Intuition · The Black Box Next Door · Deeper record

Development Record — The Black Box Next Door

A public-safe account of how the essay was drafted four ways by four AI models, chosen by feel, argued over line by line, and published — including the origin story that used to sit on the essay page.

Reader note

This is not a raw machine log. Local file paths, shell traces, and private operational details are omitted or summarized. The three outside-model drafts and the external review are excerpted at the level needed to show what actually happened; the full texts are preserved in the author's local archive.

Intuition is a lighter-rigor surface. This record shows how the essay was made and where it was pushed back on. It does not prove the essay's bet is right — the essay itself calls it a bet.

The experiment, in one paragraph

Every statistic in the essay comes from the verified source work behind the stricter Signal & Noise issue Unseen and Unenforced: Who Pays for AI's Grid? That fact base was frozen into a fact pack with its qualifiers attached, and the same assignment went to four AI models: Claude (working in-session with full context), and — via a clean-room paste packet the author ran manually — Grok 4.5, Gemini 3.1 Pro, and ChatGPT 5.6 Ultra. The author judged the four drafts on feel. Grok's landed best and became the base. It was then cut, partially restored, re-punctuated, externally reviewed, and finished with the author's own wording. The author owns the result, opinions and errors included.

Sources: the July 12, 2026 development session (author + AI editorial process), the archived model outputs and adjudication artifacts, and the issue edit-log. Private operational details are omitted.

The ask · an op-ed bet on Issue 19's facts

The author · the assignment · July 12
I would like to use the content in Signal & Noise Issue 19 as a foundation/reference for an op-ed to publish in Intuition based on my personal opinion that if people had full transparency of the potential benefits and costs, that acceptance of data centers would increase, not decrease. In other words, it's the lack of transparency that is fueling public distrust and pushback. And since this is an opinion piece, we can be more "one sided" and persuasive focusing more on the benefits than the negatives. Also I would like to aim for the following format: 600 to 800 words, persuasive argument. Also, just to see how different models write, please prepare a draft yourself and a cut and paste prompt I can provide to other SOTA AI models so we can compare writing styles and I can compare the gestalt/human landing. What do you think?
AI editorial process · assessment excerpt · July 12
The idea works, and it's a natural pair. Issue 19 proved, neutrally, that the transparency machinery is missing (no receipt, fragmented say). Your op-ed bets on what happens if it existed — a claim S&N couldn't publish (untested prediction) but which is exactly what Intuition is for. One honest caution: the evidence for "transparency → acceptance" is genuinely mixed. Gallup's 71% is siting opinion — some of it is about land, noise, and character, and would survive perfect ledgers. So the draft scopes the bet to the cost fight and frames it as your bet, not a finding. And I'd argue the persuasive one-sidedness should live in emphasis, never in the facts — the concessions are what make it persuasive, because a reader can't dismiss it as industry PR when it demands the one thing PR avoids. A transparency op-ed that cherry-picks would refute itself.
AI editorial process · first draft + clean-room packet built · July 12
Produced a first in-session draft ("The Secrecy Tax," 800 words) plus the cut-and-paste packet for the other models: the same thesis and format rules, and a ten-item verified fact pack distilled from Issue 19 with every qualifier baked in — the fair-rates-and-modeled-rise pairing, the disputed $9.3 billion attribution, funder labels on the industry studies, and the search-scoped "no rule found" phrasing. Models were instructed to use only those facts and add nothing. Both the draft and the packet were then adversarially checked by four independent verification passes against the Issue 19 corpus, which caught real errors before anything went out — including an absence claim I had overstated from "no source found in this research" to "no state requires."

