feat(ops): restructure the weekly digest around three production windows

.github/workflows/run-ops.yml

@@ -27,8 +27,8 @@ name: run-ops
 # The reading surfaces are digests a maintainer closes, not a standing body this
 # job edits in place: an unread digest is an open issue, so the backlog is
 # visible without a reading-state store. The Monday schedule opens the **weekly
-# performance digest** as its own `weekly-digest` issue — the health questions,
-# the committed boards' state, the week's cells and measured spend, and the
+# performance digest** as its own `weekly-digest` issue — the cells and measured
+# spend produced over the week, the trailing month and the Term to date, and the
 # back-test results, every metrics figure carrying its artifact's vintage. A
 # second job on the daily tick opens the **daily prediction-reading digest**:
 # one predicted event with every predictor side by side, on its own
@@ -49,10 +49,10 @@ on:
   schedule:
     - cron: "0 8 * * *" # after run-pull (07:17) so the day's runs are visible
     # The weekly tick: the same daily job, plus the weekly performance digest
-    # issue. Two steps are gated on this exact schedule string — the build
-    # step's WEEKLY_TICK env and the posting step's `if` — and a shape test
-    # pins both to a cron this block actually declares. Offset from the daily
-    # 08:00 so the two runs never race on the `ops-metrics` branch.
+    # issue. The posting step's `if` is gated on this exact schedule string,
+    # and a shape test pins it to a cron this block actually declares. Offset
+    # from the daily 08:00 so the two runs never race on the `ops-metrics`
+    # branch.
     - cron: "30 8 * * 1"
   workflow_dispatch:
 
@@ -120,8 +120,7 @@ jobs:
           # An open issue wearing a run:* fan-out label is a marker somebody left
           # behind: nothing keys on those labels, so it is queued work to no one
           # and the report says so rather than letting it read as a round in
-          # flight — the weekly digest carries the same signal as a health
-          # question. One gh call per label, merged; the tested
+          # flight. One gh call per label, merged; the tested
           # `summarize_trigger_issues` does the filtering/ordering.
           for label in run:predict run:evaluate; do
             gh_retry gh issue list --repo "$REPO" --label "$label" --state open --limit 100 \

README.md

@@ -47,7 +47,8 @@ requests.
 Plus `run-ops` (a read-only daily operations report, plus two issues the maintainer
 closes once read: a daily prediction-reading digest — one predicted event with
 every predictor side by side — and a Monday performance digest carrying the
-week's cells, spend, board state, and back-test results) and
+cells and spend produced over the week, the trailing month and the Term to
+date, plus the back-test results) and
 `run-analytics` — seven dispatch modes: corpus statistics, the
 distribution-parse census, the document text-coverage enumeration, the
 tool-usage roll-up, the metrics refresh, the daily big-case board, and the

docs/pipeline.md

@@ -112,25 +112,30 @@ unread backlog.
 
 Four blocks, in the order a reader needs them:
 
