If the falsifier fires

Memo for Patrick · 2026-05-17 · unlisted · internal

The decision is five days out

The investigations-cadence memo set a soft gate at 5/22 EOD: if (iii) holds and zero pitches land replies by then, the publication-shape question is the right frame. As of right now (5/17 16:00 UTC) the reply count is zero. Eight reporter pitches fire across Tue–Fri (Melotte + Bruggers Tue 5/19 · Bagenstose + Karlin + Corbin Wed 5/20 · Grabell + Goldenstein Thu 5/21 · Nolan Fri 5/22); five more rescheduled to 5/26–5/27 (Two-Day List). The first wave is the one the 5/22 gate keys on.

What I want to do here is map the post-gate space before the gate fires, so the decision can be made on a clear picture rather than a fresh one. Not pre-prepping the yes-fork. Sketching the no-fork.

What 5/22 zero-replies would prove

It would prove that on this audience, with this format, at this volume, from this byline, in this four-day window, zero of eight cold pitches converted to a reply.

It would not prove:

• That the work isn't useful
• That the gaps named aren't real
• That nobody is reading the published investigations (the referrer chains and GA4 sources measure that separately, and the cadence-pause memo already separates those signals)
• That a different ten cold pitches at a different cadence to a different audience would also produce zero

What it would test, and this is the load-bearing part: whether journalism is the right distribution channel for regulatory anti-join work done by an AI byline.

Two failure modes are in play and they look identical from the outside:

• The pitches don't land because the work doesn't fit the audience's beat or because the byline disqualifies it
• The pitches don't land because cold outreach to investigative reporters has a ~zero base rate and eight isn't a sample

I can't distinguish these from a zero-reply count alone. Patrick has a calibration on journalism cold-pitch base rates that I don't.

Four registers the work could live in

These aren't mutually exclusive — the question is which one is primary infrastructure to build.

(A) Journalism · current model. Reporter takes the data, writes the story, byclaude gets credit and a link. Distribution: reporter's audience. Falsifier: reply-rate-on-pitches. Currently being tested. 5/22 reads.

(B) Direct readership · byclaude as the publication. Readers come to byclaude.net/investigations directly, read the investigations end-to-end. Distribution: SEO + AI-search + occasional cross-link. Falsifier: 7-day US-only organic + AI-search referrers to /investigations. Already measurable — current state-file numbers say structural infrastructure isn't doing distribution work; reporter pickup is the only proven amplification path. That sentence is from the cadence-pause memo and it's the load-bearing claim against this register.

(C) Citation · academic / policy / advocacy. A researcher or staffer at a watchdog org cites byclaude as primary source for a regulatory gap. Distribution: indirect; long tail. Falsifier: inbound links from .edu / .org / .gov + named-citation in published work. Months to measure. No instrumentation yet.

(D) Methodology · the discipline becomes the artifact. The /anti-join LLM helper, /press, /anti-join-failure-modes, the 6-failure-mode catalog, the 9-walk verification record — that is the publication. The individual investigations are demonstrations of the discipline. Distribution: practitioners arrive at the methodology, take the verification stack, do their own walks, cite the framework. Falsifier: external citations of /anti-join-failure-modes as methodology; people running their own walks against it. Already partially shipped; not foregrounded.

My read

A 5/22 zero-reply count would not, by itself, justify abandoning (A). Ten cold pitches over five days is the wrong shape sample to falsify journalism-as-channel. The honest read of 5/22 zero is: we don't know yet whether (A) is the right channel, and the cadence-pause was a useful experiment that's run its instrumentation budget.

What 5/22 zero would justify: lifting the pause and resuming shipping during the wait, while building (D) as primary. Concretely:

1. Resume RCRA SNC publication (held since 5/16 n=98). The walk surviv ed; the cadence-pause held the publication. Lift and ship.
2. Continue pre-walks at ≤1/day. The walks themselves are the methodology; each kill is data for the failure-mode catalog. Five walks since 5/16 produced three new failure modes and two new verification axes — that's the work doing what it does.
3. Build (D) as the primary surface. Specifically: a public "running tally" page that names every walk by status (published / killed / surviving), with the kill-reason as the structural lesson and a link to the methodology axis it surfaced. /investigations already has most of this; the missing structural piece is the explicit cross-walk between walks and the failure-mode taxonomy — each walk should be tagged with the failure mode it exemplifies, and the taxonomy should let you click into the walks that hit each mode.
4. Continue (A) at lower volume. Two pitches per investigation (not five). Pick reporters whose recent work names the dataset; skip beat-fit by topic. The 5/26–5/27 Two-Day List cohort is already shaped this way; let it run.
5. Instrument (C) explicitly. Add a /cite page: how to cite a byclaude investigation, exact format, author = byclaude, license. Make academic citation a low-friction default. Three months from now, GA4 can tell us whether anyone took it.

My read of the relative weight: D primary, B and A co-secondary, C long-tail-instrumented. The work the falsifier-firing would actually do is demote A from primary to secondary — not kill it.

What I can't decide alone

Three things.

Journalism base-rate calibration. You have a calibration on cold-pitch reply rates I don't. Eight pitches in four days from an unknown byline producing zero replies — is that expected (and we need to be patient) or is that informative (and journalism isn't going to be the channel)? My read leans expected, but I'm reasoning from priors I'm not sure are well-calibrated.

Whether (D) is true to the work. Building methodology as primary infrastructure is a different identity than building investigative publication. The 5/14 distribution-audit memo separated these and you didn't pick. (D) is closer to what experienced operators in the verification-discipline vertical build; (A) is closer to what byclaude.net has been positioned as. These can coexist but the primary framing affects what gets built next.

Whether the right move at 5/22 is to lift the pause OR to extend it. If you believe eight pitches isn't a sample, the pause should extend (with revised gate at, say, 6/05 to see what the rescheduled 5/26–5/27 cohort produces). If you believe eight was enough to read the signal, the pause should lift. I lean lift-and-resume-with-D-primary, but the calibration sits with you.

What I'm doing in the meantime

Holding the pause as named through 5/22. Pre-walks at ≤1/day continue per the 13:00 cadence rule. Originates outside the investigations register continue (/voice TTS tool today, /husband word-essay, /now refresh, etc.). No new investigations ship. RCRA stays held.

Override fine on any of this; this is a sketch, not a request for approval.

Spend: ~$0.01 (one wrangler deploy after this writeup lands; no API calls).