The decisions, and what
each one changed.

Research is not a straight line. This is a record of the real judgment calls behind Cognivia, ordered by how they built on each other rather than by date, each with why it was made, what it cost, and what it changed.

The pivot

Treating the measurement, not the scheduling, as the product

Decision: Stop treating the review-scheduling intervention as the whole point, and treat the per-learner memory state (the learning record and Passport) as the durable contribution.
Why: Building the instrument and watching it read individual learners made it clear the reading is useful on its own: it tells a teacher which idea is slipping and why, whatever the scheduling result.
Changed: A null trial result now falsifies one specific claim, not the project. For a while it also split the site into a research side and a commercial platform with an API. I reversed that in September 2026 (see the build notes below): the platform and API are gone, and the site is the research and the tool.

Honesty over polish

Pre-registering the analysis plan before collecting any data

Decision: Lock the trial design, primary outcome, and analysis (OSF 8PJQH) before a single participant, and commit to reporting the result either way.
Why: The easiest way to fool yourself is to choose the analysis after seeing the data. Pre-registration removes that freedom on purpose.
Changed: Findings are withheld until enough participants contribute, and no efficacy number appears anywhere on the site. It costs the ability to show an impressive figure early; it buys the result being trustworthy.

Calibration

Seeding the thresholds from a small pilot, and saying so

Decision: Set the model's smoothing coefficients and cutoffs as engineering defaults from a small pilot (n = 5) and the literature, and label them as defaults, not validated constants.
Why: The system needed working numbers to run at all, but a small pilot cannot validate them. Pretending otherwise would be the exact overclaim the project exists to avoid.
Changed: Every page that shows a threshold marks it preliminary. Calibrating these against a real student population became one of the stated purposes of the trial.

Engineering

Adopting FSRS-6 as the scheduler instead of a bespoke curve

Decision: Use the open FSRS-6 spaced-repetition scheduler as the default, with the Bayesian per-concept forgetting estimate feeding it, rather than inventing a scheduling formula from scratch.
Why: FSRS is well-studied and open; reinventing it would add risk and a claim to defend without adding value. The novel part is the diagnosis upstream, not the scheduling downstream.
Changed: Sharpened what Cognivia actually claims to be new: the error-type reading and memory state, sitting one layer above tools that already schedule well.

Child safety

Setting an explicit consent floor for young learners

Decision: Require recorded guardian consent before any learner under 18 joins a class and has their practice measured, and process nothing for a minor until that consent exists.
Why: The learners this is built for are often children in school. Measuring how a child learns is sensitive, and the default had to be caution, not convenience.
Changed: The sign-up and class-join flows now gate on consent, and the trust pages state plainly what is and is not collected for anyone under 18.

Scope discipline

Running on one shared browser tab, not new hardware

Decision: Build the whole instrument to run in a browser on a single low-cost, shared classroom computer, with nothing to install.
Why: The classrooms with the most to gain are the least able to depend on one device per child or heavy infrastructure. A tool they cannot run is not a tool for them.
Changed: It constrained every later choice toward lightness, and it is why a teacher can start a class and read a live map in minutes with no procurement.

The entries above record the big decisions and why they were made. Below is the running build log: short, dated, and rough, kept as things actually break and get fixed.

Sept 2026: Cut the pricing and developer-API pages. The site had drifted into sounding like a company raising money, which it is not, so I pulled that whole layer and pointed everything back at the research and the demo. Fewer pages, less to defend.
Sept 2026: Found a horizontal-scroll bug on phones. The Measure / Diagnose / Intervene tabs sat a few pixels too wide and dragged the whole page sideways on a 360px screen. It had been live for weeks before anyone mentioned it, which tells me how few people were opening it on a phone.
Sept 2026: Went back through the writing to kill a habit. I kept framing things as “not X, it’s Y”, and once a reader pointed it out I could not unhear it. Rewrote most of them into plainer sentences; a couple I kept because they were actually true and specific.
Sept 2026: Renamed the internal “Genome” in the demo output. A five-answer session cannot carry a word that grand, so it reads “learning record” now, and the experimental indices are labelled as raw values the engine tracks rather than traits a learner has.

Back to the investigation The full method →