1080 Learner Design Rationale and Evidence Base

02What 1080 Learner does

2.1A session

A learner works under an alias. The first step, Tune in, sets up access before any evidence is collected: text size, read aloud, a display theme (Daylight, Dusk or High contrast) and input by tap or keys. Every setting is printed on the final page, so the adult knows how the tasks were taken. The learner never sees whether an answer was right; every response gets the same reply, “Saved.” “I’m not sure” and “Skip” are full buttons, so nobody is pushed into a guess that would look like evidence. There are no timers.

An evidence engine, a set of written rules, works from the learner’s responses. The demo is a fixed-order slice: equal-sign items come first, then fractions. Within a topic it gives a second item because one response cannot settle anything, and it gives a third equal-sign item only when the first two disagree. For each choice it records a plain reason, for example “Chosen because one response cannot settle anything; a second equal-sign item is needed.” Each subject records its own stop reason from a fixed list, such as clear evidence, an item limit, or no suitable item left. When a subject stops at its item limit, the page adds: “This does not show what the learner can or cannot do.” Choosing items by the links between skills is part of the proposed next engine (Section 3.3), not of the demo.

2.2The page

Each kind of claim has its own typeface (monospace for observation, sans serif for interpretation, serif for action) and always a text label. A mark beside each skill shows how much usable evidence stands behind it: filled (clear), half (some evidence), split (mixed), hollow (not enough yet) or dashed (not offered, for example for access reasons). The legend reads: “The mark shows how much usable evidence there is, not how well the learner did.” A “What this is not” footer closes every page. Any single response can be opened in the Prism, which splits it into What they did, Also possible and Ask next.

2.3A worked example: Heron, a sample learner

The public demo has three sample learners. Here is the math path for one of them, Heron. A verification script recomputed every key below across 192 answer-type paths through the demo’s math items (each response counted as correct, a tagged wrong answer, or another answer; paths that use Skip or “I’m not sure” were not enumerated), and it caught a wrong key that we planted to test it.

Two answers of 13 match the tag for reading “=” as “write the answer next.” (On EQ-1, 17, the total of all three numbers, carries a second tag.) Two consistent responses made the equal-sign evidence clear. The engine therefore skipped EQ-3, a true/false item it gives only when the first two responses disagree, and used the last two of the demo’s four math items on fractions. Both fraction responses add the same amount to top and bottom. One of the two is a three-option choice that can be picked by chance, so the evidence is marked Some, not clear. Heron’s page reads, in part:

• What we saw: “Wrote 13 for 7 + 6 = □ + 4.” “Wrote 13 for 5 + 8 = 6 + □.” “Both answers equal the total of the left side.” “Clear evidence that the equal-sign idea is not yet secure on these items (2 responses).”
• What it might mean: “The equal sign may be read as ‘write the answer next’.” Also possible: “Copied a familiar worksheet habit without reading the right side”; “Misread where the box was”; “Rushed after the first item”.
• Try tomorrow: “10-minute balance talk: true-or-false number sentences, then build both sides with cubes.”
• How you’ll know: “In 1-2 weeks, give 6 + 9 = □ + 8. An answer of 7 fits the balance meaning. 15 would be the total of the left side again; 23 would be the total of all three numbers. Different item, so it is not a growth comparison.”
• Equivalent fractions (marked Some): “Wrote 11 for 2/3 = □/12.” “For ‘Which fraction is equal to 3/4?’, chose ‘4/5’.” “Both answers add the same amount to the top and the bottom.” “Some evidence that the equivalent-fraction idea is not yet secure on these items. One more example would confirm it.” Its held-back check is 3/5 = □/20: “An answer of 12 fits multiplying the top and the bottom by the same number. 18 would be adding the same amount to both; 4 would be the multiplier written as the answer.”
• Stop reason (math): “Stopped because the 4-item limit for this part of the demo was reached.” “This does not show what the learner can or cannot do.”

