diff --git a/docs/README.md b/docs/README.md
index b78e870..0224bf8 100644
--- a/docs/README.md
+++ b/docs/README.md
@@ -27,6 +27,7 @@
 - [Reservation Engine Plan](./reservation-engine-plan.md)
 - [Reservation Engine Semantics](./reservation-semantics.md)
 - [Reservation Runtime Seam Evaluation](./reservation-runtime-seam-evaluation.md)
+- [Runtime Extraction Roadmap](./runtime-extraction-roadmap.md)
 - [Revoke Safety Slice](./revoke-safety-slice.md)
 - [Operator Runbook](./operator-runbook.md)
 - [KubeVirt Jepsen Report](./kubevirt-jepsen-report.md)
diff --git a/docs/runtime-extraction-roadmap.md b/docs/runtime-extraction-roadmap.md
new file mode 100644
index 0000000..4cef94d
--- /dev/null
+++ b/docs/runtime-extraction-roadmap.md
@@ -0,0 +1,148 @@
+# Runtime Extraction Roadmap
+
+## Purpose
+
+This document defines the path from the current engine family to something that can honestly be
+called a general internal DB-building library.
+
+The current state is:
+
+- `allocdb-core`, `quota-core`, and `reservation-core` all exist on `main`
+- the engine thesis is proven strongly enough
+- a broad shared runtime is still premature
+- `retire_queue` is the first justified micro-extraction candidate
+
+The goal is not to market a framework early. The goal is to extract only the runtime substrate that
+has actually stabilized under multiple engines.
+
+## End State
+
+We should only call this a general internal DB-building library when all of the following are true:
+
+- more than one runtime module is shared cleanly across engines
+- the shared-vs-domain boundary is explicit and stable
+- a new engine or engine slice can be built with materially less copy-paste
+- extraction reduces maintenance cost more than it adds abstraction cost
+
+Until then, the honest description remains:
+
+- multiple deterministic engines
+- emerging shared runtime
+
+## Milestone Shape
+
+### M12: First Internal Runtime Extractions
+
+Goal:
+
+- extract the smallest runtime pieces that are already mechanically shared
+
+Scope:
+
+- `retire_queue`
+- `wal`
+- `wal_file`
+- `snapshot_file` only if the file-level discipline stays separable from snapshot schemas
+
+Non-goals:
+
+- no public framework story
+- no snapshot schema extraction
+- no recovery API extraction
+- no state-machine trait layer
+
+Exit criteria:
+
+- extracted modules are used by all applicable engines
+- behavior is unchanged
+- tests stay green without new abstraction leaks
+
+### M13: Internal Engine Authoring Contract
+
+Goal:
+
+- define the stable boundary between shared runtime and engine-local semantics
+
+Scope:
+
+- one internal runtime contract note
+- explicit ownership of:
+  - bounded collections
+  - durable frame/file helpers
+  - snapshot-file discipline
+  - recovery helper seams, if any
+- explicit non-ownership of:
+  - command schemas
+  - result surfaces
+  - snapshot schemas
+  - state-machine semantics
+
+Exit criteria:
+
+- the contract is clear enough that another engine authoring pass is constrained by it
+
+### M14: Fourth-Engine Or Reduced-Copy Proof
+
+Goal:
+
+- prove that the extracted substrate lowers authoring cost rather than only moving code around
+
+Acceptable proof shapes:
+
+- build a fourth engine against the extracted substrate, or
+- retrofit one substantial new engine slice against the extracted substrate with clearly reduced
+  copy-paste and no correctness regression
+
+Exit criteria:
+
+- one new engine or engine slice uses the extracted substrate directly
+- the reduction in duplicated runtime code is obvious
+- the authoring contract survives contact with real implementation work
+
+## Recommended Issue Shape
+
+### M12
+
+- `M12`: Extract the first internal shared runtime substrate from the three-engine family
+- `M12-T01`: Extract shared `retire_queue`
+- `M12-T02`: Extract shared `wal`
+- `M12-T03`: Extract shared `wal_file`
+- `M12-T04`: Evaluate and, if still clean, extract shared `snapshot_file`
+
+### M13
+
+- `M13`: Define the internal engine authoring boundary after the first extractions
+- `M13-T01`: Write the internal runtime-vs-engine contract
+- `M13-T02`: Reassess whether a fourth-engine proof is still required or whether the extracted
+  substrate already lowered authoring cost enough
+
+### M14
+
+- `M14`: Prove the extracted substrate lowers engine-authoring cost
+- `M14-T01`: Build one new engine or engine slice against the extracted substrate
+- `M14-T02`: Re-evaluate whether the repository can now honestly claim an internal DB-building
+  library
+
+## Execution Rules
+
+- extract smallest-first
+- after each micro-extraction, stop and verify before continuing
