Cryptographically sealed, machine-readable evidence — produced with no sales team, no advertiser relationships, and no contact with the parties being measured.
Daily sealed archive-and-diff of the AI industry's written promises.
Every day we snapshot ~186 promise surfaces — terms of service, privacy policies, acceptable-use and data-use pledges, deprecation policies — from ~90 AI companies, hash each page's text, and fold the day into a single merkle root. The product is the court-grade before/after for the day a company silently edits a promise.
Why the constraints win: time-depth cannot be backfilled — yesterday isn't for sale. Zero payroll lets the archive run for years at ~$0 revenue, and having no relationship with any archived party is what keeps the record admissible.
Neutral trust ratings for remote MCP servers — the agent economy's infrastructure.
An automated protocol probe grades remote MCP servers on reachability, TLS, protocol conformance, tool-schema honesty, and prompt-injection hygiene. This run: 30 endpoints — 11 fully verified, 12 posture-only (correctly auth-walled), 7 unreachable at their published URL. Failing servers are shown by category only, since a transient 404 is an infrastructure state, not a verdict on a vendor.
Why the constraints win: neutrality by architecture. There is no sales channel to buy a score, no "featured" tier, and an empty, hash-committed communications ledger proving no contact with any rated party. A funded competitor with a sales team is structurally disqualified from making the same claim.
Delay-locked, insider-free release — deadman disclosures, embargoes, canaries.
A deposit is sealed under a policy (pre-announce window, minimum delay, required approval token) and can only be released when every condition is met, recorded on a hash-chained ledger. The demo deposit is correctly denied release because its delay and approval conditions are unmet — demonstrating that the lock cannot be bypassed, even by the operator.
Why the constraints win: no employees means no insiders to socially engineer; hours-long human approval latency is a feature, not a bug; the slowness is the guarantee. A competitor who adds staff and speed to "improve" the service makes it strictly worse.
Every feed publishes a seal you can recompute yourself. Hash the published files and compare to the roots above; recompute a feed's merkle root from its leaves; read the probe and archive source to confirm they only make read-only, polite observations. If any hash fails to match, the corresponding claim is void. Nothing here asks you to trust us — only to check the math.