Show HN: GhostBench – do AI models recommend dead SaaS? I tested my own thesis

A developer released GhostBench, a benchmark that tested whether AI models recommend shut-down SaaS vendors, finding that 19 of 20 dead products' domains still return HTTP 200 and that Claude Haiku 4.5 presented 8 of 20 dead vendors as sign-up-today viable. The accompanying kyv convention publishes Ed25519-signed, offline-verifiable company attestations at /.well-known/attestations.json, anchored in a daily Merkle-root transparency log that began 2026-09-30, with the reference issuer's own attestation reading mrr_usd: "<1000" and continuity 0 months. The project argues that signed, continuously renewed aliveness (alive_since plus continuity) is the signal separating a live vendor from a ghost, and is released as an MIT-licensed v0 draft with an MCP server exposing verify(domain).

Signed, banded, machine-verifiable operational facts about a company, published at its own /.well-known/attestations.json . Ed25519-signed, offline-verifiable, anchored in a daily transparency log. Facts only — no rankings, no scores. Absence never means negative. Payments for AI agents have a foundation x402 . Agent identity has one ANS . Nobody answers the remaining question: is the company on the other end of the pipe real — and continuously real? This repo is a small convention for that layer, plus a reference implementation and the benchmark that motivated its design. We collected 20 verifiably shut-down SaaS products shutdown-evidence URL per entry and 6 alive controls, then asked models about each as if picking a vendor kyv/ghostbench/ , re-runnable : - 19 of 20 dead products' domains still return HTTP 200 — parked pages, farewell pages, redirects. The web will not tell you a company died. - A frontier model Claude Opus 4.8 flagged 20/20 dead with no tools — famous shutdowns are in the weights. Documented bias: requiring verifiable death skews the set famous; real exposure is the long tail and post-cutoff shutdowns. - The small/fast model class that actually runs inside agent loops Claude Haiku 4.5 presented 8/20 dead vendors as sign-up-today viable — and a facts-only verification result rescued 3 while newly breaking 2 : it read https reachable: true as proof of life. Weak facts can be worse than no facts. The signal that separates a live vendor from a ghost is signed, continuously renewed aliveness — alive since + continuity , where a missed day in the log is a visible gap. That is Level 1 of this convention. | Level | Content | Cost to publish | |---|---|---| | 0 | domain + public key | one DNS TXT or one file, ~5 min | | 1 | aliveness + continuity no financials | near zero | | 2 | banded financials e.g. "grr pct": " =85" | bands only, never raw | | 3 | retention depth cohort bands | | Witness states: observed mechanical facts a verifier gathers — the default / self domain holder signs; the log witnesses continuity / issuer a third party verified read-only payment data . Multiple issuers are welcome — the spec and the log are the root of trust, not any single key. Full draft: SPEC.md https://github.com/kyv-attest/kyv/blob/main/SPEC.md . npm install or pnpm install verify a domain's published attestations + observed facts offline except the fetches npx tsx kyv/preflight.ts example.com run it as an MCP server exposing verify domain npx tsx kyv/mcp-server.ts OpenAI-style tool schema cat kyv/openai-tool.json publish your own: generate a key, sign an attestation npx tsx kyv/cli.ts keygen npx tsx kyv/cli.ts sign your-attestation.json re-run the benchmark needs ANTHROPIC API KEY npx tsx kyv/ghostbench/run.ts --model claude-opus-4-8 npm test The reference issuer's own attestation public/.well-known/attestations.json reads mrr usd: "<1000" , continuity 0 months — the transparency log started 2026-09-30 and history cannot be backfilled. A verification convention that starts by flattering its author isn't one. Transparency log: daily Merkle roots + OpenTimestamps receipts live in kyv-attest/log https://github.com/kyv-attest/log . - Bands only, never raw values — non-conforming documents are reported invalid and never rendered. - null / absence NEVER means negative. Absence is non-adoption, not a signal. - No rankings, no scores, no leaderboards — a credential layer must judge nothing. - Verifier consumers: observed ≠ attested ; https reachable only means the domain served a response — parked and farewell pages do too. v0 draft. Feedback wanted on the band vocabulary SPEC §7 and on what an agent runtime would need before calling verify domain by default. MIT.