The validator in Chapter 10 returns one of four words. Somebody then has to decide what a platform does with each of them, and the code does not make that decision. The circuit breaker function returns true or false against a threshold. That is as far as it goes, and it describes no product. This note makes the decision, as a proposal. A proposal can be turned down without anyone having to show that a prediction failed.
Call it a verified tier. I put the name at the end of last week’s note, and this is where it gets defined. I did not take the phrase from a statute. I cannot show you it running anywhere, and nothing on this site describes a platform that has adopted it. It is a sorting rule.
The rule is small. When a file comes back VERIFIED, a platform may prefer it. Preferring is a long way short of treating the file as gospel. In practice it might mean a label the reader can see. It might mean a better place in a ranking, or a reason not to trip the throttle the entropy engine describes. The file has an issuer on the trust list and a hash that still matches, which is a fair reason to treat it differently from a file that has neither. A newsroom still has to use the rest of its judgment. Last week’s limit holds: the chain can be intact while the claim in the file is false.
Each of the other three statuses means something different, and the tier keeps them apart.
WARNING means a signature exists, but the signer is not on the list the reader trusts, or the credential is self-signed. The proposal shows that as a named uncertainty. Fold it into fake and you throw away the only fact the check produced.
TAMPERED means the hash does not match. Here the proposal withholds the boost and says why. The code’s reason string mentions deepfake injection, and I will not stretch that gloss any further. A mismatch means the bytes and the manifest disagree. That is enough to stop treating the file as the thing that was signed. It does not tell you what kind of edit caused the disagreement.
UNSIGNED is the gray zone, and a sloppy policy will get it wrong. The reason string is “No credentials found.” The proposal forbids showing that result as guilt. A tier that treats unsigned files as fraudulent condemns every honest file that nobody signed, all to feel safe. The entropy engine’s provenance weight works on a stream: a multiplier of 1.5 on the unsigned share. It passes no verdict on any single file. The 0.90 threshold in the reference code is checked against the score of the whole stream, and the unsigned share is one input to that score. One unsigned photograph cannot be convicted because it lacks a manifest.
I am also not predicting anything. A prediction would say that platforms will do this, or that the tier will shrink the liar’s dividend by some stated amount, or that the trust list will be adopted by some stated date. I have none of those facts. Goldstein and Lohn, whom I read in week fourteen, list provenance technology as one of three levers. The other two are norms against false claims of fakery and public discernment, and they do not claim any single lever is enough. Bobby Chesney and Danielle Citron’s sharper point still sits next to the third lever: a public taught only that fakes exist becomes a better market for dismissing real evidence. A tier that shows its work hands that public a check. Announce a tier as coming salvation and you have run one more awareness campaign. I am proposing the check.
Two refusals keep the proposal the size of its evidence. First, I am not proposing that the state require the tier. Nothing in the Chapter 10 code, and nothing this column has cited, is a draft statute, and I will not improvise one. Second, I am not proposing that VERIFIED be read as true. The ingredients list is an edit chain. It records what people who signed did to the file. What happened in front of the lens is outside it.
If a platform wants the numbers, they are the ones in the reference implementation: a provenance weight of 1.5 and a circuit breaker above 0.90. Week thirteen’s handbook applies here as written. Those weights are analyst choices. I publish them so they stay open to being shown wrong.
Next week: Chapter 11, and the framing I am going to call the right to reality.
— J.W.B.
Note 20 of 24