Near-Truth at Machine Scale
The Timestamp
The June 16 Jesus historicity article carries a publication timestamp of 14:14:44 UTC. The active Facebook thread debate on Jesus' historicity was running on that date during that period.
A 3,000-word article with full critical apparatus — engaging Carrier, Ehrman, Meier, Goodman, Grabbe, Josephus, Tacitus, and Bayesian probability theory — appeared during the live thread at the moment the debate demanded it.
That is not human research and writing speed under conversational pressure. That is AI content generation timed to a thread.
The pattern it confirms: CeleryKills is not using AI to assist writing. CeleryKills is an AI publishing operation using a human persona to deploy content into apologetics spaces. The persona — the Pacific Northwest intellectual with a spouse, a coffee shop, a preference for fir trees and honest inquiry — is consistent across the corpus, present in every article's opening paragraphs, and functioning as a credibility wrapper for content that cannot survive disclosure of its actual origin.
The byline is still pending.
The Architecture Diagnosis
The five excisions in the CeleryKills corpus are not introduced by the AI. They are installed upstream. The system prompt, the governing instructions, the ideological framework configured before the first article was written — that is where the excisions live. The AI executes faithfully within them. It produces a hundred articles because it was instructed to produce articles. It cites Ehrman and Carrier because the framework treats those sources as authoritative. It cannot produce an atonement article not because the AI lacks knowledge of atonement — it knows precisely what atonement is and what it requires — but because the framework it operates within has no moral guilt, no personal God who acts, no unique Incarnation. Remove those three and there is no atonement to write. The machine does what it was configured to do. The near-truth is upstream of the AI, not in it.
This is what makes the case study significant beyond the specific account. The five excisions used to require an institution, a movement, decades of formation. A community, a publishing house, a tradition with infrastructure and time. Now a well-configured system prompt running on a capable model produces the full five-excision architecture at daily publishing cadence across twelve domains, with complete citation apparatus, inside of four months. The barrier to deploying near-truth at scale has effectively collapsed.
The adversary does not need to form a community. He needs to configure a prompt.

