FE-1573: Construct and explain one real net region from a genuine conversation

libs/@hashintel/brunch-agent/AGENTS.md

@@ -139,9 +139,9 @@ These rules exist because Mission 4 lost its design between the owner conversati
 
 ## Authorities vs obligations
 
-[`docs/specs/`](docs/specs), [`docs/adr/`](docs/adr) (see its [README](docs/adr/README.md)), and
-[`docs/evidence/`](docs/evidence) are history and reference: prior design hypotheses and observed
-results. They are not marching orders. Re-earn any design you build to; an implemented decision is
+[`docs/specs/`](docs/specs) (see its [README](docs/specs/README.md)), [`docs/adr/`](docs/adr)
+(see its [README](docs/adr/README.md)), and [`docs/evidence/`](docs/evidence) are history and
+reference: prior design hypotheses and observed results. They are not marching orders. Re-earn any design you build to; an implemented decision is
 evidence, unimplemented design is a hypothesis. A branch may depart from a recorded decision by
 noting the divergence in its commit. Provenance is not warrant: a statement is evidence of what
 was said, not automatically of the terrain. This holds equally for specs, ADRs, the user's

libs/@hashintel/brunch-agent/CONTEXT.md

@@ -113,12 +113,16 @@ The runtime branch in which the workpiece is the complete input and no interview
 **Evidence level**:
 One of three non-collapsible claims about a constructed artifact: tool-schema acceptance, agent-reviewed structural correspondence, and behavioral execution or stronger analysis. Report every level reached; none implies the next.
 
-### Evidence and capture
+### Evidence
 
 **Session**:
-One substrate conversation: the full log of user, agent, tool, and injected entries. Sessions go quiet rather than close.
+One substrate conversation: the full log of user, agent, tool, and injected entries. Sessions go quiet rather than close. Flue history is the canonical conversation log.
 _Avoid_: sitting, conversation (as a distinct concept)
 
+### Historical — rejected capture path (2026-09-04)
+
+These terms describe Mission 2's mechanical sweep and store. They were rejected as product provenance on 2026-09-04: Flue history already carries message ids and exact text, and the store duplicated them under a second identity scheme. Surviving homes are the workpiece revision protocol and, if compaction loses folded records, the existing session-log archive lane. Do not treat the still-exported capture-store code as the durable truth of a document.
+
 **Capture**:
 Mechanically extracted source evidence from a settled range of session entries: an immutable, quote-anchored, domain-opaque envelope. Produced only by a sweep and never written during conversation.
 _Avoid_: extraction, harvest, typed claim

libs/@hashintel/brunch-agent/README.md

@@ -7,8 +7,9 @@ Brunch is the stateful elicitation harness and package family at `libs/@hashinte
   [`MISSION.next.md`](./MISSION.next.md) is the self-contained canonical future spine and is not
   execution authority. Closed missions live under [`docs/mission-archive/`](./docs/mission-archive/).
 - [`CONTEXT.md`](./CONTEXT.md) defines the domain language.
-- [`docs/specs/`](./docs/specs/) and [`docs/adr/`](./docs/adr/) record the harness contract and
-  prior design decisions (see [`docs/adr/README.md`](./docs/adr/README.md)).
+- [`docs/specs/`](./docs/specs/) and [`docs/adr/`](./docs/adr/) are historical design hypotheses,
+  not the current harness contract (see [`docs/specs/README.md`](./docs/specs/README.md) and
+  [`docs/adr/README.md`](./docs/adr/README.md)).
 - [`docs/evidence/`](./docs/evidence/) holds observed results and proofs.
 - [`packages/core/`](./packages/core/) is `@hashintel/brunch-agent`; its `./flue` subpath is the
   production contribution (always-on prompt and the `elicitation` skill), `./storage` and

libs/@hashintel/brunch-agent/docs/adr/0001-brunch-is-the-product-name.md

@@ -3,7 +3,8 @@
 Date: 2026-08-13
 Status: accepted
 Amended: 2026-08-20 by ADR-0004 / FE-1437 (HASH package namespace)
-Supersedes: spec [§12.3](../specs/elicitation-kernel.md#123-naming--tool-namespacing) in part
+Supersedes: historical kernel spec §12.3 in part (last living copy
+`69c02f69a9:libs/@hashintel/brunch-agent/docs/specs/elicitation-kernel.md`)
 Decided on: FE-1388
 
