name: synth-set-fields
description: How `set_*` (multi-select) fields are profiled per option, captured jointly per option pair, read safely on the wide rungs, and generated deterministically with a pre-registered dictionary.
type: guide
kind: design
applies_to: inspect, predict, manifest
covers: [profile create, synth from-profile, set_bernoulli, set_u128, set_u256, conditional]
requires: [capability:synth]
Synth set fields
Set (multi-select) field profiling
set_* fields are “select all that apply” bitmasks over one shared dictionary — bit i ↔ dictionary entry i. Each entry profiles as an independent Bernoulli sub-field for MARGINAL purposes: set.options[i].frequency is P(bit i set | field non-null).
profile create always captures set = {n, options: [{value, count, frequency}]} (no flag). options lists EVERY dictionary entry in bit order — never sorted by frequency like categorical.top — so two profiles of one schema compare position-for-position. n is the non-null row count. SpecFromProfile reconstructs the field as set_bernoulli from options, bounded only by MaxSetEntries().
Joint capture. --conditional extends to set_* pairs by running the per-pair-type machinery ONCE PER OPTION:
| Section | One entry per | Shape |
|---|---|---|
conditional.set_categorical_pairs | (set option, categorical field) | {set, option, categorical, cells, n}; a_value fixed "selected"/"not_selected", b_value the categorical value (own --top-k collapse) |
conditional.set_numeric_pairs | (set option, numeric field) | {set, option, numeric, categories, n}; categories keyed "selected"/"not_selected" |
conditional.set_set_pairs | option pair between two DIFFERENT set fields (never two options of one field) | {set_a, option_a, set_b, option_b, cells, n}, a 2x2 table |
The option axis never needs a top-K collapse (fixed two-value domain). Bounded by FieldType.MaxSetEntries() (8/16/32/64/128/256) and ContingencyCellCap (128). Thin cells reuse thinPairWarning.
Wide rungs
set_u128 / set_u256 are captured member-for-member, and the read is the whole risk. A narrow rung arrives in the decoder’s wide map as a uint64; a wide one arrives as an encoding.SetMask. Every site — the marginal, all three conditional pair kinds, the --fit-models row snapshot and its dummy design columns, the recovery refit’s reader and all four set fidelity sections — resolves it through the shared setMaskFromWide / setMaskFromWideEntry helpers and NEVER type-asserts .(uint64). A failed assertion yields the zero value, which for a set field is a perfectly plausible “selected nothing”: every member captures frequency: 0.0, every conditional cell lands in the not_selected arm, every model row is deleted listwise (the target reads as carries no predictors), every fidelity delta reads exactly 0.0, and nothing errors. setMaskFromWideEntry is the presence-reporting half — ok == false for a nil map, an absent key (the decoder deletes the entry for a null set field) or a non-mask value, i.e. MISSING rather than empty, which is what a model fit’s listwise deletion needs and a marginal does not.
A dictionary wider than its own rung’s mask is reported, never repaired. Members past MaxSetEntries() get no marginal, no pair capture and no fidelity comparison, which is indistinguishable in the document from an option nobody selects — so capture appends set field "x": dictionary carries N option(s) but the set_uK mask addresses K … to Profile.Warnings, grouped as its own ATTENTION kind (set options beyond the mask width) in the terminal summary. Widening is pulse widen’s job; the profiler states the loss and moves on.
Rules are width-agnostic — a rule addresses a member by option NAME (the row holds a map[string]bool), never by bit. But set_expr assigns a WHOLE selection: there is no bit-count, single-member edit or mutual-exclusion primitive, so a cardinality constraint over many members is unstated rather than expressible.
Generation
Each pair resamples ONE option’s bit conditioned on the paired field’s already-drawn value, in the fixed order set-set → set-categorical → set-numeric, after categorical joint structure and before the correlator. Two options on the SAME set field are two DIFFERENT claim targets, never a conflict (synth-conflicts).
A set_* dictionary is pre-populated ONCE from params.options’ declared order at schema-build time, never lazily during generation — the row value is a map[string]bool and Go randomizes map iteration per process, so pre-registration is what keeps bit assignment (hence determinism) independent of it. AugmentFromProfile requires both dictionaries to name the identical option universe in the identical ORDER (PULSE_SYNTH_PROFILE_SCHEMA_MISMATCH). A marginal-only profile still generates correctly — joint structure is additive, never required.
See
op-synth-set-bernoulli·type-set-u8·type-set-u256.synth-fidelity-report— theset_fieldsand set pair sections.synth-profile-capture— the other conditional pair kinds.