diff --git a/.github/workflows/lint_sqlfluff.yml b/.github/workflows/lint_sqlfluff.yml
index 0986785b..d6e1c3f0 100644
--- a/.github/workflows/lint_sqlfluff.yml
+++ b/.github/workflows/lint_sqlfluff.yml
@@ -1,11 +1,17 @@
 name: SQLFluff
 
+# Lints changed mimic-iv/concepts/*.sql on pull requests and posts GitHub
+# annotations. Fork PRs skip Annotate — GITHUB_TOKEN cannot create check runs
+# from forks (Resource not accessible by integration / 403).
 on:
   - pull_request
 
 jobs:
   lint-mimic-iv:
     runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      checks: write
     steps:
       - name: checkout
         uses: actions/checkout@v7
@@ -35,8 +41,14 @@ jobs:
         shell: bash
         run: sqlfluff lint --format github-annotation --annotation-level failure --nofail ${{ steps.get_files_to_lint.outputs.lintees }} > annotations.json
       - name: Annotate
+        # Same-repo PRs only: fork tokens get 403 creating check runs, and
+        # ignore-unauthorized-error does not treat that as unauthorized.
+        if: >
+          steps.get_files_to_lint.outputs.lintees != '' &&
+          github.event.pull_request.head.repo.full_name == github.repository
         uses: yuzutech/annotations-action@v0.6.0
         with:
           repo-token: "${{ secrets.GITHUB_TOKEN }}"
           title: "SQLFluff Lint"
-          input: "./annotations.json"
\ No newline at end of file
+          input: "./annotations.json"
+          ignore-unauthorized-error: true
diff --git a/README.md b/README.md
index 53fe9f9f..756f7726 100644
--- a/README.md
+++ b/README.md
@@ -114,3 +114,7 @@ By committing your code to the [MIMIC Code Repository](https://github.com/mit-lc
 ### Coding style
 
 Please refer to the [style guide](https://github.com/MIT-LCP/mimic-code/blob/main/styleguide.md) for guidelines on formatting your code for the repository.
+
+### MIMIC-IV timing notes
+
+- [Admission deathtime vs transfers](mimic-iv/docs/DEATHTIME.md)
diff --git a/mimic-iv/README.md b/mimic-iv/README.md
index 919b196c..adda4a6d 100644
--- a/mimic-iv/README.md
+++ b/mimic-iv/README.md
@@ -105,4 +105,8 @@ mimic_utils convert_folder mimic-iv/concepts mimic-iv/concepts_duckdb --destinat
 
