This article was written as part of the Kiroween 2026 hackathon.

Public promises have a short life online. A date is announced, the news cycle moves on, and nobody checks. I live in Bunia, in eastern DR Congo, and I wanted a place that remembers.

Promesses Fantômes (Ghost Pledges) turns each verifiable promise into a ghost. The ghost is not decoration: it is the state of the promise. Its opacity shows the evidence level, from L0 (rumour) to L4 (official data). Chains hold it while evidence is missing and break when it grows. A kept promise turns solid in a burst of golden petals; a broken one falls to embers and a gravestone rises from the mist.

Players stake virtual Souls on "kept" or "broken". Souls have no monetary value. A free one-tap poll lets anyone give an opinion without staking. The feed mixes real, sourced content verified on 16 November 2026 (Champions League contenders, public-policy announcements in DR Congo, a celebrity space-launch promise), live football fixtures, and clearly labelled fictional scenarios.

What Kiro did I wrote the rules as steering files and a spec: Souls are virtual, nothing resolves below two independent reliable sources, quotes are never invented, sensitive topics need a human. Kiro implemented the engine and the UI task by task, with property-based tests (fast-check) for the invariants: payouts never create or lose a Soul, no result below evidence L3, nothing is published without criteria, deadline and source, and a model can never lower a safety tier. In total 118 tests pass. Four agent hooks check fictional data, secrets, engine tests and session notes. Kiro also implemented the fixes for real production problems found during verification: Node ESM imports in serverless functions, a wrong season folder in a public dataset, and a changed model name in a text-analysis API.

Disclosure. The planning drafts of the spec and steering, and the independent checks I ran on the pushed repository after each step, were done with Claude; Kiro implemented, tested and corrected the app. Gemini is only a runtime feature inside the deployed app, and its output is treated as untrusted.

My favourite feature: the Cemetery of AI Mistakes. When a human corrects an AI resolution, the app keeps the original claim, the correction, the reason and the date on a tombstone. Admitting an error is part of the product, not something to hide.

Limits. Live political extraction sometimes returns nothing, which is safer than inventing a quote. State is stored on the device in this build; a Supabase schema is prepared but not connected.

Try it: https://promesses-fantomes.vercel.app/ Code: https://github.com/olivierkamali9-blip/promesses-fantomes
Devpost project:https://devpost.com/software/promesses-fantomes-ghost-pledges