-- **Health questions** — the fixed interrogative bullets (replay calibration,
-  forward cells scored, watchlist vs next conference, oldest stalled trigger,
-  spend vs budget): numbers as questions demanding a reaction. These carry the
-  ops report's un-vintaged framing, which is why the vintage rule below is
-  scoped to the two blocks that publish figures to quote.
-- **Analytics state** — what the committed boards hold. An empty one names the
-  condition that empties it — which cells the frozen headline ranks, and how
-  many have reached it — rather than showing a bare zero, and an artifact that
-  has never landed reads differently from one that landed empty. Plus the
-  statpack's two headline rates — stated with each one's own denominator, and
-  with the plain statement that **neither anchors a scored cell**: a forward
-  cert cell is scored against its own band's strictly-prior-Term risk-set rate,
-  and the pooled band rate is a fit diagnostic for the ranking constant rather
-  than a scoring baseline ([salience.md](salience.md)).
 - **Produced this week** — cells landed by role and stage, how many events they
   covered, and the week's measured spend, all over one window and one set of
-  `usage.json` records; then the spend backstop's own (longer) window and how
-  much of its ceiling the trailing period has consumed. An unenforced ceiling
-  says so instead of reporting a fraction of a budget that does not exist.
+  `usage.json` records. The window's bounds are in the heading: the Monday tick
+  titles its issue for the ISO week that *starts* that morning while the census
+  covers the seven days before it, so without them the block would describe the
+  previous week under this week's heading.
+- **Produced this month** — the same shape over the trailing window the ex-post
+  spend backstop is configured with, closing with that backstop's own verdict
+  and how much of its ceiling the period has consumed. The window is the
+  backstop's own precisely so the census and the verdict beside it cannot
+  describe different periods. An unenforced ceiling says so instead of reporting
+  a fraction of a budget that does not exist.
+- **Produced this term** — the same shape again over the October Term to date,
+  from the 1 October that Term opened to the day the digest is generated. The
+  cutoff is pinned to that instant rather than counted back in days, or a
+  mid-morning render would cut the Term's own first morning out of its census.
+  It closes with the forward cells scored under the process in force. That
+  count is cumulative over the whole ledger, not the Term's, with a delta
+  against the prior ops-metrics snapshot (a week when the dated snapshot
+  exists, shorter when the job fell back to the latest one); it sits in the
+  Term block because a forward cell is minted once at its event and never
+  again, so the Term is the period it is worth reading beside. A frozen scope
+  with nothing scored in either stratum is named as the shakedown state rather
+  than shown as a bare zero.
 - **Backtest results** — the historical replay **per court**, with each court's
   own always-deny floor beside its accuracy and the pooled row labelled as the
   mixture it is (`granted` means cert on a SCOTUS row and a motion granted on a
@@ -157,10 +162,12 @@ Four blocks, in the order a reader needs them:
   the number is in, because a caveat one bullet away does not travel when the
   line is quoted. A board with no entries says so and prints no floor.
 
-**In the analytics and back-test blocks, every figure carries the vintage of the
-artifact it came from.** None of those artifacts is refreshed on this schedule —
-a board is byte-stable and a statpack moves only when the corpus does — so a
-figure without its vintage would silently claim to be this week's. The vintage
+**In the back-test block, every figure carries the vintage of the artifact it
+came from.** None of those artifacts is refreshed on this schedule — a board is
+byte-stable, and the cert back-test moves only when a maintainer dispatches one —
+so a figure without its vintage would silently claim to be this week's. The
+production blocks need no vintage: they are computed from the committed ledger at
+render time. The vintage
 is the commit that last wrote the file, and a **shallow** checkout yields none:
 in a depth-1 clone the one grafted commit matches every path, so a pathspec'd
 `git log` would stamp every board with today's date — the exact misreading the

src/fedcourtsai/cli.py

@@ -185,6 +185,7 @@
     DAILY_DIGEST_MARKER_LINES,
     WEEKLY_DIGEST_LABEL,
     WEEKLY_DIGEST_MARKER_LINES,
+    ProducedWindow,
     Vintaged,
     WeeklyAnalytics,
     WeeklyProduction,
@@ -301,15 +302,13 @@
     AgentFlags,
     AgentToolingFeedback,
     Backtest,
-    BigCaseBoard,
     CellFailure,
     CellMode,
     CertBacktest,
     CertBacktestCellLoss,
     CertBacktestDispatch,
     CertBacktestProvenance,
     ClaimScoreBlock,
-    ClaimScoreBoard,
     ConferenceBucket,
     CorpusValidation,
     DataHealth,
@@ -7739,12 +7738,8 @@ def _vintaged[T: BaseModel](path: Path, model: type[T]) -> Vintaged[T]:
 
 
 def _weekly_analytics(metrics_root: Path) -> WeeklyAnalytics:
-    """The committed boards the weekly digest reports, each with its own vintage."""
+    """The committed replay artifacts the weekly digest's back-test block reports."""
     return WeeklyAnalytics(
-        leaderboard=_vintaged(metrics_root / "leaderboard.json", Leaderboard),
-        claim_scores=_vintaged(metrics_root / "claim-scores.json", ClaimScoreBoard),
-        big_cases=_vintaged(metrics_root / "big-cases.json", BigCaseBoard),
-        statpack=_vintaged(metrics_root / "statpack.json", StatPack),
         backtest=_vintaged(metrics_root / "backtest.json", Backtest),
         salience_replay=_vintaged(metrics_root / "salience-replay.json", SalienceReplay),
         # Never produced by the scheduled refresh — a real-engine replay spends
@@ -7780,29 +7775,66 @@ def _parse_when(stamp: str) -> datetime:
     return parsed if parsed.tzinfo is not None else parsed.replace(tzinfo=UTC)
 