 # To PostgreSQL:
 mimic_utils convert_folder mimic-iv/concepts mimic-iv/concepts_postgres --destination_dialect postgres
-```
\ No newline at end of file
+```
+
+## Mortality timestamps
+
+`admissions.deathtime` can precede later `transfers` rows — see [docs/DEATHTIME.md](docs/DEATHTIME.md).
diff --git a/mimic-iv/concepts/documentation/README.md b/mimic-iv/concepts/documentation/README.md
new file mode 100644
index 00000000..15576507
--- /dev/null
+++ b/mimic-iv/concepts/documentation/README.md
@@ -0,0 +1,4 @@
+# Documentation helper queries
+
+Ad-hoc SQL that illustrates known data quirks. These are **not** part of the
+official derived concept build — run them against a local or BigQuery copy.
diff --git a/mimic-iv/concepts/documentation/icu_outtime_after_deathtime.sql b/mimic-iv/concepts/documentation/icu_outtime_after_deathtime.sql
new file mode 100644
index 00000000..4fadf65a
--- /dev/null
+++ b/mimic-iv/concepts/documentation/icu_outtime_after_deathtime.sql
@@ -0,0 +1,13 @@
+-- ICU stays whose outtime is after the admission deathtime.
+
+SELECT
+  COUNT(*) AS icustays_out_after_death,
+  COUNT(*) FILTER (
+    WHERE ie.outtime > a.deathtime
+      AND ie.outtime <= a.deathtime + INTERVAL '48 hours'
+  ) AS within_48h
+FROM icu.icustays AS ie
+INNER JOIN hosp.admissions AS a
+  ON ie.hadm_id = a.hadm_id
+WHERE a.deathtime IS NOT NULL
+  AND ie.outtime > a.deathtime;
diff --git a/mimic-iv/concepts/documentation/transfers_after_deathtime.sql b/mimic-iv/concepts/documentation/transfers_after_deathtime.sql
new file mode 100644
index 00000000..26482544
--- /dev/null
+++ b/mimic-iv/concepts/documentation/transfers_after_deathtime.sql
@@ -0,0 +1,11 @@
+-- Count hospital transfers whose intime falls after admissions.deathtime.
+-- Illustrates the #1945 ADT-lag pattern (not a stay_id join failure).
+
+SELECT
+  COUNT(*) AS transfer_rows_after_deathtime,
+  COUNT(DISTINCT t.hadm_id) AS admissions_affected
+FROM hosp.transfers AS t
+INNER JOIN hosp.admissions AS a
+  ON t.hadm_id = a.hadm_id
+WHERE a.deathtime IS NOT NULL
+  AND t.intime > a.deathtime;
diff --git a/mimic-iv/docs/DEATHTIME.md b/mimic-iv/docs/DEATHTIME.md
new file mode 100644
index 00000000..751e6717
--- /dev/null
+++ b/mimic-iv/docs/DEATHTIME.md
@@ -0,0 +1,61 @@
+# Interpreting `admissions.deathtime`
+
+Hospital admissions can show a `deathtime` while `transfers` (and ICU
+`icustays`) still list care-unit activity **after** that timestamp. That is a
+known pattern in MIMIC-IV, not usually a join bug.
+
+See [issue #1945](https://github.com/MIT-LCP/mimic-code/issues/1945).
+
+## What `deathtime` is
+
+| Column | Table | Meaning |
+|--------|-------|---------|
+| `deathtime` | `hosp.admissions` | Time of death recorded for **that hospital admission**, when death occurred in-hospital |
+| `hospital_expire_flag` | `hosp.admissions` | 1 if the patient died during the admission |
+| `dod` | `hosp.patients` | Date of death at the **patient** level (coarser; may come from hospital or Social Security sources depending on version/docs) |
+
+`deathtime` is an admission-scoped clinical/administrative death time. It is
+**not** forced to equal the last `transfers.outtime` or `icustays.outtime`.
+
+## Why transfers can continue after `deathtime`
+
+Common explanations (often several at once):
+
+1. **ADT lag.** Transfer / location rows are written when the bed-management
+   system updates, which can trail the documented time of death.
+2. **Post-mortem logistics.** Patients may still appear to "move" (e.g. ICU →
+   PACU / morgue-related locations) in the transfer chain after death is
+   recorded.
+3. **Clock disagreement.** Death documentation and ADT systems are not always
+   on one synchronized clock in the released tables.
+4. **Stay construction.** Derived ICU stays use transfer/careunit rules; they
+   are not clipped to `deathtime` in the PhysioNet loaders.
+
+So for `hadm_id` examples like the one in #1945, prefer treating `deathtime`
+as the mortality timestamp for that admission, and treat later transfer rows as
+ADT/administrative trail — unless your study specifically needs location-based
+definitions of "end of care."
+
+## Practical guidance
+
+- For **in-hospital mortality**, use `hospital_expire_flag` / `deathtime` on
+  `admissions` (and `patients.dod` when you need patient-level death date).
+- Do **not** require `deathtime >= max(transfers.outtime)` as a data-quality
+  filter without documenting that you are excluding real deaths with late ADT.
+- If you need "alive in ICU until outtime," define that rule explicitly (e.g.
+  ICU outtime, or `least(outtime, deathtime)`), and report sensitivity to the
+  choice.
+- This repository will **not** rewrite transfer out-times from `deathtime` in
+  the CSV loaders.
+
+## Related notes
+
+- ICU charting after outtime: `mimic-iv/docs/CHARTEVENTS_TIMING.md`
+- ED intime vs pyxis: `mimic-iv-ed/docs/TIMING_NOTES.md`
+
+## What this repository will not change
+
+MIMIC-Code does **not** clip `transfers` or `icustays` to `admissions.deathtime`
+in the PhysioNet loaders. Aligning ADT trails with mortality timestamps belongs
+in analysis code or an explicitly documented derived concept.
+