diff --git a/NEXT.md b/NEXT.md
index 091f38c..41f1e41 100644
--- a/NEXT.md
+++ b/NEXT.md
@@ -478,7 +478,39 @@ run one lane at a time; item 4 is disjoint and runs alongside any of them.
    because `check_dotimes` folds the step into the body and a `continue` branching to the header would skip it and
    hang.
 
-8. **More than one error per compile.** Raised by the author's workflow: write everything, compile at the end, squash
+8. **Errors: a structured value with spans and notes, and more than one per compile.** One piece of work, not two —
+   both need `Loc.Error` to stop being a single location plus a string.
+
+   **Today:** `lib/loc.ml` carries a point location and a message, and `Loc.Error` is the frontend's *one* exception, so
+   the first error aborts the run. The author's workflow is write everything, compile at the end, squash the list —
+   which cannot work when there is never a list. The *content* of the messages is already good; they state the reason
+   and name what to write instead. What is missing is structure and volume.
+
+   **jank is the model** (`~/Repositories/jank`, `compiler+runtime/include/cpp/jank/error.hpp`). It is a Lisp on LLVM
+   with unusually good diagnostics and three things worth taking:
+
+   - **A named `kind` per error** — roughly a hundred, `lex_unterminated_string`, `parse_odd_entries_in_map` — each with
+     a stable string id. Machine-readable classification with no JSON mode and no prose parsing.
+   - **A source *span*, not a point.** This is what draws Elm's squiggle: you underline a range. A column number cannot.
+   - **Notes: an error carries zero or more, each with its own span and its own severity** (info/warning/error), sorted
+     by position. **This is the actual secret of Elm-quality messages** — "this is wrong *here*" plus "because of *that*
+     over there", two places highlighted and each explained. One location and one string can never express it.
+
+   jank also carries the **macro expansion** an error came from, which this project will want once macros land, and it
+   is worth building the field now rather than retrofitting it.
+
+   **Then collect rather than raise:** finish the function, finish the file, report everything found. Error recovery in
+   a checker is real work — the hard part is resynchronising after a bad form without cascading nonsense — and it is
+   what the workflow actually needs.
+
+   **No editor work is required.** Flan already prints `file:line:col: message`, the GNU format Emacs's
+   `compilation-mode` parses with no configuration, so `M-x compile` gives a clickable list and `next-error` free.
+   Flycheck and a structured JSON report were both considered and are **not** wanted — the workflow is
+   compile-at-the-end, not live linting.
+
+   **Cannot run beside the current lanes**: it touches every file that raises, which is the whole frontend.
+
+8b. **The old entry, kept for its one extra fact:** Raised by the author's workflow: write everything, compile at the end, squash
    the list. That does not work today — `Loc.Error` is the frontend's **one** exception, so the first error aborts the
    run and you get them one at a time, which is exactly the loop that workflow exists to avoid.