The receipt-register exchange

The author · a register worry · July 12
While "receipts" has become a popular slang term for proof (as in "show me the receipts"), in a serious essay about infrastructure, taxation, and civic policy, it can feel slightly out of place or too colloquial for the target audience. Consider swapping out the metaphor to rely on more grounded financial and civic terms — like final tally, actuals, and hard data — which maintain the punchy, authoritative rhythm of your writing without distracting the reader. what do you think?
AI editorial process · adjudication excerpt · July 12
The slang sense is "receipts," plural — this piece uses "the receipt," singular and literal: the itemized record of what was actually paid, issued after the fact. That pairing with "the price tag" is doing precise conceptual work. And the colloquialism is partly the point: the essay's core demand is accounting "in words a household can read." Price tag and receipt are household objects. "Actuals" and "final tally" are the vocabulary of the people who currently control the ledger; adopting their register would subtly concede the piece's own argument. The middle path: keep the metaphor, add the grounded terms as ballast — one gloss at first use ("the receipt after — the audited actuals") and swap the single highest slang-cadence spot.
AI editorial process · applied on the author's direction · July 12
The author took the middle path, and the ballast survives in the published Grok-based text as the "audited ledger" and reconciliation language. The exchange set a rule the rest of the edit kept returning to: keep the felt register, anchor it with one precise term where it matters.

The four-model experiment

The author · returns the outside drafts · July 12
Before we proceed, I would like your feedback on the op-ed's from other SOTA AI Models (Grok, Gemeni, and Chat GPT) [three full drafts pasted — preserved in the local archive and excerpted below]
AI editorial process · comparison excerpt · July 12
The clean-room packet held: zero fabricated numbers across all three — every figure traces to the fact pack with its qualifiers intact. Because the facts were fixed, what you're actually comparing is pure editorial judgment. Grok 4.5, "The Black Box Next Door": the most felt voice, the radical-candor move ("I cannot prove that is the industry standard, and I will not pretend I can"), the best aphorism of the batch ("Opacity turns a contested model into a rumor. Sunlight turns it into an argument we can have") — and 1,179 words against your 800 cap, a 47% overage. Gemini 3.1 Pro: in-range but a briefing memo wearing an op-ed hat; its close pastes the assignment thesis verbatim — the biggest human-landing tell in the experiment. One genuinely good line: "you turn them down because the process treats you as an obstacle." ChatGPT 5.6 Ultra: the most editor-shaped op-ed, with the single best line anyone produced — "earn consent instead of demanding faith" — but it quietly dropped the two strongest benefits cards to buy its pace. Convergence was the quiet finding: all four drafts independently landed on secrecy → suspicion → "the worst version of every number," and receipts as the remedy. With facts held constant, the models agree on the argument; they differ in voice, curation, and closing craft.
The author · the selection · July 12
These are all excellent and a hard choice, but honestly, maybe because it's the most "felt", the Grok essay lands the best with me personally. While keeping it's essence, radical-candor move, aphorisms and genuinely authored ending, any tweaks you would make before publishing? perhaps changing the middle paragraphs in some way?

Compression, restoration, and the em-dash register

AI editorial process · compression pass · July 12
Cut the Grok base from 1,179 to 800 words under one rule: cut detail, never honesty. Every binding qualifier survived — the fair-rates/modeled-rise pairing, the auction-shortfall-not-blackout framing, the $9.3 billion attribution with the dispute visible, funder disclosures, the scoped absence claim. What paid the bill was specification: named companies, megawatt figures, the study millions, statutory mechanics.
The author · restores the middle · July 12
Even though we will go over the 800 word limit, let's restore the Grok's original, version of paragraph 3 starting with "The costs that do show up in market machinery" and paragraph 4 "Benefits get buried the same way costs do" because extra words to increase instant understanding. Also, if possible, we should try to reduce at least 2/3 of the em dashes. I know em dashes are a versatile punctuation marks, but seeing them over and over throughout the essay is kind-of jarring. What do you think?
AI editorial process · both applied · July 12
Agreed on both counts — Grok's original two-sentence handling of the $9.3 billion figure is objectively clearer than the compressed version, and twenty em dashes in a short essay had become the prose's default connective (and an AI-writing tell). Both paragraphs restored verbatim; em dashes cut from 20 to 6, keeping only the thesis punch, the studies range, the receipt-definition pair, and the closing peak. Replacements varied — colons, parentheses, a semicolon, sentence splits — so no substitute became the new tic. Body settled at 971 words with the cap consciously relaxed by the author.

