IntelligenceBriefGKR
What it is¶
IntelligenceBriefGKR compiles the Claim Intelligence Brief: one page per claim composed from the governed section each pipeline tool publishes. It is a composition, not an analysis. It reads the thesis records the tools built from their traces, applies pipeline-level invariants as assertions across them, and renders; it decides nothing of its own . The richness of a section comes from the tool's trace shape, not from the renderer: a tool can contribute a section only once its trace carries cited facts with four-status, every rule evaluated and recorded, indeterminates naming fields, money in heads, and deadlines with anchors .
The compiled brief is built only from tools that have passed their evaluation kit. It is step 9, the last, in the Trace Contract Program lineup for Auto and Property .
What it is not, on the carrier path¶
Until the program brings other tools through their kits, the carrier deliverable is the subrogation thesis brief rendered by the Part C renderer from build_thesis_record, not this compiled brief. Delivering the 13-section full-pipeline brief in place of the thesis brief was ruled the wrong artifact twice and is listed in the repository's standing rules as a mistake already made once . The second instance, produced on the sample carrier payload, showed why: its subrogation section invented a recovery range, left the at-fault driver indeterminate where the thesis confirms it, omitted two theories, and printed enum literals and dict reprs in prose . Findings logged from that run: the adapter recomputed a rental total and set a stated medical reserve from a BI reserve, both previously reverted by ruling; IntelligenceBriefGKR and ObligationEngineGKR print status envelopes into prose; coverage and the detector report error on a carrier payload where they must report not_run with a reason; and the glossary defines Stated as span-verified extraction, which is wrong on the carrier path.
The false affirmatives¶
The false-affirmative class is the only defect category that loses a POC outright; everything else produces a recoverable gap. The bar for the brief is that a page full of honest gaps passes and a confident page with one fabrication does not. Three were found in one readiness audit, two of them certain to fire on every Property claim.
The first was in this tool. recommended_actions closed with "no escalation required, proceed with standard handling" when every upstream section had errored. The synthesis step treated the absence of findings as a finding. The second was a coverage verdict of covered reached with zero exclusions evaluated because an empty catalogue read as no exclusions applying . The third was "no indicators fired" printed after the detector was handed empty content .
A related output-layer defect: a Property subrogation section reporting NO REFERRAL when a vocabulary mismatch meant no theory could ever confirm. Could-not-evaluate and considered-and-rejected must never render the same. And a figure that contradicts its own qualifier is read as the figure: a reserve rendering $0 beside 0 percent completeness tells an adjuster the claim is worth nothing.
When the compiler cannot compose, it says so. SYNTHESIS REFUSED is the correct output when an upstream section that should have run has errored; the defect that was ruled was not the refusal but that coverage and the detector reported error on a payload they had no narrative to run on, where the state is not_run with a reason.
Three states, not two¶
The code kept collapsing a three-way distinction into two states. A section is skipped when it does not apply to the LOB, not_configured when it applies but is unbuilt, and error when it should have run and failed. An empty dict meant all three, so a deliberate Property liability skip rendered as a stage error and permanently blocked action synthesis. Binary status fields hide distinctions that matter downstream, and the collapse is invisible until a consumer acts on it.
| State | Meaning | Rendering |
|---|---|---|
skipped |
Does not apply to this LOB | Not-applicable renderer |
not_configured |
Applies but the governed content is unbuilt | Not-configured renderer |
error |
Should have run and failed | Error renderer, with the named error |
The dependency check is positive: a section is satisfied only when its struct says ran or skipped, never by the absence of an error key. The three separate renderers and the positive check are in code. The obligations section carries "Deadlines will compute once anchor date is available, do not treat as no deadline", the false-affirmative guard written into the output rather than into a review comment.
Composition rules¶
The renderer reads only the thesis record built by build_<tool>_record(trace); nothing reads a trace directly . Every sentence comes from a template record keyed by an outcome key; no engine basis, reason, text or verify string is printed, and no identifier, enum literal, snake_case key, bracketed tag or Python repr appears in prose. Every rendered paragraph and row carries data-trace and data-status. A record and a render committed together carry the same run_id, and renders come from seed runs, never fixtures. On the carrier path not_returned renders as "not supplied" and not_stated as "not stated"; they are different facts and are never rendered with the same words.
Pipeline invariants are assertions the compiler runs across sections. The first of them land at the authority gate step and are extended as each tool joins; which invariants exist today is not yet documented.
Open items¶
The status-envelope printing (E6) is Trace Contract Program step 1 work. The not_run state for coverage and the detector on the carrier path (E7) and the glossary's definition of Stated (E8) are open findings. The compiled brief is not a carrier deliverable until the lineup completes.