@@ -0,0 +1,92 @@
+# N3 bounded pilot, gate G0 (scope identity) — pre-registration
+
+Written 2026-10-20, before the G0 filter was written or run. What had been read before this file:
+
+- the D11 triage of the 7 holdout DI001 findings (`evidence/holdout/ownnet/di_sample.tsv` of the direction
+  research, private), which names the cause of each FP;
+- `ownlang/di.py` (`find_captive_dependencies`);
+- the field names of one `services` fact in the OrchardCore holdout facts. No DI001 result was recomputed.
+
+Authority: owner decision `ACCEPT-PHASE0 / GO-N3-BOUNDED-PILOT / NO-PRODUCTION`, given on PR #414. G0 budget:
+0.25 agent-day. G1–G4 start only after a G0 PASS, and get one agent-day in total. The first failed mandatory gate
+stops the pilot (`KILL-N3-PILOT`). The budget is not extended.
+
+## Question
+
+Can product registrations be told apart from test registrations, and can each registration be resolved through its
+own implementation, using only the facts the extractor already emits and the repository's project files? The DI
+graph must not be rebuilt.
+
+N3 needs this because its warning must rest on a proven scoped lifetime of a concrete object. If the registration
+graph cannot say "scoped" without the 7/7 FP of D11, a detached-work rule built on it inherits them.
+
+## Known failure (the target)
+
+7 of 7 product DI001 findings on OrchardCore (holdout, target `5ddc4594`, Own.NET `e7d33e33`) are FP. The D11 triage
+gives two causes:
+
+- **test registrations merged in (6).** `IExtensionManager` (5 findings) and `IMediaTokenService` (1) are
+  `AddSingleton` in product code. The `scoped` lifetime comes from test registrations (`StubExtensionManager`,
+  `MediaTokenTests.cs`). The core keys the graph by service name, and the last registration wins.
+- **implementations merged by service name (1).** `IDeploymentTargetProvider`: the singleton implementation
+  `FileDownloadDeploymentTargetProvider` takes only `IStringLocalizer`. The scoped dependency belongs to another
+  implementation of the same interface. The core starts each walk from `by_name[s.name]`, not from the registration
+  `s` itself.
+
+## Method (fixed now)
+
+A research script outside the product tree reads saved `facts.json` files and the repository sources. Production
+code, OwnIR, CI and goldens are not changed. The script applies four rules:
+
+1. **Test project.** A registration belongs to the project whose `.csproj` is in the nearest ancestor directory of
+   its `file`. The project is a test project when its `.csproj` sets `<IsTestProject>true</IsTestProject>` or
+   references one of `Microsoft.NET.Test.Sdk`, `xunit`, `xunit.v3`, `NUnit`, `MSTest.TestFramework`, `MSTest`,
+   `TUnit` or `BenchmarkDotNet`. Path names are not used for this decision. The path regex of the direction
+   research (`sweep_summary.py`) is reported next to it, as a comparison only.
+2. **Drop test registrations.** Registrations in test projects are removed. A registration with no owning
+   `.csproj` is kept and counted.
+3. **Own implementation at the root.** The walk for a singleton registration starts from that registration's own
+   `deps`, not from the node of the same service name.
+4. **Proven lifetime only.** A dependency name counts as scoped only when every remaining registration of that
+   name is scoped. It is followed as transient only when every remaining registration is transient. Mixed
+   lifetimes make the name unknown, and unknown adds no edge. The number of mixed names is reported.
+
+The DI001 rule itself (singleton → scoped, through transients) does not change.
+
+## Inputs
+
+- **Target.** OrchardCore holdout facts as saved by D11 (Own.NET `e7d33e33`). They are not re-extracted.
+- **Positive controls (must keep firing).**
+  - `frontend/roslyn/samples/DiCaptiveSample.cs`: exactly four DI001 (`EmailSender -> AppDbContext`,
+    `ReportService -> UnitOfWork -> AppDbContext`, `CacheService -> IRepo`, `PrimaryCtorService -> AppDbContext`)
+    and silence on `Metrics -> Clock` and `PublicCtorOnly`, as its CI assertions pin today.
+  - Corpus pair `rrr-pi-rate/StockPriceSheetPrintService#60` (C10): DI001 present on `base`, absent on `fix`, as in
+    the D6 ledger.
+- **New-finding check.** The other 14 holdout repositories and the 13 design-B sweep repositories, all from saved
+  facts. Rule 3 can add findings that the current core does not report, so every finding that the current core
+  does not report is listed and triaged by reading the code.
+
+## Verdict rules
+
+`PASS-G0` needs all of:
+
+- all 7 known OrchardCore FPs are gone, each by a rule named above;
+- both positive controls hold exactly;
+- every new finding is triaged, and none is a FP. An `UNCLEAR` counts as a FP;
+- elapsed work within 0.25 agent-day.
+
+`KILL-N3-PILOT` at G0 if any of the following:
+
+- a known FP survives the four rules and removing it needs another mechanism;
+- a positive control is lost;
+- a new finding is a FP or `UNCLEAR`;
+- the budget runs out before a verdict.
+
+No rule may be added or changed after the first run. A fifth rule invented after the run is a G0 failure, not a
+fix.
+
+## Out of scope for G0 (input to G1)
+
+- Framework registrations (`AddDbContext`, `AddHttpClient`, controllers, Razor components). Their lifetime comes from
+  the framework, not from an `Add*` call in the repo. G1 must say how N3 gets "scoped" for them; G0 does not test it.
+- Open generics, factory lambdas, assembly scanning: unknown, as today.