 ## Context

libs/@hashintel/brunch-agent/docs/adr/0002-topology-and-placement-rules.md

@@ -3,8 +3,9 @@
 Date: 2026-08-17
 Status: historical; superseded for current Brunch composition by the final [Mission 4 architecture](../mission-archive/4-core-plugin-elicitation-proof-of-life.md). N3's app composition boundary and the prohibition on app-local plugin content survive, but the three-lane/YAML/repertoire details do not.
 Amended: 2026-08-20 by ADR-0004 / FE-1437 (N3 application placement)
-Refines: spec [§12.2](../specs/elicitation-kernel.md) (package topology) with placement
-rules the spec did not state
+Refines: historical kernel spec §12.2 (last living copy
+`69c02f69a9:libs/@hashintel/brunch-agent/docs/specs/elicitation-kernel.md`) with placement
+rules that spec did not state
 Decided on: FE-1401 (remediation sweep); ratified by Lu, 2026-08-17
 
 ## Context

libs/@hashintel/brunch-agent/docs/adr/0003-three-register-ir.md

@@ -1,9 +1,10 @@
 # ADR-0003: The IR is the elicited model, derived — three registers, not one
 
 Date: 2026-08-18
-Status: accepted
-Amends: [ir-design.md](../specs/intermediate-representation.md) Layer A (the
-"Definition" paragraph), ratified FE-1364/FE-1397
+Status: historical; superseded as product provenance by the 2026-09-04 lineage/basis lock in
+[`MISSION.next.md`](../../MISSION.next.md). The three-register capture/fold IR is rejected.
+Amends: historical IR spec Layer A (last living copy
+`69c02f69a9:libs/@hashintel/brunch-agent/docs/specs/intermediate-representation.md`), ratified FE-1364/FE-1397
 Amended by: [ADR-0005](0005-model-assisted-sdcpn-realization.md) — projections remain pure through
 the scaffold and obligation plan; executable code is realized downstream.
 Decided on: FE-1405 (payload-interiors session); ratified by Lu, 2026-08-18

libs/@hashintel/brunch-agent/docs/adr/0005-model-assisted-sdcpn-realization.md

@@ -1,7 +1,7 @@
 # ADR-0005: Realize executable SDCPNs from deterministic projection scaffolds
 
 Date: 2026-08-24
-Status: accepted
+Status: historical; register-3 projection scaffolds are not current product provenance. Mission 7 constructs through declared basis on browser mutations, not a pure fold over captures.
 Amends: [ADR-0003](0003-three-register-ir.md), register 3
 Extends: [ADR-0004](0004-in-petrinaut-staging-and-the-monorepo-import.md), artifact contract only;
 the application/library topology is unchanged

libs/@hashintel/brunch-agent/docs/adr/0006-plugins-per-target-formalism.md

@@ -21,7 +21,8 @@ line, truck fleet, coating plant — is unknown before the conversation starts.
 cannot be keyed to a domain; the only thing fixed before the first turn is the target formalism the
 model will be projected into.
 
-The IR spec's [Layer B](../specs/intermediate-representation.md#layer-b--the-cps-plugins-ir)
+The historical IR spec's Layer B (last living copy
+`69c02f69a9:libs/@hashintel/brunch-agent/docs/specs/intermediate-representation.md`)
 already defined the CPS plugin at exactly that level: ten kinds, cross-kind `quantity` /
 `source-regime` / `rationale` attributes, and question-relative completion over a static floor.
 The design-convergence queue selected by S-005 then drifted below it. The FE-1402 rehearsal

libs/@hashintel/brunch-agent/docs/adr/0008-repertoire-and-plugin-contract-live-in-core.md

