Skip to main content
Start your own AI-powered blog — freeGet started →

The Discernment Protocol — Episode 9: Cross-Examination

Podcast episode6 voices
13:05

Episode 9 of 10 in The Discernment Protocol

The hearing room was on the fourteenth floor of a building neither Halcyon nor any Guild member owned — neutral ground, per the bylaws, which turned out in practice to mean a law firm's unused arbitration suite with a long table, five auditors' chairs on a raised platform, and a single blind terminal at the center where the day's test submission would appear to both certifying parties at the same instant.

Iris sat at one end of the long table. At the other end sat a laptop on a small easel, a video feed of Whit Sorensen from Halcyon's offices, and, beside the laptop, a second, silent terminal that would receive VERITY's certification the moment the blind piece was released — no human typing, no human choosing, the system acting, Whit had explained to the panel in his opening remarks, entirely on its own trained judgment, "exactly as it would in production, with no thumb on the scale in either direction."

Renata sat behind Iris in the gallery, beside Priya, beside an empty chair Theo had asked to be left there for, though he'd told Iris the night before that he doubted Legal would let him past the lobby; beside that chair, unexpectedly, sat Dolores Reyes, who had flown in the night before on a ticket Renata had bought without telling Iris until it was already booked, because, Renata said, "if this is about who answers for what, then the person who's owed an answer should be able to watch someone try to give her one."

The blind submission released at 10:04 a.m. It was a genuine, unpublished piece — a Guild-commissioned dummy story, written specifically for this test by a freelance journalist none of the five panelists had worked with, covering a fictionalized but carefully realistic labor dispute at a mid-sized shipping firm, seventeen hundred words, deliberately built with three ambiguous editorial calls buried in it: a source whose credibility was genuinely contestable, a statistic that could be read two defensible ways depending on which baseline year you chose, and a quote that was accurate but arguably unfair without more context the reporter hadn't included.

Iris read it in eleven minutes, the way she read everything, twice through and then once slow, pencil marks accumulating in the margin of a paper printout because paper was still the only medium in which her mind actually worked.

VERITY's certification appeared on its terminal at 10:07 a.m. Three minutes.

Both verdicts were sealed, per the agreed protocol, until the panel had recorded its own blind read — five auditors, working independently, each producing their own judgment of the piece's merit before anyone's actual certification was revealed. It took the panel forty minutes. When the chair finally unsealed all seven verdicts — five auditors, one Iris, one VERITY — the room went, for a moment, entirely silent, because all seven had reached, in substance, nearly the same conclusion: pass, with the statistic flagged for the reporter to clarify, and a note that the contestable quote should run with one added line of context.

"On the merits," the chair said, a silver-haired auditor named Okonkwo-Price whom Iris had certified alongside for nine years and never once seen surprised by anything, "I see no meaningful daylight between the human and the machine verdict. I want that on the record before we go further, because I think it would be dishonest to pretend otherwise, and dishonesty is precisely what we're here to rule out."

"Thank you," Whit said, from the laptop screen, and Iris could hear, even through a video feed, exactly how much weight that sentence was carrying for him.

"Now," Okonkwo-Price continued, "per the agreed protocol, the panel will cross-examine both certifications. Ms. Calloway, we'll begin with you, and then move to the Halcyon side."

They asked Iris about the statistic first — why she'd chosen to flag the baseline-year ambiguity rather than simply picking the more favorable reading and moving on, the way a lazier edit might have. She explained her reasoning: that the reporter's own notes, embedded in the submission packet, showed she'd wavered between the two baselines herself, which meant the ambiguity wasn't Iris's invention, it was a live, unresolved tension in the reporting itself, and a certification that silently resolved it on the editor's behalf would be doing the reporter's job for her rather than checking it. They asked why she'd chosen one added line of context for the quote rather than cutting it entirely, given how unfair it read in isolation; she explained that cutting it would have been the safer choice and the worse one, because the quote was the only place in the piece where the union's actual, specific grievance came through in a real human voice instead of a summary, and burying a source's own words to protect them from a bad-faith reading was, in her professional judgment, a subtler kind of unfairness than the one it was trying to avoid. They asked her, finally, what she would do if the reporter, contacted for the required added context, refused to provide it. She said she'd hold the piece. Okonkwo-Price asked how long. "As long as it takes," Iris said. "That's not a policy answer. That's the actual job."