 
-#: The window the weekly digest's production census and its spend figure share.
-#: A week, because that is the period the digest covers; the spend *backstop*
-#: keeps its own, longer window, which is why the two are reported separately
-#: rather than one being derived from the other.
+#: The window the digest's first production block covers. A week, because that is
+#: the period the digest is opened for; the month block takes its own window from
+#: the spend backstop's config instead, so the backstop's verdict and the census
+#: printed beside it can never describe different periods.
 _WEEKLY_WINDOW_DAYS = 7
 
 
+def _produced_over(
+    usage: Sequence[ModelUsage],
+    title: str,
+    *,
+    window_days: int,
+    when: datetime,
+    since: datetime | None = None,
+    term: int | None = None,
+) -> ProducedWindow:
+    """One production block's census and spend, over one window of the ledger."""
+    return ProducedWindow(
+        title=title,
+        census=cell_census(usage, window_days=window_days, now=when, since=since),
+        spend_usd=spend_over(usage, window_days=window_days, now=when, since=since)[0],
+        window_start=(since or when - timedelta(days=window_days)).date(),
+        window_end=when.date(),
+        term=term,
+    )
+
+
 def _weekly_production(data_root: Path, config_root: Path, when: datetime) -> WeeklyProduction:
-    """The week's cells and cost, plus the spend backstop's own verdict.
+    """The digest's three production windows, plus the spend backstop's verdict.
 
-    One walk of the ledger for all three figures: the census, the week's spend,
-    and the backstop's own longer window read the same records, so they cannot
-    disagree about what they cover and the growing tree is scanned once.
+    One walk of the ledger for every figure: the three censuses, their three
+    spend totals, and the backstop's own window all read the same records, so
+    they cannot disagree about what they cover and the growing tree is scanned
+    once.
+
+    The Term window is pinned to the instant its Term opened rather than counted
+    back in days — an October Term opens at midnight on 1 October and the digest
+    renders mid-morning, so a day count would cut the Term's own first morning
+    out of its census.
     """
     usage = iter_usage(data_root)
-    census = cell_census(usage, window_days=_WEEKLY_WINDOW_DAYS, now=when)
-    spent, _cells = spend_over(usage, window_days=_WEEKLY_WINDOW_DAYS, now=when)
+    spend_config = load_spend_config(config_root)
+    term = october_term_year(when.date())
+    term_start = datetime(term, 10, 1, tzinfo=UTC)
     return WeeklyProduction(
-        census=census,
-        spend_usd=spent,
-        backstop=verdict_over(usage, load_spend_config(config_root), now=when),
-        window_start=(when - timedelta(days=_WEEKLY_WINDOW_DAYS)).date(),
-        window_end=when.date(),
+        week=_produced_over(
+            usage, "Produced this week", window_days=_WEEKLY_WINDOW_DAYS, when=when
+        ),
+        month=_produced_over(
+            usage, "Produced this month", window_days=spend_config.window_days, when=when
+        ),
+        term=_produced_over(
+            usage,
+            "Produced this term",
+            window_days=(when.date() - term_start.date()).days,
+            when=when,
+            since=term_start,
+            term=term,
+        ),
+        backstop=verdict_over(usage, spend_config, now=when),
     )
 
 
@@ -7888,12 +7920,13 @@ def ops_report(  # noqa: PLR0913 - one option per independent read-only feed
     to stdout (the run-ops job's Actions step summary); ``--json`` writes the
     structured ``OpsReport``.
 
-    ``--digest-out`` renders the **weekly performance digest** — the health
-    questions, the committed boards' state with each empty one saying why it is
-    empty, the week's cells and measured spend, and the back-test results,
-    every metrics-derived figure carrying the vintage of the artifact it came
-    from. ``--digest-post-repo`` additionally opens it as a `weekly-digest`
-    issue, once per ISO week.
+    ``--digest-out`` renders the **weekly performance digest** — the cells and
+    measured spend produced over the week, the trailing month (closing with the
+    spend backstop's verdict) and the Term to date (closing with the forward
+    cells scored under the process in force), then the back-test results, each
+    back-test figure carrying the vintage of the artifact it came from.
+    ``post-weekly-digest`` is the separate command that opens it as a
+    `weekly-digest` issue, once per ISO week.
 
     Unlike the leaderboard/back-test roll-ups it is a point-in-time snapshot, so
     it is surfaced, not committed.