself.
Trust annotations and policy evaluation are first-class.
Denied then eligible purchase states demonstrate nuanced control.
03 Review highlights
Standouts across reviewers
Strong authority model.
Prompt-injection resistance and idempotent transaction handling in one concise demo concept.
Idempotent transaction and revocation are unusually concrete.
Red flags
Sandbox nature limits evidence of real booking integrations; frame details were not fully viewable due to vision rate limiting.
Frames do not show raw tool names or protocol payloads.
Sandbox behavior does not establish production airline integration.
04 Evidence
What each artifact proves is labeled on the artifact itself. A Devpost page capture is packaging evidence, not proof the product runs;
video frames are evidence from the submitted demo, not live verification.
Contact sheets from the ?s video the team submitted. This is what reviewers were shown;
it demonstrates the product in motion but is not independent verification.
· watch the original
EX-V1 Sheet 1 of 3 · reported video evidenceEX-V2 Sheet 2 of 3 · reported video evidenceEX-V3 Sheet 3 of 3 · reported video evidence
