Skip to Content

Trust & Privacy

Amal works with measurements of children’s reading, and it makes recommendations a teacher acts on. That only earns trust if the system is honest about what it knows, careful with what it holds, and clear that the teacher, not the software, is in charge. This page sets out how that works, in depth.

The teacher decides; the system only proposes

Every output of the platform is a proposal. The status it computes, the profile it composes, the support plan it recommends, the grouping it suggests, and the response decision it reaches are all suggestions the teacher reviews and can accept, change, or set aside. Nothing activates on a child without the teacher, and nothing escalates a child to more intensive support on its own.

When a teacher overrides a recommendation, the reason is recorded by choosing from a fixed list of reason codes, never by typing free text about a child. The same holds for context a teacher wants to add: it comes from a controlled set of context flags, not a typed note. This keeps the record clean and respectful, and it keeps the teacher’s judgment central without ever asking the teacher to write something that could follow a child.

Growth language only, no labels

No clinical term and no external program label appears anywhere a person can read. The platform describes a child in terms of what they can do and what to work on next. It does not name diagnoses, and it does not borrow the names of international reading frameworks or scales. The wording for parents is stricter still: no scores, no internal terms, only plain, encouraging language about growth.

This is not a tone preference. It is enforced. A language-safety layer screens user-facing copy and blocks the banned terms before anything reaches a screen or a report.

No single overall score, by design

There is no overall reading number anywhere in the platform, and this is deliberate. A single percentage would hide exactly the detail a teacher needs, and it would invite comparing children on one axis. Instead, reading is always reported per skill and per reading area, with each measure resting at a clear state such as at, near, below, or well below the benchmark, or not yet assessed. When a measure has no comparable evidence, it is marked “not assessed” and excluded from any summary, never counted as a zero.

Children’s data is handled with care

A few commitments govern how a child’s record is kept:

  • No free-text notes about a child are ever written. Overrides and context use fixed codes and flags, so the system never stores a typed opinion that could be misread later.
  • Records are append-only. Measurement-relevant records are never edited or deleted in place. A correction is a new entry, so the history a decision was based on stays intact and faithful for later review.
  • Access follows roles and the school structure. A teacher reaches only their assigned classes, and every record is scoped to its organization so one school’s data cannot bleed into another’s.

For the full picture of what is collected and how long it is kept, see the foundations pages on roles and access and quality and determinism.

Scientific honesty: the system declines to decide on too little data

The most important honesty rule is that the platform stops when it should not decide. When the evidence for a skill is split or too thin, it marks the skill “not enough data yet” and waits, and everything downstream waits with it. No profile is built on a not-yet-decided skill, no plan is recommended from it, and no alert fires from it. The platform would rather say “I do not know yet” than guess.

The same caution governs progress monitoring. No status changes on a single data point. A sharp drop, defined as a fall of at least twenty percent across two comparable sessions, raises a review for a teacher to look at, never an automatic failure. And before a plan can be judged “not working,” the platform checks whether the plan was actually delivered, so a child is never pushed to more intensive support because of a plan that was never really taught.

No AI touches measurement or the child’s experience

Language models do not participate in measurement, and they do not appear in front of a child. No language model estimates a child’s reading ability, picks an item, scores a check, or runs alongside a child’s session. Statistics measure, and people decide. Language models are confined to helping a teacher draft their own writing, such as a note home, and only after the teacher reviews and owns what is sent. The child’s tasks, scores, and the short warm-up before a session are all rule-based and deterministic.

Where to go next

How Amal Works