Approval, verification, and the external review

The author · approves with checks · July 12
I approve v0.3 with one last tweak [the dek's "bill math" became "the potential bill"]. Can you re-verify/check CL-012 and the pJM 2028/20 auction. Let's include a brief one line note re multi model origin story. Otherwise ready to proceed to next step.
AI editorial process · verification results · July 12
Two checks before publication. The essay's absence claim — no state found requiring a recurring, public, after-the-fact reconciliation households can read — was re-verified beyond the original research corpus: Oklahoma's new act (pre-project notice, separate billing), five newly enacted state ratepayer-protection laws (Alabama, Tennessee, South Dakota, Nebraska, Florida — all upfront allocation mechanisms), and California's one-time cost-shift assessment. No counterexample; the claim stood as scoped. And a timing hazard: the region's first uncapped capacity auction had closed, with results expected July 14. Publishing before that date kept the essay's "latest capacity auction" passage accurate as of publication.
External review (run manually by the author in ChatGPT) · excerpt · July 12
Verdict: publish after fixes. Its best catch: despite the careful qualifiers, the sequencing made the Gallup result feel like corroboration of the secrecy explanation — "the poll establishes the breadth of opposition, not its cause." It also pressed on certainty: the bookended prediction "present[s] the increase in acceptance as established rather than as the essay's personal wager," and lines like "that is how you manufacture reflex rejection" "outrun the opening disclosure that the author cannot prove the secrecy pattern."
AI editorial process · adjudication · July 12
Accepted: the evidence/inference separator ("The poll cannot say why."), the scope-once fix ("My argument is simple" → "My bet is simple"), a paragraph split, and moving disclosure material out of the essay's closing flow. Rejected, with the author's veto rights noted: a structural reorder that would have gutted the process paragraph's payoff, and the softening of the aphorisms the author had selected the Grok base for.
The author · overrides one rejection with his own wording · July 12
I agree with your accepted changes and also would incorporate Finding 4 suggestions - something close to: "I think that is how…," and "When a process works this way, the worst story becomes the easiest one to believe." Otherwise, please proceed to the packaging set.
AI editorial process · final text · July 12
Both formulations applied as written: "and I think that is how you manufacture reflex rejection," and the process paragraph now opens "When a process works this way, the worst story becomes the easiest one to believe" — which turns the essay's closing "the worst one people can invent" into a deliberate escalation: believe, then invent. Final body: 983 words, six em dashes, every fact-handling rule intact.

Publication, and the disclosure the page shed

The author · publish + card + gates · July 12
Yes, please commit and push, and please create a social card - perhaps generate a typographic house-pattern card from the brand assets, and proceed to the post-publication gates.
AI editorial process · shipped · July 12
Published inside the pre-results window, with the homepage and issues indexes updated. The social card reuses the family typographic pattern — navy, Iowan serif, gold rule — carrying Intuition's spiral mark and coral dot. The audio companion was rendered through the established synthetic-voice chain (4:47) and handed off for the author's listening check and upload. The stricter sibling's public record now links back here: the Issue 19 origin page carries a "Paired opinion essay" section.
The author · fresh eyes after publication · July 12
Looking at Intuition Issue 3 with fresh eyes post-publication, and it looks to me like we have "disclosure and disclaimer" overload. To Mitigate, please make something close to the following modifications For Intuition Issue 3: [shorten the companion note] And let's remove the Origin note... It's excellent information, but it really belongs in The World Behind the Words territory, if anywhere. And Please proceed to create the development record.
AI editorial process · this page · July 12
The essay page was simplified the same day: a one-line companion note, and the origin note removed from the essay's foot. Its contents now live here — which is where a story about showing your work belongs.
The origin note, relocated

Origin note. This piece began as an experiment: the same verified fact pack, drawn from the reporting behind Signal & Noise Issue 19, went to four AI models. The draft that landed best (Grok's) became the base; I cut it, reworked it, and own it, opinions and errors included.

Originally the final line of the essay page; moved here on the author's post-publication judgment that it belongs with the development record.