Notice what the page does not say. It never states that Heron “has a misconception.” The rule for naming a likely misconception requires every plausible setting of the engine’s assumptions to agree, and in the demo it never passes, even on 13 and 13. The page reports the literal feature of the responses and offers the misconception as one reading among several. We will not loosen that rule to make pages sound more confident.

Replay as another learner swaps in a different sample response set: different evidence gives a different page and a different move. Of the 192 answer-type paths, 189 stop at the four-item math limit and carry the caveat. Cedar, the sample learner who answers all four math items correctly, stops because “every idea checked has clear evidence,” and that stop has no caveat. One response is never enough for clear, but it can give a tentative Some evidence reading. Heron’s single correct literacy answer reads “Some evidence on these items: 1 of 1 correct. One more example would confirm it,” and its move is to wait: “Nothing to change yet. Ask one more question on the same idea before reading anything into this.” The sample learner Delta gave one fraction answer (4 for 2/3 = □/12). The page marks it Some evidence, not clear, but still offers a reading (“The multiplier may have been found and then written as the answer.”) and a reteach move. Cedar’s and Delta’s single literacy answers, both wrong, read “Not enough evidence yet (1 response).” A thinking task (tracking a hidden marker across a grid) is reported only as literal counts, such as “Matched 1 of 1 one-move trial, 1 of 1 two-move trial, 1 of 2 three-move trials.” It comes with a question to ask and a free support idea, and never with a “because,” a comparison or a link.

2.4The research app (private research app, not yet open)

The research app holds 210 items across 63 sampled skills: 117 math (30 early, 54 middle, 33 secondary) and 93 literacy (30 early, 33 middle, 30 secondary). It also has three thinking tasks, groups of up to 30 learners, and exports as text, JSON, print or a ZIP of separate profiles. The demo items were written for the demo and are not from this bank. The demo runs the evidence rules proposed for the app’s next version (engine v2) with a fixed item order. The research app has not adopted engine v2.

03Design principles and the research behind each

3.1Wrong answers that carry information

DesignEach item is written so that a common wrong rule produces a distinct wrong answer, and that answer is tagged to a written error pattern. For 2/3 = □/12, 11 is tagged “added the same amount to top and bottom” and 4 “wrote the scale factor.” On the multiple-choice item, 3/8 is tagged “changed only the bottom.”

ResearchIn the US federally funded DIMS project, every multiple-choice item was written so that at least one wrong option maps to a known misconception (Wylie & Wiliam, 2007) [F10]. The same authors argue that items should start from the teacher’s next decision, so that wrong answers show what is not understood and ideally what to do about it (Wylie & Wiliam, 2006) [F14]. The NRC (2001) describes the Facets work of Minstrell and Hunt, in which each answer choice maps back to a way of thinking [D06]. Wang et al. (2020) describe questions whose wrong answers each reveal a common misconception, so educators are not left to guess why [F12]. Sadler’s astronomy items, as described by Wylie and Wiliam (2007), show that a misconception-based wrong option can attract students just below average more than the lowest performers, so right/wrong totals can hide what students believe [F11]; and Wiliam (2005) uses his TIMSS pair to argue that a wrong rule should lead to a wrong answer [F13]. The literature calls these “diagnostic questions.” We say error-pattern items, because 1080 Learner is not a diagnosis of any learner.

The patterns come from math research. Equal-sign studies use open equations and true/false structure items [MATH-05], and equations with operations on both sides are generally the harder kind [MATH-04]. 1080 keeps operations on both sides in its session items and its fresh check, so the two stay comparable. For fractions, “whole number bias” describes whole-number knowledge interfering with fractions (Siegler, Thompson & Schneider, 2011, describing Ni & Zhou, 2005) [MATH-10]. The IES fractions guide names treating numerator and denominator as separate whole numbers as a common misconception [MATH-11]. It also describes changing a denominator without the matching change to the numerator [MATH-12], and calls it critical to find students who hold such misconceptions and discuss them (Siegler et al., 2010) [MATH-13].

