DSL_CTE Staged Analysis Capability Reference

Version: v2.0

State: narrow contract available

Implementation scope: Java engine

DSL_CTE is the v2.0 structured stage plan for complex analysis. It does not expose arbitrary SQL CTEs to the LLM; the Java engine validates, compiles, and executes a governed multi-stage plan.

Position

v1.0 JSON Query DSL is suitable for one filtering, grouping, aggregation, and ordering operation. DSL_CTE covers more complex but still governable analysis:

• aggregate, then derive a metric;
• output aliases, then apply result-stage filtering or ordering;
• calculate cumulative contribution or ranking under a signed window contract;
• use limited bridge templates for period comparison, funnel, SLA, or attribution scenarios.

Every stage field, alias, formula, and output must be verifiable.

Signed capabilities

Stage/contract	Capability	v2.0 position
aggregate	Governed filtering, grouping, and metric aggregation	May feed a downstream stage
derive	Signed formula over aliases from the previous stage	No physical-field reference
window_derive	Narrow windows such as cumulative contribution and signed rank()	No arbitrary window function
postSlice	Filter result-stage aliases	Must not be silently pushed into source filtering
orderBy / limit	Order/truncate by output, groupBy, or signed aliases	No invisible fields
bridge template	Month/quarter comparison, CRM funnel, SLA, target-event/target-month	All template parameters are required

Plan boundary

A stage plan should retain source model or previous-stage input, filters/systemSlice/policy context, grouping grain, metrics and aggregate aliases, derived aliases and dependencies, window contract, ordering/frame/postSlice policy, output aliases, and evidence metadata.

Missing time range, grouping grain, alignment key, version, or scenario is a clarification or rejection case.

Fail-closed rules

Reject requests that:

• reference physical tables, physical columns, or unauthorized fields;
• reference an alias absent from the previous stage output;
• disguise result-stage filtering as source filtering;
• use an unsigned stage type;
• use an unsigned window function, arbitrary SQL expression, or free JOIN;
• omit a Memory Grid alignment contract for cross-model analysis.

Not included in v2.0

• Arbitrary SQL CTE.
• Arbitrary cross-model SQL join.
• Free physical-table query.
• Generic period comparison, funnel, or attribution.
• Automatic inference of fiscal calendars, custom calendars, or business phases.

Relationship to other capabilities

• Governed Expression provides the signed formulas used by derive and result-stage calculations.
• Memory Grid provides bounded alignment for cross-model results instead of expanding DSL_CTE into a free Join surface.
• Pivot is an independent contract and should not be mixed into arbitrary DSL_CTE shapes.
• Experience Recipe may capture frequent analysis, but it cannot bypass the DSL_CTE validator.