Schema Instantiation
TL;DR
Schema instantiation is Bartlett’s schema and Minsky’s frame put to work: understanding a new case by fitting it to a known template whose slots get filled from context — and whose unfilled slots quietly take default values. That last clause is the governance problem, because defaults are where unevidenced facts enter a determination: recognizing a file as “a standard disputed-ACH matter” imports a dozen assumptions nobody stated. The graph retrieves the template with its slot list, binds each slot to evidence, and checks provenance slot by slot. The binding gate admits no slot that is neither evidenced with a citation nor marked an unfilled default; the fit gate escalates a case that resists the template rather than trimming it to fit. The schema stays; the imports become countable.
The file that was understood before it was read
An intake analyst opens a dispute and knows within seconds what it is: a standard unauthorized-ACH matter. She fills the fields the file gives her — date, amount, originating company, the customer’s own sentence about the debit. The fields it does not give her fill themselves, not by anyone typing but by the shape of the category: a single debit rather than the sixth of a series, a consumer account rather than one being run as a business, notice inside the window, because inside the window is what standard means.
The same eleven minutes would have produced the same determination for the file two rows down, where the debit is the sixth of a series and the window closed before the customer wrote in — facts the template has no field for, and so no reason to look for. Nothing was invented. Part of what the write-up asserts was read off the evidence and part off the schema, and nothing in it says which is which.
An unguided model is a schema engine with no inventory: it fills every gap with whatever the category makes fluent, and states the fill in the voice of a fact. The schema graph must name the template it used, bind each slot to evidence or mark it unfilled, and let the case argue back.
How the graph works, step by step
Retrieve the candidate schema. The template arrives as an artifact rather than as recognition: the procedure, the claim type, the typology entry, with its slot list and its source. Retrieval also makes rivalry visible — the schemas a case nearly fits are as informative as the one it is filed under. A recalled category carries its defaults invisibly; a retrieved one carries them as a list.
Generate the slot bindings. A dedicated node walks the slot list and binds each slot to a span of the record: this field, that statement, this ledger line. Its output is a binding table rather than prose — one row per slot, the value, and where the value came from. Slots the evidence does not reach stay empty, which is the point: an empty row is cheap to see and expensive to miss.
Verify provenance, slot by slot. Each binding is classified, not merely accepted: evidenced with a citation, derived from another slot by a stated rule, or unfilled default — the value the template supplies when the case is silent. Defaults are not forbidden here; forbidding them would make the strategy useless. They are counted, and the count travels with the determination.
Critique the fit and gate it. The critique node asks what a filled form never asks itself: what does this case contain that the schema has no slot for, and which bindings does the evidence strain? A case that resists is admitted as atypical and routed to a human with the resistance named — the alternative to the quiet edit in which the awkward fact is dropped for want of anywhere to put it.
The gates, operationally
| Gate | What it checks | Fails when |
|---|---|---|
| Schema gate | The template is retrieved with its slot list and its source, and the near-fitting rivals it was chosen over are named | The category is recognized rather than retrieved, so its defaults never become inspectable |
| Binding gate | Every slot is evidenced with a citation or explicitly marked an unfilled default, with the default count carried forward | A value supplied by the template is stated in the same voice as one read off the file — an assumption entering as an observation |
| Fit gate | The residue is recorded — what the case holds that the schema has no slot for — and a resisting case escalates as atypical | The case is trimmed to fit: the fact with no home is dropped, and the file reads as standard because the non-standard part was discarded |
Where it fits — three use cases
1. Intake triage and categorization
Categorizing a complaint is not a filing convenience; it silently decides which rulebook, which clock, and which rights apply. Run as schema instantiation, that decision becomes an artifact — the category, the slots it filled, the slots it defaulted — so a misrouted matter is found by reading the trace rather than by waiting for the customer to escalate.
2. Underwriting and claims templates
The standard case is standard because its defaults are usually true, which is why they survive into the file where they are not. The graph does not stop a template from being used on an atypical exposure or an awkward claim sequence; it makes the inherited assumptions countable, so review can ask whether those particular four hold for this risk.
3. Typology-driven narratives
SAR narratives are often drafted against a typology, and a typology is a schema with unusually strong defaults: what such a scheme typically involves. The binding discipline keeps those expectations out of the observed-facts column, so the narrative asserts what the records show and characterizes the rest as pattern-consistent.
When to reach for it
Reach for schema instantiation wherever categorizing a case is most of the work of deciding it — intake, triage, templated underwriting, form-driven adjudication — and especially where volume means the category is assigned in seconds.
The honest limit is that schemas are what make expert work fast, and this strategy does not ask anyone to stop using them. What it reduces and interrupts is narrower: the silent passage of template defaults into the record as observed fact. The imported assumptions become visible and countable, so the difference between observed and assumed-by-template is on the record and can be challenged.
- Slot and default
- A template’s fields and the values it supplies when the case is silent — how recognizing a case type imports facts nobody stated.
- Per-slot provenance
- Each bound slot carries its origin — evidenced, derived by a stated rule, or unfilled default — and the default count travels with the determination.
- Schema residue
- What the case contains that the template has no slot for; the fit gate’s unit of escalation, and the alternative to trimming a case until it looks standard.