LimitThe IES examples are about fraction addition. Extending them to equivalence, as our “added the same amount” and “changed only the bottom” tags do, is our inference [MATH-11, MATH-12]. Siegler, Thompson and Schneider (2011) also argue that fraction arithmetic errors often reflect confusion about the right strategy, “rather than a consistent whole number bias or other systematic misunderstanding” [MATH-10b]. That is one reason 1080 never treats a single answer as a pattern. Good items of this kind are hard to write: Wang et al. (2020) note that it is challenging to write good items of this kind, in which every wrong answer is a plausible distractor [F12b]. No study has yet confirmed that real learners who give a tagged answer are using the tagged rule.

3.2Evidence before labels: “Also possible”

DesignNo skill is called clear on one response. One response can give only a tentative Some evidence reading, which still comes with a reading and a move and says that one more example would confirm it. For clear, the engine needs at least two responses, and it works out each skill’s state under eight combinations of unknown settings: a careless-error rate treated as a range (0.08 to 0.30), a guessing rate set by item format, and how consistently a pattern appears. A skill is “clear” only if all eight agree. Every interpretation carries at least two “Also possible” explanations, and the build fails if one is missing. Pattern wording stays literal: the page reports the feature the responses share (for example, “Both answers equal the total of the left side.”). Opening a response in the Prism gives an Ask next question.

ResearchThe NRC (2001) calls assessment a process of reasoning from evidence [D01]. It notes that assessment information is typically incomplete, inconclusive and open to more than one explanation [D02]. A result is an estimate from a small sample that can vary with format, timing or a student’s health that day [D03]. Naming a specific flawed rule takes more than one item [D05], and the model of learning, the tasks and the interpretation method must be designed together [D04]. Wylie and Wiliam (2007) define a misconception by consistency, not by a single wrong answer [F08]. Matthews et al. (2012) warn that either/or results on one item type hide real differences [MATH-06]. The DINA model (Junker & Sijtsma, 2001, as described by Lee, 2017) has a “slip” term (wrong despite mastery) and a “guess” term (right without it) [D11]. 1080’s error and guessing ranges make the same allowance.

LimitNo opened source tests 1080’s gate. Applying Matthews et al. to it is our inference.

3.3Adaptive questioning and honest stop reasons

Design(proposed: engine v2, not yet in the research app) The proposed engine picks the item most likely to settle the next teaching step and gives priority to skills whose earlier steps look secure. Each item appears at most once. Fresh-check items are held back. No strand takes more than 40% of items, and each subject is capped at 16 items or a time limit. The skill graph chooses items but never sets a reported label. In the demo, the item order is fixed (equal-sign items, then fractions), the limits are four math items and one literacy item, fresh checks are held back, and each subject records its own stop reason.

ResearchEvidence-centered design separates evidence rules, which read a single response, from the models that combine evidence across tasks (Mislevy, Almond & Lukas, 2004) [D07]. It also names an activity-selection process that decides the next task and when to stop [D08], and an assembly model that sets evidence targets and balance constraints [D09]. 1080 writes its targets and constraints down as the rules above. A simulation study of adaptive tests found that tests that were too short (a mean length under 10 items) did not give good ability estimates for test takers in the low range, and it recommends combining variable stopping rules with a minimum number of items (Babcock & Weiss, 2012) [D10]. The 2014 Standards published jointly by AERA, APA and NCME ask that reports say in simple language what was covered, what results mean and how they are meant to be used (American Educational Research Association et al., 2014) [D12]. 1080 prints its evidence counts, stop reason and coverage.

Internal simulation(synthetic learners only) We tested the proposed rules on simulated learners (250 to 400 per setting) on an eight-skill test graph. These results describe our design under our own assumptions, not real students. With typed-answer items, wrong confident calls stayed at or below 2.2% for every simulated learner type, in about 12 to 13 items per subject. With four-option multiple choice, a next step was named for only 20 to 27% of simulated learners, which is why typed answers are the authoring target. When we broke the engine on purpose (one response settles a skill, and no guessing is assumed), wrong confident calls rose from 0.9% to 5.0%. The test harness flagged it against a 5% check; the proposed 3% threshold catches it with more margin. Babcock and Weiss studied ability estimates, not per-skill readings, so their work does not support our item counts [D10].