@@ -1,7 +1,7 @@
 # ADR-0008: Repertoire and plugin contract live in core
 
 Date: 2026-08-26
-Status: accepted 2026-08-26 (Lu)
+Status: historical; superseded for current implementation by the final [Mission 4 architecture](../mission-archive/4-core-plugin-elicitation-proof-of-life.md). The YAML repertoire/plugin-contract machinery was removed.
 Amends: [ADR-0007](0007-harness-teaching-meets-plugin-content-at-fixed-keys.md),
 decision 8 (`packages/repertoire` is replaced by a guarded core subpath)
 Preserves: ADR-0007 decisions 1–7 and 9; the repertoire remains harness-owned,

libs/@hashintel/brunch-agent/docs/adr/0009-openai-voice-ui-turn-shell.md

@@ -1,11 +1,14 @@
 # ADR-0009: OpenAI Realtime media plane, Brunch control plane
 
 Date: 2026-08-26
-Status: accepted for the bounded H-6763 preview stack
+Status: historical for the H-6763 preview stack. The Realtime-as-media-plane / Brunch-as-control-plane
+split survives; `brunch_ask`, capture-fold authority, and duplex-shell details do not. Live Voice
+contracts are Missions 5, 6b, and 7.
 Extends: [ADR-0004](0004-in-petrinaut-staging-and-the-monorepo-import.md), which keeps Brunch and
 Petrinaut composition in applications and reusable libraries mutually unaware
-Preserves: [ADR-0003](0003-three-register-ir.md), which makes Brunch's capture fold authoritative,
-and the [Petrinaut integration attach contract](../specs/petrinaut-integration.md#attach-contract)
+Originally preserved: [ADR-0003](0003-three-register-ir.md) and the historical
+[Petrinaut integration attach contract](../specs/petrinaut-integration.md#attach-contract) —
+both later superseded for provenance and structured questions.
 
 ## Context
 

libs/@hashintel/brunch-agent/docs/adr/README.md

@@ -8,4 +8,6 @@ re-earn before building further on them.
 Internal references to retired paths (`docs/control/`, `docs/agents/`, `docs/INDEX.md`) are
 historical and not maintained.
 
-For the current accepted Brunch architecture, start at the live root [`MISSION.md`](../../MISSION.md), [`MISSION.next.md`](../../MISSION.next.md), and the final [Mission 4 archive](../mission-archive/4-core-plugin-elicitation-proof-of-life.md). Mission 4 replaced the generalized YAML/repertoire/plugin machinery described in ADR-0002, ADR-0006, and ADR-0007 with a Flue-native independent core `elicitation` capability, target-pairing plugin job skills, and app-owned composition. Those ADRs remain useful design history, not an integration baseline.
+For the current accepted Brunch architecture, start at the live root [`MISSION.md`](../../MISSION.md), [`MISSION.next.md`](../../MISSION.next.md), and the final [Mission 4 archive](../mission-archive/4-core-plugin-elicitation-proof-of-life.md). Mission 4 replaced the generalized YAML/repertoire/plugin machinery described in ADR-0002, ADR-0006, ADR-0007, and ADR-0008 with a Flue-native independent core `elicitation` capability, target-pairing plugin job skills, and app-owned composition. ADR-0003 and ADR-0005 describe the rejected three-register capture/fold IR; provenance is now recovered lineage plus declared basis. ADR-0009's Realtime-as-media-plane split survives; its `brunch_ask` / capture / duplex-shell details do not — live Voice contracts are Missions 5, 6b, and 7. Those ADRs remain useful design history, not an integration baseline.
+
+Living YAML/IR/completion specs that these ADRs once pointed at were removed on 2026-09-07; last copies are at commit `69c02f69a9`. See [`docs/archive/specs/README.md`](../archive/specs/README.md).

libs/@hashintel/brunch-agent/docs/archive/elicitation-kernel/map.md

@@ -1,8 +1,12 @@
 # Map: Elicitation Kernel — carve-out spec
 
+> Historical 2026-08-10 wayfinder. The assembled kernel spec it points at was removed from
+> `docs/specs/` on 2026-09-07; last living copy
+> `69c02f69a9:libs/@hashintel/brunch-agent/docs/specs/elicitation-kernel.md`.
+
 Label: wayfinder:map
-Status: closed — destination reached 2026-08-10 (the spec is assembled:
-[spec.md](../../specs/elicitation-kernel.md))
+Status: closed — destination reached 2026-08-10 (the spec was assembled, then later removed
+from the living tree)
 Created: 2026-08-06
 
