Every week I pick one product, use it first-hand — twice — and publish a teardown where every opinion traces to a screenshot, a review, or a number. Six months later each prediction gets scored publicly — being transparently wrong beats being vague.
The log is empty — on purpose. The first teardown drops Mon · Jul 20, 2026.
Slot open — picked from the backlog. Candidates need a 14+ curiosity-and-evidence score to make the cut.
Slot open.
Slot open.
Every phase has a hard definition-of-done — a checklist that must be fully green before the next phase starts. If a week compresses, usage evidence and the pre-publish gate are protected first; if even that doesn’t fit, the week is skipped rather than published thin.
Hard gates before any scoring: I can reach the core loop myself, there’s enough public material (a founder long-form plus 15+ user data points), no undisclosed conflicts. The opening question is written down before research starts — and it has to be one smart people would disagree on.
Fresh account, every onboarding step screenshotted, time-to-value clocked, paywall hit on purpose. Then a second session 72+ hours later: does the core loop still pay off, and what pulls me back? One session shows onboarding; two show retention.
The original launch post plus the most recent long-form interview — if the story changed between them, that divergence is a finding. Extract the stated problem, wedge, and target user, then judge whether the product expresses it, citing the exact screen.
Competitor list built by a selection rule, not memory. Pricing tables from same-day screenshots of live pages. Distribution claims backed by observable signals — ad libraries, SEO footprint, sales headcount — never intuition.
20+ data points across 3+ platforms, last 12 months. A complaint only counts as recurring at 5+ independent instances on 2+ platforms — below that it’s labeled an anecdote. Each piece discloses how the sample is skewed, in one line.
Every number has a source or a stated assumption — zero silent guesses. Fresh headcount and funding data, per-user inference costs for AI products, ranges instead of false precision, and the one assumption that breaks the conclusion if it’s off by 2x.
What I’d steal, what I’d worry about — every worry with a steelman: what would have to be true for it not to matter. The verdict answers the opening question, and the prediction gets an outcome, a deadline, a resolution criterion, and a confidence %.
On publish day: volatile facts re-verified, every negative claim classified as sourced fact, attributed quote, or clearly-framed opinion, screenshots scrubbed of anyone’s personal data — and the founder gets a right of reply 24–48 hours before the piece ships.
Reader corrections issued visibly within 48 hours — never silent edits. Every prediction lives in a public ledger and is scored quarterly against its resolution criterion; calibration stats only once 20+ predictions have resolved.
Traceability. Every opinion points to a screenshot, review, quote, or number. No evidence, no claim.
First-hand only. If I can’t sign up and use it myself, I don’t tear it down.
Right of reply. Founders see the worry list before publish and get 24–48 hours to respond. Their answer — or their silence — goes in the piece.
Visible corrections. Reader corrections are handled within 48 hours as visible edits, never silent ones. The correction habit is the trust engine of the whole log.
Scored predictions. Six months after each teardown, the prediction gets scored against its resolution criterion — publicly, hit or miss, with the reason. Being transparently wrong builds more trust than being vague.