3.4Next moves and fresh checks

DesignEvery skill block ends in a suggested move, shown next to its evidence, and a held-back check in the same format, to be given after a delay. The educator decides whether to use the move, change it or set it aside [D17].

ResearchFormative assessment is defined by use [F01] and works through specific guidance rather than marks [F03]. Growing Success asks for specific next steps [F16]. The IES data-use cycle [MATH-17] maps onto 1080’s last three questions. Its evidence is rated low, so we call the loop sensible practice, nothing more. For the equal-sign move, a randomized study (N = 90, mean age 8) found that practice with nontraditional formats such as 17 = 9 + 8 improved children’s understanding of equivalence more than traditional or no extra practice (McNeil et al., 2011, ERIC abstract) [MATH-08]. The What Works Clearinghouse (2014) rated that study “Meets WWC standards without reservations” for a contrast of 60 students in grades 1 to 3 [MATH-09]. That is one study, and our balance talk itself is untested; the research supports only its non-standard-format ingredient. Equal-sign understanding is linked to equation solving in middle school, even after controlling for students’ results on standardized mathematics tests (Knuth et al., 2006) [MATH-02], and to basic algebra in Grades 2 to 6 (Matthews et al., 2012) [MATH-07]. Both findings are correlational. For fractions, the IES guide recommends number lines [MATH-14], and area models and other visual representations for understanding fraction computation [MATH-16], but notes that the number-line evidence comes mainly from whole-number studies [MATH-15]. For timing, an IES guide recommends re-exposure after several weeks to several months (Pashler et al., 2007) [MATH-18]. Our 1 to 2 weeks is shorter, and that guide concerns remembering, not whether a re-check is a sound gauge.

LimitAs described by Shute (2007), Sleeman et al. (1989) found that simply re-teaching algebra content worked as well as feedback built on detailed error analysis [F07]. Researchers describe feedback findings as inconsistent and highly variable [F06]. Black and Wiliam’s (1998) typical effect sizes of 0.4 to 0.7 come from a 1998 review of formative classroom practice in general and say nothing about 1080 [F02]. We therefore claim no effect. “Try tomorrow” is a practical re-teach, and the fresh check decides whether it worked.

3.5Families

DesignOne press turns a page into a Family Note: What we saw, Try at home (5 minutes), and What this is not. It targets roughly a Grade 4 reading level, prints on paper, reads on a phone and can be edited by the educator. For now it is in English only. Families receive it through the educator; there are no parent accounts.

ResearchUS Department of Education guidance recommends that parents’ requests for records held by a provider go through the school or district (Privacy Technical Assistance Center, 2014) [EQ-14]. In a randomized field experiment, weekly one-sentence teacher messages to parents reduced the share of students failing to earn credit, mainly by preventing drop-outs. Messages about what to improve had the largest effects (Kraft & Rogers, 2015, published abstract) [EQ-16]. School email to families at or below the poverty line stayed flat after 2007 (Kraft, 2016) [EQ-17], which is one reason the note works on paper. READY4K! sent parents short, specific activities built on family routines, and its evaluation reports gains in parental involvement (York & Loeb, 2014) [EQ-18].

LimitThese studies involved US high-school credit recovery and preschool texts. None tested a note like ours.

3.6Access and Universal Design for Learning

DesignAccess comes before evidence: text size, themes, read aloud, tap or keys, a tap or key alternative for every drag, no timers and Esc to pause. A task a learner cannot access is recorded as “Not offered: access,” never as wrong. Plain mode removes all art and motion.

