The grammar of what happened.

The fade: support that plans its own exit

Every response in a WellDash plan step is a configuration — four dials recorded together: strategy, cue, assistance, manner. Independence work is the discipline of moving two of those dials one rung at a time, on evidence.

• The assistance ladder, faded downwards: hand_over_hand → partial_physical → supervision → none. A permanent floor where safety requires it is legitimate — floor is not failure.
• The cue ladder: none · environmental · gestural · modelled · verbal_enabling · verbal_direct. modelled is the demonstration rung — show the person first, hands off entirely. A demonstration is a teaching dose, never doing-for.
• Who did the doing, every time: performed_by — person, shared or carer_did_for. Doing the task for someone is a different kind of record, never a heavier rung of assistance. The doing-for → shared → person curve is the reablement line, dated, computed from the same rows.
• What the hands did, when hands were on: touch_kind — bearing (taking the weight) and/or guiding (directing the movement). Most real hands-on care is both at once, and the record says so.
• The working rules: one axis per revision — change the physical rung or the cue, never both, so the record can attribute what worked. Struggle steps support up now; success steps it down next time. Acquisition runs most-to-least; maintenance runs least-to-most. Never fade past the authored floor. Consent rides every touch.

The evidence behind the shape: scoring physical assistance and cognitive prompting concurrently, as two rails, is warranted by the performance-assessment literature (PCPT-ALF, Bowen 2017; PASS’s nine-level assist hierarchy, Chisholm 2014). The fade mechanics are classical: graduated guidance fades touch (Foxx & Azrin 1973); time delay fades the cue (Touchette 1971); support shifts contingently with success and struggle (Wood & Middleton 1975); fading direction differs between acquisition and maintenance (Demchak 1990; Libby et al. 2008). Over-helping is iatrogenic — prompt dependence and learned non-use are the recorded harms of helping too much. And UK reablement carries the same posture at guideline level: NICE NG74 tells practitioners to observe and guide rather than automatically intervene, the SPRU evaluation names the practice “standing and watching”, and SCIE contrasts reablement’s doing-with against traditional care’s doing-for.

The easier ask: same wage, not same shape

Functional communication training (Carr & Durand 1985) is among the best-evidenced behavioural interventions, and it is a recommendation of NICE NG11: develop an alternative behaviour that achieves the same function. In WellDash the replacement is a first-class plan step, not a note in a document.

• The ask is taught as its own step (behaviour_response) with strategy teach_fct, paired with reinforce_success — honour the ask the moment it lands on first uses, or the teaching loop breaks. Thin the acknowledgement as the ask becomes habitual.
• The height of the arc is the ask the person starts themselves, unprompted — visible in the live record as the cue rung falling to none on dated rows for the taught step.
• The law: equivalence means the same wage, not the same shape. The record checks it as same occasion, different form, same settle — the index holds each form’s settle record under the same function, so an equivalence claim can be examined against the person’s own rows rather than believed.
• At the pre-verbal end, the lineage runs through Intensive Interaction (Nind & Hewett): the ask does not have to be a word to earn its wage.
• Behind the palette sits an evidence-attached strategy library — forty authored entries covering ten functions across four kinds of response, each carrying its example, its non-example and its citations, so the same equivalents can be offered consistently across shifts rather than reinvented on each one.

The context pieces: an ABC whose A can be counted

An ABC record is only as useful as its antecedent limb, and in most software the A is prose. In WellDash the context is typed at capture, so it computes.

• Setting events are recorded as typed tokens across five classes — physiological, sensory and environmental, social and relational, routine and temporal, emotional — beside typed physical rule-outs.
• Each episode carries a ctx_key — its first setting token — and the empty value is itself meaningful: a recorded “nothing unusual” is information, not absence.
• Three body clocks run continuously from the person’s own recorded numbers — night_wakings, days_since_last_movement, refusal_day_count — and annotate every episode when it is read. The bands on those clocks are declared reading conventions on the person’s own history, never clinical thresholds, and never part of a key: annotation is joined at read.
• The differentiation law: which background state loads which behaviour — whether a shut bowel loads tapping while broken sleep loads hitting out — is learned from the person’s own co-occurrence counts, never decided by a taxonomy. Distal and proximate layers are recorded for honesty but are never load-bearing.
• Flag, never force: when the vocabulary has no token for what the carer said, the gap is flagged so the vocabulary can grow. Nothing is squeezed into a wrong token or abandoned to prose.

Body first: the safety spine the mortality reviews demanded

People with learning disabilities die avoidably when bodily illness is read as behaviour — diagnostic overshadowing is the standing finding of CIPOLD (2013) and the NHS LeDeR programme, and NICE NG11 puts physical causes first. WellDash builds that priority into the record’s structure, not into a training slide.

