diff --git a/CONTEXT.md b/CONTEXT.md
index 9d60223d..33bf900a 100644
--- a/CONTEXT.md
+++ b/CONTEXT.md
@@ -105,7 +105,7 @@ A Scenario constraint that takes a measure off the table for the optimiser. The
 _Avoid_: inclusions, allowed measures (as a stored concept — UI may present "allowed/excluded" toggles, but what's captured is the exclusion set)
 
 **Modelling run**:
-One user-initiated act of triggering modelling: a set of Scenarios crossed with one filter-defined property selection within a Portfolio. An insert-only record of *what was asked* — the Run filters, the Scenarios, who triggered it, when, and the matched-property count shown at preview — so Plans can be traced back to the selection that produced them. Execution progress is a separate concern (the pipeline's task records own it); a Modelling run is the request and its provenance, not the work. A Scenario is the *question*; a Modelling run *asks* it (possibly again, possibly for a different property selection). Re-running an already-modelled property appends a newer Plan; latest wins on read.
+One user-initiated act of triggering modelling: a set of Scenarios crossed with one filter-defined property selection within a Portfolio. A durable record of *what was asked* — the Run filters, the Scenarios, who triggered it, when, and the matched-property count shown at preview — so Plans can be traced back to the selection that produced them, alongside how the work is going. A Scenario is the *question*; a Modelling run *asks* it (possibly again, possibly for a different property selection). Re-running an already-modelled property appends a newer Plan; latest wins on read.
 _Avoid_: job, batch, trigger request, modelling task (the pipeline's execution records)
 
 **Run filter**:
diff --git a/docs/adr/0008-modelling-runs-filters-to-distributor.md b/docs/adr/0008-modelling-runs-filters-to-distributor.md
index 9d72e06f..dab4e11e 100644
--- a/docs/adr/0008-modelling-runs-filters-to-distributor.md
+++ b/docs/adr/0008-modelling-runs-filters-to-distributor.md
@@ -23,14 +23,23 @@ explicit marker when trigger wiring lands. This is that wiring.
 
 ## Decision
 
-- **A `ModellingRun` is recorded, insert-only, by Next.js at trigger time**:
-  portfolio, the selected scenario ids, the Run filters, the matched-property
-  count shown at preview, who triggered it, when. It is the request and its
-  provenance — Plans become traceable to the selection that produced them.
-  It carries no status column (mirrors ADR-0003's derived-state rule).
-- **The distributor receives `{run_id, portfolio_id, scenario_ids, filters}`**
+- **A Modelling run is recorded as an app-created task** — the BulkUpload
+  trigger convention (`triggerAddressMatching`): Next.js inserts the `tasks`
+  row (service `modelling_run`, `source_id` = portfolio, status `waiting`)
+  plus one config `sub_task` whose `inputs` JSON carries
+  `{scenario_ids, filters, previewed_property_count, triggered_by}`, then
+  passes `task_id` to the distributor, which attaches its execution subtasks
+  to the same task. **No new table, no migration.** Run history and the
+  per-scenario in-flight badge read the portfolio's own `modelling_run` tasks
+  (portfolio-scoped, a handful of rows) and parse the app-authored inputs
+  JSON; execution progress ("61 of 104 ready") falls out of the existing
+  subtask-count summary pattern.
+- **The distributor receives `{task_id, portfolio_id, scenario_ids, filters}`**
   — filters, never property-id lists. The payload stays tiny at any portfolio
-  size; the run row is the durable copy of the same config.
+  size; the config subtask is the durable copy of the same request.
+- **Dispatch failure marks the task `failed`** (an improvement on the
+  bulk-upload flow, which leaves failed triggers at `waiting`), so history
+  never shows a ghost run as pending.
 - **The modelling backend is the single authority for resolving filters to
   properties.** Resolution precedence: landlord-override fact → EPC-derived
   value (new-approach EPC graph, lodged over predicted, or legacy property-row
@@ -46,12 +55,11 @@ explicit marker when trigger wiring lands. This is that wiring.
   the static canonical enums plus Unknown. Filters are bounded (cap the
   postcode multi-select) as a UX guard; "all" is an absent filter, never an
   enumeration.
-- **Run status is derived by correlation, not stored**: the distributor stamps
-  `run_id` on the task records the pipeline already writes
-  (`tasks`/`sub_tasks`, service `plan_engine`). Scenario badges gain a third
-  derived state — "Modelling in progress" — and a per-portfolio run history
-  lists each run's filters, counts, initiator and derived status. No live
-  progress bars in v1.
+- **Run status is the task's status** — no correlation machinery needed: the
+  run and the execution share one record by construction. Scenario badges
+  gain a third derived state — "Modelling in progress" — and a per-portfolio
+  run history lists each run's filters, counts, initiator and derived status
+  with subtask-count progress.
 - **Concurrent runs warn, never block**: triggering a scenario with a run in
   flight shows who started it and when, but proceeds. Plans append and
   latest-per-property wins on read, so overlap is safe; a hard lock could leak
@@ -59,6 +67,13 @@ explicit marker when trigger wiring lands. This is that wiring.
 
 ## Alternatives considered
 
+- **A dedicated `modelling_run` table (+ scenarios join table)**: relational
+  scenario linkage and typed columns for who/filters/counts, but a migration,
+  a second record system beside tasks, and correlation machinery (either the
+  backend stamping run ids into tasks, or the app storing returned task ids).
+  The task system already records app-initiated pipeline work with inputs,
+  progress and status — and the badge/history queries are portfolio-scoped
+  over a handful of rows, so the JSON parse costs nothing that matters.
 - **Pass resolved property ids to the distributor**: makes preview-vs-run
   drift impossible by construction, but ~50k ids is a payload the async entry
   point can't take, and it moves the resolution rule into the app — the
@@ -74,9 +89,10 @@ explicit marker when trigger wiring lands. This is that wiring.
 
 ## Consequences
 
-- Two backend asks must land with the distributor: accept + persist `run_id`
-  into task records, and expose the dry-run/count endpoint sharing the
-  resolution code path.
+- Two backend asks must land with the distributor: accept an app-created
+  `task_id` and attach execution subtasks to it (rather than creating its own
+  task, as the current `plan_engine` entrypoint does), and expose the
+  dry-run/count endpoint sharing the resolution code path.
 - The preview's already-modelled callout and matched counts are only as fresh
   as trigger time; a property added between preview and trigger is picked up
   by the run (filters re-resolve) — the run row records the previewed count so