ResearchCAST’s UDL Guidelines 3.0 (2024) list adjustable text size, spacing and display [EQ-01], name text-to-audio as one of the easiest ways to widen access [EQ-02], and recommend several physical ways to respond [EQ-03]. They also call for informative rather than comparative feedback [EQ-04] and for flexible pace, session length and breaks [EQ-05]. W3C’s Accessibility Guidelines Working Group (2024) recommends WCAG 2.2 as the conformance target [EQ-06]. WCAG 2.2 requires single-pointer alternatives to dragging [EQ-07] and says sign-in should not depend on remembering a password unless an alternative or assisting mechanism is provided (the criterion lists further exceptions) [EQ-08]. We therefore plan for the adult to open each session and the learner to choose an alias, with no learner password.

LimitThe current build specification aims at WCAG 2.1 AA. We plan to raise the target to 2.2 AA and publish the gap list against it. Voice and switch input are not built yet. Read aloud on a decoding item changes what the item shows; that point is our inference, so the page records when read aloud was on.

3.7Privacy by design

DesignLearners use aliases, and educators never need to enter a learner’s name; adults own the workspace and can export or delete it; each learner has a separate profile; and Milestone Teachers does not sell learner data. The published privacy notice identifies the processors and United States data storage, the retention period and deletion route. The public sample shows status only, with no correctness, grouping or ranking. It uses sample learner responses; it is not a real student record. 1080 makes no predictions.

ResearchIn October 2025, Canada’s federal, provincial and territorial privacy regulators jointly called for privacy-protective defaults and support for pseudonyms in classroom technology [EQ-09]. They asked vendors to collect only what is necessary and proportionate [EQ-10], named predictions that could shape a child’s educational path as a key risk [EQ-11], and asked for language children can understand [EQ-12]. The Information and Privacy Commissioner of Ontario’s Digital Privacy Charter for Ontario Schools (2024) includes a pledge to make privacy-protective settings the default [EQ-13]. The FTC’s COPPA guidance, which concerns children under 13, counts as personal information “a persistent identifier that can be used to recognize a user over time and across different websites or online services” [EQ-15]. The guidance does not settle whether a classroom alias meets that definition. We therefore treat aliases as personal information and do not rely on them as an exemption. Labels carry their own risk. The NRC (2001) warns that fixed-ability beliefs can lead educators to label students “bright” or “dull” [D15], and teacher expectations have been shown to affect outcomes (Rubie-Davies & Hattie, 2025) [D16]. So 1080 gives no ability labels, and every reading is provisional and paired with a check that can overturn it.

LimitThe privacy notice is published, and learners see a plain-language explanation before opening their session. The notice describes United States storage, retention and deletion. This paper does not independently evaluate legal compliance or the service’s live data handling.

04What 1080 Learner is not

• Not a diagnosis. It names no conditions and no causes in a learner. Thinking tasks are reported as what happened in the task, never as a trait.
• Not a score. It gives no scores, norms, percentiles, grade equivalents, levels or IQ, and it never compares learners [F05, EQ-04].
• Not for high-stakes decisions. It is not for placement, eligibility or special-education planning. The Standards say that a decision with major impact on a student should not rest on one test alone [D13], and that where adequate evidence is lacking, users should be told to weigh results accordingly [D14].
• Not validated. Validation: none yet. Studies are still needed on consistency, fairness across groups of learners, retest effects and usefulness.
• Not a replacement for the educator. A page is one student product. Ontario policy asks for triangulation with observation and conversation [D18]. Opening a response in the Prism gives an Ask next question, and educators decide what to do with every suggestion [D17].
• Not a claim about effect. We do not claim that using 1080 Learner improves learning.

