Demonstration · Synthetic fixture data · No real reports, notifications, or payments
sentinelEvidence

Safety & fairness

Enforced in code.Not promised in copy.

Each safeguard below is backed by an adversarial test in the project's suite, run against a real Postgres database, and by a written policy. Where the system falls short, we say so.

Evidence frames orbiting a glowing seal inside a glass sphere

Safeguards

Twelve rules.No score can bend them.

The test file named under each rule tries to break it: self-approval, AI-only awards, edited seals, conflicted scouts, private-network fetches. Each attempt is expected to fail.

01

No accusation from a score

A case reaches owner outreach only when a named reviewer advances it with all 13 checks true. A rejection always marks it a false positive, and every finding is audited.

tested reviewerChecklist.test.ts
02

No payout from a score

Lead, standard and resolution awards each require their own human-sourced record. Without that record, the award cannot be created.

tested bountyAwardGating.test.tsBounty policy
03

A separate person approves every payout

Approval needs payout permission, verified identity and a linked account. A scout can never approve their own payout, even with administrator rights. Approval is idempotent.

tested payoutApproval.test.tstested payoutSafetyHardening.test.tsPayout policy
04

First valid reporter, no conflicts

One award per case and outcome. A scout tied to the rights-holder contact is blocked on that case, and the case itself continues.

tested bountyAwardGating.test.tsFraud & conflict policy
05

Evidence cannot be rewritten

Seals are immutable at the database level. Replacing a file creates a new version. Uploading never marks a work as registered with the Copyright Office.

tested creatorRegistry.test.tsCreator registry policy
06

AI cannot verify ownership

Upload sets 'claimed owner' only. Matching and scout submissions never promote it, and funding a pool never affects verification.

tested creatorRegistry.test.ts
07

Neutral language, enforced

Accusatory language is rejected before an outreach message is stored. Sending needs approval permission, respects opt-outs, and stops under the emergency pause.

tested outreachPolicy.test.tstested terminology.test.tsTerminology guide
08

Money never steers outcomes

Donations never touch case status or award eligibility. An owner's actions never satisfy the reviewer gate. A platform removal alone never creates a legal determination.

tested payoutSafetyHardening.test.ts
09

Public pages only

Collection is fail-closed per source and pausable. Private, loopback, link-local and cloud-metadata addresses are refused, with or without demo mode.

tested sourcePolicy.test.tstested fetchGuard.test.tstested ssrfGuardWithDemoModeOff.test.ts
10

Untrusted media stays contained

Video, audio and images are decoded in a locked-down sandbox with size, pixel-count and time limits, and filenames are never taken from the caller.

tested mediaSandbox.test.tstested imageSandbox.test.ts
11

Holds, clawbacks and appeals

A hold blocks approval until it is released. A clawback needs permission and a paid award. Appeals need a distinct human and never reverse anything automatically.

tested payoutApproval.test.ts
12

A demonstration cannot touch the real world

Demo mode blocks hosts outside a small allowlist, any payment or identity provider that is not simulated, and any mail sink but the local one.

tested campaignDemoIsolation.test.ts

What we do not claim

Honest aboutthe edges.

A rights platform earns trust by stating its limits before anyone has to find them.

  • We do not claim to find every instance of reuse, on every platform.
  • A platform removal is never treated as proof that a use was unlawful.
  • We make no legal determinations. Only a court or authoritative process can, and the record says so when none exists.
  • An evidence seal is not a government registration, and Sentinel never files one for you.
  • Benchmark misses: small picture-in-picture flush against the frame edge; short segment embedded in a longer compilation; heavy crop (about 35% removed).
  • The benchmark corpus is synthetic. Real-world accuracy has not been measured yet.
  • Text and plagiarism matching is not built.
  • No live platform is crawled. Discovery runs against a local fixture server; real sources stay behind a legal activation checklist.
  • Payments, identity checks and owner email are simulated or local-only in this build.

Plain language

Rights work goes wrong when a tool sounds more certain than it is. These three statements appear in the product wherever they apply.

Seals are private records Every match goes to a person The rights holder decides Courts decide infringement

Transparency

See the numbers,with their sources.