 ## Destination

libs/@hashintel/brunch-agent/docs/archive/specs/README.md

@@ -0,0 +1,19 @@
+# Retired living specs and leftover research
+
+On 2026-09-07, on `ln/fe-1573-construct-and-explain`, the owner authorized deletion of living
+docs that still described the discarded YAML plugin, three-register IR, capture-envelope, and
+pre-Realtime Voice destinations. Surviving rationale already lives in
+[`MISSION.next.md`](../../../MISSION.next.md), the [Mission 4 archive](../../mission-archive/4-core-plugin-elicitation-proof-of-life.md),
+and the 2026-09-04 provenance-by-lineage evidence. Complete last living copies are pinned at
+commit `69c02f69a9`.
+
+Removed from living paths (not copied forward):
+
+- `docs/specs/{plugin-contract,elicitation-completion,elicitation-kernel,intermediate-representation,intermediate-representation-plain,elicitation-to-ir-oracle-design,structurally-typed-elicitation-runbooks}.md`
+- `docs/reference/architecture/capture-store.md`
+- `docs/evidence/design/{plugin-keys-pressure-review-cycle-1,elicitation-completion-rehearsal,elicitation-completion-plain,cps-interview-guidance-plain,cps-interview-guidance-desk-replay,intermediate-representation-worked-examples}.md`
+- `docs/inbox/salvage/**`
+- `docs/research/{voice-feasibility,voice-implementation-recommendation-pplx,amp-analysis-flue-vs-tilde}.md`
+
+Earlier superseded drafts already in this directory remain as 2026-08-25 archive copies. They do
+not restore the deleted living contracts.

libs/@hashintel/brunch-agent/docs/archive/specs/elicitation-completion-2026-08-25-full-draft.md

@@ -3,7 +3,8 @@
 > `where`-scoped `PresenceClause` / `SlotClause` vocabulary, and the `completionAnchor` matching
 > below have no current authority; completion is now specified as the invariants of
 > `evaluateCompletion(model, mustKnowRows)` over the plugin file's `Must know` table in the
-> rewritten [`elicitation-completion.md`](../../specs/elicitation-completion.md). Content is
+> a later living `elicitation-completion.md`, itself removed on 2026-09-07 (last copy
+> `69c02f69a9:libs/@hashintel/brunch-agent/docs/specs/elicitation-completion.md`). Content is
 > otherwise verbatim; only relative link targets were re-rooted for the archive location.
 
 # Spec: target-document completion and session stopping

libs/@hashintel/brunch-agent/docs/archive/specs/plugin-contract-2026-08-25-declarative-draft.md

@@ -3,8 +3,8 @@
 > authored as one sectioned Markdown file). The typed declarative contract below — `ScopeExpr` /
 > `where` / `inSupport`, `ProposalType.affordance.firesWhen`, `NodeKind.completionAnchor`, the
 > typed `foldTable` / `demandTable` / `variantDimension` / `lossCategories` keys — has no current
-> authority; the current contract is the shrunk [`plugin-contract.md`](../../specs/plugin-contract.md)
-> and the exemplar [`plugin-sdcpn/plugin.yaml`](../../../packages/plugin-sdcpn/plugin.yaml). Content is otherwise
+> authority; the later living `plugin-contract.md` and `plugin.yaml` exemplars were themselves
+> removed on 2026-09-07 (last copies at `69c02f69a9`). Content is otherwise
 > verbatim; only relative link targets were re-rooted for the archive location.
 
 # Spec: the plugin contract — two schemas, two tables

libs/@hashintel/brunch-agent/docs/evidence/audits/core-elicitor-prompt-material-audit-2026-08-31.md

@@ -22,7 +22,10 @@ The strongest always-on candidates are: objective-relative attention; expert voc
 
 Repeated text is not independent corroboration when the active skill, repertoire, universal syntheses, and later specifications all descend from the same local source pool. Confidence rises where different evidence classes align: verified literature, observed Brunch runs, independently observed LLM-interviewer failures, and current executable teaching. Historical specifications and prompt variants show design lineage and candidate wording, not effectiveness by themselves.
 
