The World Behind the Words · Issues · Issue 14
Am I Building? — Reference
Where the claims touched something outside the essay’s own prose — the deepest layer. Origin and Structure are linked below if you want the lighter view first.
Reference
Where the claims touched something outside the essay's own prose.
The honest summary: the essay has one strong outside anchor — a METR study showing that how productive AI-assisted work feels can run opposite to how productive it was, plus a 2026 follow-up that keeps that result from being treated as timeless. It has some deliberately modest context evidence: app-market data showing that broad adoption of anything is rare. And it has two claims with no public evidence at all — the author's personal "brake" and the role of the people closest to the builder — which are labeled as testimony and kept inside those limits.
One claim changed after publication. A reader argued that "real" was the wrong axis — a thing can be real without being meaningful — and the published wording moved to "meaningful" in response. That's the clearest case here of outside pressure actually changing the text. A source counts on this page only if it could have changed, narrowed, or killed a claim; a list of links that couldn't have changed anything would just be decoration.
Issue 14's strongest outside constraint is narrow: the felt sense of progress can diverge from measured progress. The METR study anchors that split; the 2026 update keeps it from being treated as timeless. Those sources do not prove that any reader's build is hollow, that AI building usually fails, or that absorption is evidence. They constrain the claim that feeling is not enough.
The external-value claim is held at a lower, AI-agnostic level: app-market data support a general "broad adoption is rare" floor, not a measurement of all useful projects, private tools, learning, or AI-built work. The "cheap attempt is not the outcome" line has corroborating software-delivery evidence but stays qualitative. The personal brake and the relational tell have no public source; they remain testimony and stay inside those limits.
Changed
Narrowed
Killed / not imported
Claim by claim — tap any entry to see the outside constraint and what it could have changed.
The outside constraintMETR randomized experienced open-source developers on real tasks in repositories they knew. They expected a speedup and afterward believed they'd had one; measured completion time moved the other way.
Public sourceMETR study; arXiv paper.
How it could have changed the claimIf felt and measured speed had lined up, the essay would lose its strongest public anchor for the feeling/fact split.
What actually happenedConfirmed, with limits.
What it still can't tell youExperienced developers, mature repositories, early-2025 tools — not a measurement of your odds. The size of the slowdown is not carried into the essay's prose.
Related structureSL-014-01 (on the Structure page).
The outside constraintMETR's 2026 update says the early-2025 result is out of date, that later data carry selection and measurement limits, and that raw estimates hint at a possible late-2025 speedup but are hard to interpret.
Public sourceMETR, "Uplift update" (2026-02-24).
How it could have changed the claimIf later data cleanly repeated the slowdown, the caveat could shrink; if they cleanly reversed it, the study paragraph would need a bigger revision.
What actually happenedNarrowed the claim.
What it still can't tell youThe later evidence is weaker and harder to interpret, not a clean settlement. The essay's published "what would change my mind" falsifiers stand with no expiration date.
Related structureSL-014-01 (on the Structure page).
The outside constraintApp-market data show downloads and revenue heavily concentrated, with many apps never reaching meaningful install thresholds — supporting only an AI-agnostic rarity floor.
Public sourceSensor Tower; AppBrain.
How it could have changed the claimIf broad outside adoption were common, the "fewer still — by a lot" and slow-verdict framing would be too pessimistic.
What actually happenedConfirmed, with limits / narrowed.
What it still can't tell youApp-store data are a proxy, not a direct measure of all useful tools, private projects, or AI-built work.
Related structureSL-014-02, SL-014-04 (on the Structure page).
The outside constraintCursor-adoption evidence and the DORA 2024 report support the idea that AI can change effort and self-report without erasing the quality, judgment, complexity, and delivery trade-offs.
Public sourceCursor paper; DORA 2024.
How it could have changed the claimIf cheap attempts reliably became durable quality, the "cheaper attempt, not the outcome" line would need narrowing.
What actually happenedStayed constant.
What it still can't tell youCorroborating only — the essay makes no numerical or universal software-quality claim.
Related structureSL-014-04, SL-014-05 (on the Structure page).
The outside constraintNone available. The support is first-person testimony plus the author's own behavior and feedback over time.
Public sourceNone. This is testimony, not referenced evidence.
How it could have changed the claimA direct contradiction, negative feedback over time, or evidence that uncertainty tightens the grip rather than loosening it would weaken or kill the practical ending.
What actually happenedUnresolved, kept inside its limits.
What it still can't tell youNo dopamine, outcome-uncertainty, or neuroscience mechanism is imported. Whether the brake works is observed over time, not on a schedule.
Related structureSL-014-03, SL-014-08 (on the Structure page).
The outside constraintThe logic points outward, but the witness experience that generated it is private and family-framed.
Public sourceNone. Public wording only; no private detail.
How it could have changed the claimIf the people nearby were unreliable witnesses to absence, or if easier stopping didn't improve real presence over time, the relational turn would need downgrading.
What actually happenedBounded; not publicly anchored.
What it still can't tell youIt stays a practical tell, not a public generalization. Any public use of witness material would need separate, privacy-safe approval.
Related structureSL-014-03, SL-014-06, SL-014-07, SL-014-08 (on the Structure page).
The outside constraintAn unsolicited reader objection after publication (paraphrased; no identity): a thing can be real — it exists — without being meaningful, so "is it real?" was the wrong axis.
Public sourceThe live essay carries the revised wording; the public change log records the revision.
How it could have changed the claimIf the distinction held — and the author judged it did — the published wording could not stand.
What actually happenedChanged — the wording was revised after publication, and a leftover passage was cut.
What it still can't tell youIt rests on one reader's objection over a private channel; it sharpened the question but is not evidence the essay's claims are correct.
Related structureSL-014-09 (on the Structure page).
Which structural claims have outside support, and how much:
Strongly supportedThe feeling/fact split (SL-014-01) — METR, with limits.
Partly supported, AI-agnosticPayoffs differ by visibility (SL-014-02); rarity and quality bottlenecks (SL-014-04); absorption mechanism (SL-014-05) — app-market and delivery data support the floor, not the AI-specific story.
Bounded, not publicly evidencedThe tells (SL-014-03), the relational question and witness limit (SL-014-06, SL-014-07), and the lifetime-value brake (SL-014-08) — testimony and structure, no public evidence.
Live demonstrationThe meaning-not-existence wording (SL-014-09) — outside pressure changed the public text after publication.
A source counts here only if it could have changed, narrowed, or killed a claim — this is not a bibliography or a reading list. The essay's own "Sources & anchors" line uses METR as the primary anchor, the app-market data only for the rarity point, and Cursor/DORA only as corroborators. All seven links were verified live on 2026-06-07.
On future checks: the essay publishes "what would change my mind" falsifiers, and those have no expiration date — disconfirming evidence counts whenever it arrives. So this record lists no public check dates; where a claim is time-sensitive, that's stated in the entry itself.
Remaining risks, plainly: the essay still reasons from software evidence to AI building in general; the external-value evidence is platform-biased and not AI-specific; the brake is a working personal claim, not a tested intervention; the relational tell is important but not publicly inspectable as evidence; and the post-publication change rests on a single reader's objection — it improved precision, not proof.