• The override: when anything is unexplained, new or out of character, the body question outranks every behavioural reading — body_first in the record. Constipation until proven otherwise.
• PINCH ME at capture: seven first-pass prompts — pain, infection, nutrition, constipation, hydration, medication, environment — each mapping onto typed causes in a 24-cause differential (rule_out_type). The mapping fails closed: a prompt with no authored mapping is an error, never a guess.
• The convergence lanes: what happened now, what is different from this person’s usual, and what is known about their life are shown together for a human — never merged by the machine.
• The dechallenge is the strongest evidence the record holds: treat, then resolve, dated. A synthetic example: laxative → bowels open → the tapping stops. When the same chain has happened before, the record shows it as n-of-n on this person’s own history.
• A clinician-authored graph of symptom → consideration → disposition triples is used lawfully: as typed vocabulary and as authored disposition edges — human-activated only, with provenance on every row.
• Escalation is an authored, human-activated edge armed with the evidence. The machine never routes an emergency call.
• And medication is part of the same record: STOMP is the standing warrant that a chemical response is never the answer to a body problem.

The function index: what the record computes

The four pillars share one engine. Every episode is recorded as a transition — what was happening before, what was done, what happened after — and the record keeps how the person settled afterwards, read against their own baseline. The person’s history is the control; there is no population norm anywhere in the loop.

• Evidence is filed by what the behaviour was for, and by the exact response configuration that was used — never pooled by what the behaviour looked like. That is the difference between a count and a piece of evidence.
• Function tokens come from observations only, held in parallel until the person’s own outcomes resolve them. “Unhypothesised” is a real, recorded value — never guessed away, and never inferred backwards from the response used.
• Responses are ranked by what the record shows settles this person. A tested configuration outranks an untested one; a brand-new plan variant starts unproven and earns its rank only as episodes accumulate. A plan change adds a variant beside the baseline — it never replaces the history the ranking needs.
• Capture types itself: a carer speaks, and the words become a typed structured row — form, parallel function hypotheses, the four-dial response, who did it, before and after — with the carer’s own words preserved beside the tokens. The evidence was going to be spoken anyway; the record is rich without the carer paying for richness.

The indexing and the ranking are the reason the four pillars are computations here rather than paragraphs in a plan document. How they are computed is ours; what they produce belongs to the practitioner — evidence about one person, with its sources attached, arriving as the work happens rather than being assembled the night before it is needed.

What is different, stated carefully

Four premises, each an objective property of the software rather than an adjective:

• Capture fidelity is the floor of everything. A dropdown loses the sentence; a form captures what the form designer imagined. Typed narrative from speech captures what the carer actually saw.
• Teaching already happens — the record just loses it. Staff teach behaviour swaps constantly, informally, brilliantly. The software’s job is to capture where the teaching happened, transfer it to the team as a consistent palette, and re-teach the same swap next shift. Consistency is a palette problem, not a person problem.
• Evidence for a funding panel is computed or it is assembled by hand. A dated dechallenge, outcomes filed by function, a recorded fade and a proposed plan change carrying its sources arrive as the work happens — not the night before, from notes, at overtime rates.
• Change management is structural. When change lands as data and the plan visibly re-orders with the reason attached, adoption stops being a training problem and becomes a by-product of use.

In an August 2026 timestamped sweep of UK care-management platforms, UK PBS-specific tools and US ABA platforms, we found no deployed system that indexes behaviour evidence by function, computes functional equivalence from outcomes, or dates the fade of assistance on real-life response configurations. Voice capture itself is commoditising quickly across the sector; the typed grammar and the function index are the parts we found nowhere else.

What the system never does

• It does not diagnose, and it does not predict. It notices when this person departs from their own baseline and hands a human the evidence.
• It never places an emergency call. Escalation edges are authored by people and activated by people; the system’s job is to arm the human with the record.
• It never silently changes a live plan. Repeated outcomes can become a review proposal with its sources attached; only an authorised person can turn a proposal into operational guidance.
• It never offers a restrictive configuration. The bar is structural — built into what can be served — not a preference setting.
• It never resolves a function hypothesis by preference or protocol. Hypotheses stay multiple until the person’s own evidence resolves them.
• Population evidence shapes what is watched, never what is concluded about one person.
• Every worked example in this article is synthetic. No real person’s care record is quoted, here or anywhere on this site.

One note per pillar, and four beside them

This article is the map. Each pillar has its own note that goes a level further into the mechanism, the evidence and the boundaries — and one note explains where a service actually starts.

• Documents in, evidence out — the first version: bring the records you already have, and talk to the record.
• The fade — the two ladders scored together, who did the doing, and what the hands were doing.
• The easier ask — teaching a replacement and proving equivalence from the record.
• An ABC whose A can be counted — typed context, the body clocks, and what a band is.
• Body first — the override, PINCH ME at capture, and the dated dechallenge.
• The independence journey in Down syndrome — the same mechanisms across a lifespan.
• Fifty years to find the function — the history, and the chapter still open.
• Storage or evidence — five premises for judging any care record, including this one.

Questions, answered the way we answer them