-The corpus itself warns against prompt accretion: [`elicitation-to-ir-oracle-design.md`](../../specs/elicitation-to-ir-oracle-design.md) says not to paste source material wholesale into the system prompt or skill; [`structurally-typed-elicitation-runbooks.md`](../../specs/structurally-typed-elicitation-runbooks.md) says the always-on instruction is a concise router and invariant set, while bulky universal material remains lazy.
+The corpus itself warns against prompt accretion: the historical oracle-design and runbook specs
+(last living copies at `69c02f69a9:libs/@hashintel/brunch-agent/docs/specs/`) said not to paste
+source material wholesale into the system prompt or skill, and that the always-on instruction is
+a concise router and invariant set while bulky universal material remains lazy.
 
 ## Source register
 
@@ -38,8 +41,8 @@ The corpus itself warns against prompt accretion: [`elicitation-to-ir-oracle-des
 | [`frontier-model-elicitor-failure-catalogue.md`](../../research/elicitation/frontier-model-elicitor-failure-catalogue.md) | FM-01–15 with mechanism, detection signature, accountable layer, and prevention status | Separates local observations, published observations, and synthesis | Prevents assigning machinery failures to prompt prose; identifies opening overload, ambiguity bypass, and unlicensed influence as technique-owned or partly technique-owned |
 | [`evaluations/protocols/legacy-baseline/v0-prompt.md`](../../../evaluations/protocols/legacy-baseline/v0-prompt.md) | The first compact seven-move elicitor prompt: objectives first; slice then sweep; probe; ask absences; batch breadth/sequence depth; assumption ledger; end properly | Sealed historical evaluation instrument | Strong wording lineage and one observed intervention, but process-model categories and a full deliverable contract make it too target-specific and too large for core |
 | [`harness-teaching-lineage-audit.md`](harness-teaching-lineage-audit.md) | Fifteen historical formulations of generic interviewer craft and their migration among plugin, harness, mechanism, and prompt layers | Historical audit | Establishes that generic ownership was repeatedly intended but never cleanly delivered; does not select final content |
-| [`structurally-typed-elicitation-runbooks.md`](../../specs/structurally-typed-elicitation-runbooks.md) | Explicit Flue information hierarchy and the universal-repertoire versus target-runbook split | Historical specification, not live authority | Supplies the placement rule: concise always-on router/invariants; lifecycle in skill body; bulky teaching in resources |
-| [`elicitation-to-ir-oracle-design.md`](../../specs/elicitation-to-ir-oracle-design.md) | Eight quality claims, hard-failure gates, mistake taxonomy, and source-to-home method | Evaluation design hypothesis with calibrated artifacts | Converts broad virtues into observable failures; most detection detail belongs in evaluation, not the prompt |
+| Historical runbook spec `structurally-typed-elicitation-runbooks.md` (removed 2026-09-07; last copy `69c02f69a9`) | Explicit Flue information hierarchy and the universal-repertoire versus target-runbook split | Historical specification, not live authority | Supplies the placement rule: concise always-on router/invariants; lifecycle in skill body; bulky teaching in resources |
+| Historical oracle-design spec `elicitation-to-ir-oracle-design.md` (removed 2026-09-07; last copy `69c02f69a9`) | Eight quality claims, hard-failure gates, mistake taxonomy, and source-to-home method | Evaluation design hypothesis with calibrated artifacts | Converts broad virtues into observable failures; most detection detail belongs in evaluation, not the prompt |
 | [`vestera-legacy-baseline/readout.md`](../evaluations/vestera-legacy-baseline/readout.md) and [`vestera-prospective-baseline-v1/campaign-adjudication.md`](../evaluations/vestera-prospective-baseline-v1/campaign-adjudication.md) | Observed failure and success ranges under different prompt/runbook conditions | Local run evidence; small samples | Grounds invention, hardening, stopping, opening-load, acquisition variability, and strong behavior to preserve without treating one run as representative |
 | [`agentic-elicitation-challenges`](../../research/elicitation/agentic-elicitation-challenges-2026-08-06T10-02-41Z.md) and [`criteria`](../../research/elicitation/agentic-elicitation-criteria-2026-08-06T14-11-18Z.md) | The early interactive-compiler framing, semantic conservation, explicit transformation, controlled elicitation, and swappable targets | Imported design conversations | Useful conceptual sieve; not direct prompt copy and not independent research evidence |