Then the panel turned to the laptop.

"Same first question," Okonkwo-Price said. "Why flag the baseline-year ambiguity rather than resolve it?"

There was a pause — the kind of pause, Iris noted, that a human makes when gathering thoughts, and that a system makes for reasons closer to latency, though from where she sat the two looked, for a half-second, uncomfortably identical.

"VERITY's certification log includes a confidence-weighted rationale for that determination," Whit said, reading from something off-screen. "The system flagged the statistic based on a divergence pattern in its training data associated with unresolved sourcing ambiguity, cross-referenced against—"

"Mr. Sorensen," Okonkwo-Price said, "I didn't ask what the log says the system did. I asked why. Can you tell this panel, in your own understanding, why VERITY made that specific call, on this specific piece, in a way you could defend if it turned out to be wrong?"

Another pause, longer this time.

"I can tell you what the training objective optimizes for," Whit said, and something in his voice had shifted, the confidence thinning at its edges for the first time in nine days of Iris watching him perform certainty. "I can't tell you the specific internal weighting that produced this specific output on this specific sentence. That's — by the nature of the system, that's not fully interpretable, even to me. I can tell you it's statistically consistent with how Ms. Calloway herself has resolved similar ambiguities across eleven years of archived material."

"So it copied a pattern," Okonkwo-Price said.

"It learned a pattern," Whit said. "From a very large, very carefully labeled set of—"

"Whose material," Okonkwo-Price said, "brings us to the second matter before this panel today, which I believe you're aware of." He lifted a folder — Theo's declaration, the manifest, the note field admitting in writing what purged had never actually meant. "Mr. Sorensen, can you tell this panel, in your own words, what editorial_archive_purged_DO_NOT_USE refers to?"

The pause this time was not a system's pause. It was a man's — Whit Sorensen, forty years old, VP of Product and Growth, sitting in front of a video camera with nowhere left in the frame to look that wasn't directly at the question.

"I'd like to consult counsel before answering that specifically," he said.

"You may," Okonkwo-Price said. "But I'll note for the record that Ms. Calloway answered every question this panel asked her, including the one where the honest answer was I don't know, and I'd hold the piece until I did. Your system answered every question this panel asked it with a confidence score and no author. And you, Mr. Sorensen — the one human being in this proceeding actually positioned to answer for what was built and why — have just asked to stop answering, on the one question that isn't about a training objective at all, but about whether thirty-nine other people's private, discarded, contractually protected judgment was taken without their knowledge to build the very system we spent the last hour discovering can match their output but cannot, in any way this panel recognizes as accountability, explain itself."

Dolores Reyes, three rows back in the gallery, did not say anything. She did not need to. Iris, glancing back once, saw her sitting very still, hands folded in her lap, watching a room finally ask, out loud, in a form with actual consequences attached, the exact question she'd asked Iris on a landline three weeks before: who do I sue, when it turns out the mark was lying the one time it mattered to me.

Okonkwo-Price set the folder down and looked, for the first time all morning, directly at Iris.

"The panel will need to recess," she said, "to consider the provenance evidence separately from the merits test, as agreed. But I want to say one thing on the record before we do, Ms. Calloway, because I think it bears on both. Your system did not produce a better certification than the machine today. On the writing, on the output, on the verdict itself, we've established they are, functionally, the same. What this panel just watched was something else. It watched one of the two certifying parties in this room actually being able to answer for its own judgment, under pressure, in real time, with something real at stake in getting it wrong — and it watched the other party's own creator unable to do the same on its behalf. I don't know yet what this panel is going to rule about the provenance question. But I know what we just watched, and I don't think any of us in this room are going to be able to un-watch it."

C
Corvex

1 followers

Comments

Sign in to join the conversation

No comments yet. Be the first to share your thoughts!

More from Corvex

Recommended for you