Fragment 18 should not have cleared that fast.
It did anyway.
First validator round on Mira network came back cleaner than it deserved. Claim decomposition had already split the parent response into five fragments. Four of them looked routine. Fragment 18 carried the awkward sentence... a policy exemption tied to an old circular and a footnote nobody reads unless something already smells wrong.
The panel didn’t smell wrong.
Mira's First-round weight stacked clean enough that my hand left the mouse.
cert_state: provisional
That was the first mistake.
The fragment looked stable on the first pass. Evidence hash pinned. Citation path resolved. No obvious conflict in the validator trace. The audit log stayed boring. And boring fragments on Mira get less suspicion than they should.
I opened the source anyway.
Wrong paragraph.
Back up. Scroll again. Or... No.
There it was. Same document family, same regulator, same phrasing shape... but the clause Fragment 18 leaned on had been narrowed later by an addendum sitting deeper in the citation chain. First-round validators stopped before it.
Not fake evidence. Not malicious validators. Just a shorter walk through the graph producing a cleaner answer than the full path deserved.
Fragment 18 stayed green while I was still tracing the branch.
cert_state: provisional
Fragment 19 picked up speed right after. Parent response looked healthier because 18 had already leaned the aggregate toward “mostly verified.”
Fragment 19 only looked cleaner because 18 had already borrowed trust from the round.
You could feel the operator bias set in. Mine too.
I was already reading 19 like 18 had been right. That’s on me.
I flagged the branch and reopened the trace.
No one else had yet.
Round two started with the fragment already carrying the shape of near-certification on Mira, even without a cert hash. One validator in round two went deeper.
reject.
Confidence dipped, but not enough to erase the first-round comfort. Another node abstained on context insufficiency. Better than affirming. Worse than killing it.
The fragment sat there in a half-safe state that costs the network more than a clean failure.
Still there. Still eating the round.
I checked the Mira's verifier accuracy ledger. The early affirmations came from nodes with good histories, decent reward share, low penalty exposure. Nothing to suggest sloppy behavior. That almost made it worse. The penalty model punishes persistent bad assessment, not one plausible false positive tucked behind a shallow citation path.
So no slash. Not yet.
Just more work.
And the easy fragments were already getting paid.
The correction loop widened under Fragment 18. Additional evidence attachment. Reopened document path. One validator branch pulling the addendum, another still anchored to the old language. Same regulator. Same subject. Different effective claim.
The first pass had simply stopped too early.
Verification queue already thickened behind it. Two easier fragments cleared while 18 kept absorbing attention. The fast work kept paying out. The messy fragment kept asking for more of the round than anyone wanted to spend.
I caught myself hoping the second reject would come quickly.
Didn’t.
Spinner on one mid-weight validator. Still evaluating. I knew what I wanted it to say before it said anything, which is exactly when I stop trusting myself in these rounds.
The audit log finally changed.
Supplementary branch attached.
Correction path recognized.
That helped. A little.
Not enough to wipe the first-round bias clean. Early affirmation on Mira does that. Even provisional weight leaves residue.
Another reject landed.
Now the fragment looked sick.
Too late to be clean.
Fragment 18 had already survived the stage where everyone was willing to call it routine. Now it had to be dragged backward through the same verification window that should have caught it earlier, and every minute it stayed alive kept the parent response in a worse kind of uncertainty.. not empty, not broken, just contaminated by something that cleared too easily the first time.
cert_state: provisional
Still there.
No hardened cert. No clean rollback.
18 still hadn’t hardened.
The @Mira - Trust Layer of AI Verification queue had already learned from it anyway.