05Honest limits

1. Not field-tested. No real learner has used 1080 Learner in a pilot. Every engine figure here comes from simulated learners under our assumptions.
2. Our routines are untested. The balance talk and number-line reteach are supported only in their ingredients [MATH-08, MATH-14, MATH-15, MATH-16].
3. The fresh check is not a validated gauge of learning. Its support is indirect [MATH-17, MATH-18], and our 1 to 2 week delay is shorter than the spacing guide’s.
4. Targeting an error may not beat good re-teaching [F07], and feedback findings vary widely [F06]. That is why the check decides.
5. Correlation only. The equal-sign link to algebra is correlational [MATH-02, MATH-07].
6. Two tags are our inference. “Added the same amount” and “changed only the bottom” for equivalence extend addition examples [MATH-11, MATH-12].
7. Small item sets. The demo has nine math items (five session items, two held-back fresh checks and two extension items), a practice item, one literacy item and four spatial trials. The app’s 210 items sample skills; they are not a curriculum.
8. Format limits. If items stay multiple choice, an honest engine names a next step for only about one learner in four (simulation).
9. A strict naming rule. The demo never names a likely misconception. We will loosen the rule only if consented response data justify it.
10. Thinking tasks. No research cited here supports them. They are reported as literal observations and lead only to free support ideas.
11. Access gaps. Voice and switch input are not built. Read aloud on decoding items needs a sourced policy.
12. English only. Items and directions are in English. We are not promising a French version.
13. Data handling has limits. The current privacy notice explains aliases, United States storage, retention, deletion and the groups the product does not serve.
14. Evidence from elsewhere. The workload data come from England [F18], and the family studies come from other age groups [EQ-16, EQ-18].
15. The author is the developer and seller.
16. One response can give a tentative reading. Only clear needs two responses. A Some evidence reading can come from a single response, and it still carries a reading and a move.

06Future evidence

These are questions for any future research, not results and not a condition of access. No real-learner field study data exist yet.

Participation. Any future study would use a separate invitation and an approved data-handling plan. It would not be designed to show that 1080 Learner improves learning or reported as if it did.

Questions a future study could examine.

1. Educator-rated usefulness. After using a page, the educator rates its usefulness for planning on a short fixed scale. They also say whether they used the move, changed it or set it aside, and why.
2. Agreement with the educator’s own read. Before opening a page, the educator records their own view of the same skills (secure, not yet or unsure) and any error pattern they have seen in class. We compare the two records. Agreement does not prove either is right, so each disagreement is discussed with the educator and checked against later work.
3. Change on fresh checks. Where a move was tried, the educator gives the held-back check after the delay. We record whether the response fits the target idea, repeats the pattern or shows something else. With no comparison group, change cannot be credited to the move, because ordinary teaching, practice and chance also produce change. That question needs a later comparison study.
4. Time cost. Learner minutes come from session timestamps. Educator minutes spent reading the page and preparing the move come from brief self-reports.
5. How the engine works on real responses. Counts of each stop reason, of “not enough evidence yet” results, of cases where “Also possible” changed what the educator did, and of access settings used. Every reported problem is logged.

Reporting. Any future report should give counts with their denominators, include null and contrary results, keep observation separate from interpretation, contain no individual learner data, and name no school without that school’s written permission.

07Access and current terms

Explore a sample report: https://milestoneteachers.com/1080/sample-report. It uses sample data; it is not a real student record.

1080 Learner access. Prices are in Canadian dollars. One-time payment; no automatic renewal.

• Educator access, CAD 149: one educator account; 40 pathway starts; 365 days from first adult activation.
• School access, CAD 990: one school campus, up to 8 named educator accounts; 240 shared pathway starts; 365 days from first adult activation.

A pathway start is consumed when an educator creates a learner session; resuming it does not consume another start. After payment, find the included code in My purchases, sign in with the educator Google account that will manage the workspace, and redeem the code at milestoneteachers.com/1080/access. Your checkout email and educator Google account can differ. A full refund can be requested within 30 days of purchase at admin@milestoneteachers.ca. A refund ends active access; existing reports remain readable and exportable for 30 days, subject to retention requirements or earlier deletion.

Access terms (scope, refunds and data handling): https://milestoneteachers.com/legal/1080-access/2026-10-12-r1

Questions: admin@milestoneteachers.ca