+- if one extraction introduces awkward generic plumbing, stop and reassess rather than force the
+  sequence
+- keep domain logic local even if runtime discipline is shared
+
+## Current Recommendation
+
+Do this next:
+
+1. `M12-T01` shared `retire_queue`
+2. `M12-T02` shared `wal`
+3. `M12-T03` shared `wal_file`
+4. only then decide whether `snapshot_file` is still clean enough to extract
+
+Do not do this next:
+
+- public framework branding
+- generic state-machine APIs
+- generic snapshot schemas
+- extracting recovery entry points before the lower layers stabilize
diff --git a/docs/status.md b/docs/status.md
index b2ef7a6..d545312 100644
--- a/docs/status.md
+++ b/docs/status.md
@@ -1,6 +1,6 @@
 # AllocDB Status
 ## Current State
-- Phase: replicated implementation with external Jepsen gate closed, M9 lease-kernel follow-on live-validated, M10 second-engine proof merged, and M11 third-engine proof merged
+- Phase: replicated implementation with external Jepsen gate closed, M9 lease-kernel follow-on live-validated, M10 second-engine proof merged, M11 third-engine proof merged, and M12 runtime-extraction roadmap staged
 - Planning IDs: tasks use `M#-T#`; spikes use `M#-S#`
 - Current milestone status:
   - `M0` semantics freeze: complete enough for core work
@@ -16,6 +16,7 @@
   - `M9` generic lease-kernel follow-on: implementation merged on `main`
   - `M10` second-engine proof: merged on `main`; shared runtime extraction deferred
   - `M11` third-engine proof: merged on `main`; broad shared runtime still deferred, first micro-extraction now justified
+  - `M12` first internal runtime extractions: planned
 - Latest completed implementation chunks:
   - `4156a80` `Bootstrap AllocDB core and docs`
   - `f84a641` `Add WAL file and snapshot recovery primitives`
@@ -212,9 +213,8 @@
   simulation coverage are now all in the mainline implementation
 - PR `#97` merged issue `#96`, extending Jepsen history generation and analysis for bundle
   reserve, revoke/reclaim, and stale-holder lease paths, then closing the loop with live KubeVirt
-  `lease_safety-control` and full `1800s` `lease_safety-crash-restart` evidence on `allocdb-a`,
-  both with `blockers=0`
+  `lease_safety-control` and full `1800s` `lease_safety-crash-restart` evidence on `allocdb-a` with `blockers=0`
 - the next recommended step remains downstream real-cluster e2e work such as `gpu_control_plane`, not more unplanned lease-kernel semantics work; the current deployment slice covers a first in-cluster `StatefulSet` shape, but bootstrap-primary routing, failover/rejoin orchestration, and background maintenance remain operator work, and the current staging unblock path is to publish `skel84/allocdb` from GitHub Actions rather than relying on the local Docker engine
-- PR `#107` merged the `M10` quota-engine proof on `main`: `quota-core` now proves a second deterministic engine in-repo with bounded `CreateBucket` / `Debit`, logical-slot refill, and snapshot/WAL recovery; the `M10-T05` seam evaluation still concludes that shared runtime extraction is premature, with `retire_queue` the closest candidate and the rest still engine-local
-- PRs `#116`, `#117`, and `#118` merged the full `M11` reservation-core chain on `main`: scaffold, deterministic hold lifecycle, logical-slot overdue expiry, and expiry/recovery proof are now in the mainline implementation
-- PR `#118` also closes the third-engine readout: `retire_queue` is now the first justified internal extraction candidate across all three engines, while a broad `dsm-runtime` or public DB-building library is still premature; `wal`, `wal_file`, and `snapshot_file` are the next likely internal seams only after that micro-extraction lands
+- PR `#107` merged the `M10` quota-engine proof on `main`, and PRs `#116`, `#117`, and `#118` merged the full `M11` reservation-core chain on `main`: the repository now has a second and third deterministic engine with bounded command sets, logical-slot refill/expiry, and snapshot/WAL recovery proofs
+- the `M10-T05` and `M11-T05` readouts still defer broad shared-runtime extraction: `retire_queue` is the first justified internal extraction candidate, while `wal`, `wal_file`, and `snapshot_file` remain the next likely seams only after that micro-extraction lands
+- the next roadmap is now explicit in `runtime-extraction-roadmap.md`: start with `retire_queue`, then `wal`, then `wal_file`, and only then decide whether `snapshot_file` is still clean enough to extract before defining the internal authoring contract and asking for a fourth-engine or reduced-copy proof