Engine Changelog
The capability record of the LEF engine behind this site: what it can do that it could not do before, in plain language. Current engine: v2.5.0. Patent Pending.
v2.5.0
Each student read against their own level, and every read held to its own floor
2026-10-18- Cross-subject leads are now read against each student's own usual level in each subject, not pooled across students. The lasting link (students strong in one subject tend to be strong in the others) is shown apart and never read as a lead, and a lead is named in words only when it clears the margin of noise. The pooled method is known to report false or wrong-signed leads.
- Fixed: a room's effect on a pattern of learners is now judged against that pattern's usual movement over the same number of test windows the students actually spanned. It had assumed three, so a one-year, fall-to-spring read made every room look below usual.
- Fixed: a pattern with no measured baseline now gets no effect claim. It had been read against a baseline of zero, which reported a group's raw change as if it were the room's effect.
- Placement suggestions now need at least five students who entered at a similar level, the same floor every other read holds, and they show how many improved as a count ("8 of 8 improved"), never as a percentage.
- New: the engine has a place to record in-school supports (who received one, when it started, how many sessions, whether family took part) and a repeated behavioral, mental and emotional measure. Its read scores each support per pattern of learner against similar students who did not receive it, and guards against students referred at a low point, who rebound with or without help. It waits for a provider's records.
- Wording: the reads are rooted in single-case research (each student against their own baseline) and add a comparison with students who learn the same way in other rooms; the board now names that comparison for what it is.
v2.4.0
The engine watches itself (recorded late)
2026-07- A student's pattern now carries how differently they move across subjects (one subject fading faster, recovering in one but not another), not only whether their subjects move together.
- On every re-derivation the engine grades its own pattern library: whether the groups blurred, how many students changed group, and how many split by subject, with its own plain-language notes.
- Within a grade and subject, students whose records stall at the same goal area are grouped so one re-entry point can serve them; the teacher decides every move.
- A ledger of first sightings: the first time each pattern shape appeared, how often it has been seen since, and whether it was seen in synthetic, trial or district data.
- Fixed: a pattern is marked as seen in district field data only when a real district has joined; a public trial upload no longer counts as field data.
- The board and its data open only to signed-in pilot or owner sessions.
- These arrived between 13 and 25 July 2026 and were entered here on 18 October 2026.
v2.3.0
Joining a district never relabels a student
2026-07-08- When schools join a district, their pattern libraries merge: shared patterns combine and strengthen, patterns unique to one school carry over. Every student keeps the pattern group their own data earned; only the vocabulary that names it widens. Verified: the student-to-pattern mapping is identical before and after the join.
- The engine tells an administrative join apart from a data event using its own evidence: unfolded new data always re-derives the pattern boundaries, paperwork alone never does.
- Fixed: a legacy counting path could bleed member tallies across separately-scoped populations. Counts are now written only by the derivation that measured them.
v2.2.0
The engine keeps receipts
2026-07-04- Every roster ingestion cycle now writes a structured, append-only record of the engine’s forward projections BEFORE outcomes arrive; the next test window grades them. Validation becomes a chain of timestamped receipts, not a claim.
- Memory decay is now computed at read time: mastery ages continuously between uploads instead of freezing at the last file, so silent erosion accumulates the way memory actually does.
- Cross Subject names the students its lead points at right now, not just the pattern.
- Pathway Topology reads the same prerequisite graph at single-student resolution: pick a student, see their own gates and exactly what blocks them.
v2.1.0
The engine grades itself
2026-07-03- Local calibration loop: on every ingestion the engine re-grades its own past projections against what actually happened next, side by side with the no-signal baseline. Unflattering results are kept.
- Profile assignment history: every recluster is recorded, so type stability is measured across time instead of asserted.
- Seventh profile dimension: each student’s movement relative to their own classroom, so a student who outruns or trails their room is visible as a pattern.
v2.0.0
Types come from the data
2026-07-02- Profile types are no longer designed; they are derived from the student population itself, re-measured on every ingestion. Names come from each type’s strongest measured traits.
- Every type carries its own MEASURED expectation (typical movement per test window, with spread); contradiction flags are calibrated per type, so volatile patterns need a larger break to be genuinely off-pattern.
- A student who fits no pattern is reported as an outlier, never boxed into the nearest type.
- Every module on the feature board is driven by live engine signals; nothing on screen is staged.
v1.0.0
The pathway kernel
2026-04- Mastery health model with memory decay: what a student demonstrably had, aged by how long it has sat idle.
- Prerequisite graph routing: topics unlock and block on real prerequisite standing, per student.
- NWEA MAP ingestion and MTSS tier language as the shared vocabulary with school leadership.
Entries describe engine capability, never customer data. Numbers shown elsewhere on this site come from the engine live at view time; this page only records when each capability arrived.