The enforcement machinery

Rigor isn't a promise here. It's a build step.

Every productivity tool says it's “science-backed.” This page shows what we do differently: the tests and linters that fail our build when a claim drifts, a citation is folklore, or a number ships without its evidence record. No competitor can rehearse this demo — it's our CI.

Runs on every commit

Three jobs stand between a claim and production.

content-policy

Greps every user-facing string in the web app and backend for banned citations and overclaim patterns.

Percentage productivity claims, folklore citations, replication-crisis casualties — the build fails before the copy ships.

$ python scripts/check_content_policy.py

citation-compliance

Walks the AST of every Evidence record in the codebase and checks each citation against the allowed-anchor list.

Any scientific claim without a post-2002 peer-reviewed anchor, and any anchor on the banned list — at the data-structure level, not the copy level.

$ pytest tests/test_citation_compliance.py

rendered-claim tests

Renders the actual marketing pages in CI and asserts on the DOM the visitor sees.

Claim drift between what the evidence system says and what the page renders — the failure mode most evidence pages quietly develop.

$ vitest run src/__tests__

# the gate, verbatim

$ python scripts/check_content_policy.py

content-policy: OK

$ pytest tests/test_citation_compliance.py -q

every Evidence record checked · anchors post-2002 · banned list clean

Banned by name

The claims we refuse to ship — and what we cite instead.

These are the productivity industry's favorite numbers. Each is on our banned list because it fails our evidence bar; several are pattern-matched by CI so they cannot ship by accident.

“It takes ~23 minutes to refocus after an interruption”

Not cleanly sourced to any paper — HCI folklore that citation-laundered its way into half the productivity industry. We use a per-switch figure only as an internal planning estimate, always labeled operational, never cited.

What we use instead

Leroy & Glomb 2018 (attention residue, four studies) for the science of switching cost.

“You can only do 3–4 hours of deep work a day” (as a cited ceiling)

The meta-analysis this usually cites (Macnamara 2014) is about variance in deliberate practice — it establishes no within-day focus cap. Our weekly soft-bound is labeled an internal default.

What we use instead

An operational planning bound, graded ‘op’, disclosed in the formula.

Willpower as a depletable fuel tank (“decision fatigue”, ego depletion)

A central casualty of the replication crisis — large pre-registered replications found effects near zero. We never cite it, in any form.

What we use instead

Nothing. Some popular claims just don’t survive scrutiny.

90-minute “ultradian rhythm” work cycles

Kleitman’s basic rest-activity cycle is a sleep-research construct; the daytime 90-minute work-cycle version is empirically weak. We frame breaks as “after sustained focus,” not on a clock.

What we use instead

Albulescu 2022 (meta-analysis: micro-breaks restore vigilance).

“Mornings are the analytical peak” (for everyone)

Chronotype swamps the population mean — peak windows shift 4+ hours between extreme larks and owls. A universal window is wrong for most individuals.

What we use instead

Adan et al. 2012 (review) + your own rMEQ score, mapped continuously.

Any specific % productivity gain (“30% more productive”)

Unsupportable at that precision. Our CI literally pattern-matches this claim shape and fails the build.

What we use instead

Predicted minutes per move, calibrated against what your past moves actually delivered — then measured.

The determinism contract

Same calendar, same preferences, same plan. Bit for bit.

plan_id

Every plan is identified by a hash of its inputs. Identical inputs produce an identical plan — which makes every placement auditable and every write idempotent.

bit-identity tests

New features ship behind flags, and tests pin that the flag-off path produces byte-identical plans to the previous release. Personalization is opt-in, never drift.

measured, then promised

Every rescheduling move records a predicted gain and, later, the realized outcome. Predictions are calibrated against your own history — the promise shrinks until it's honest.