batches: packets/ and reviews/ — the archived review surface

.github/workflows/pages_dashboard.yml

@@ -20,6 +20,7 @@ on:
     branches: [main]
     paths:
       - "dashboard.html"
+      - "batches/packets/**"
   workflow_dispatch:
 
 permissions:
@@ -48,6 +49,9 @@ jobs:
         run: |
           mkdir -p _site
           cp dashboard.html _site/index.html
+          # Batch review packets (ruled public 2026-08-31) — served under /packets/
+          mkdir -p _site/packets
+          find batches/packets -maxdepth 1 -name "*.html" -exec cp {} _site/packets/ \;
       - uses: actions/upload-pages-artifact@v3
         with:
           path: _site

batches/AGENTS.md

@@ -21,12 +21,16 @@ record it.
 - usage-window-at-dispatch: <5h %, weekly %, opus %>
 - heart-ack:                  # the reason set acknowledged FOR THIS SHIFT
   - <reason line>
+- expected-effects:           # human-written at dispatch, one line per member with reach
+  - <member>: <what it may legitimately change, e.g. "merges into PyAutoArray and may stale release validation">
 - members:
   - <slug>: <prompt path> — <tier> — <review-minutes> — <outcome>
 - collected: 2026-09-04T08:30Z
 - reviewed-at: 2026-09-04T08:30Z   # when they actually sat down
 - usage-window-at-collect: <5h %, weekly %, opus %>
 - delivered: <n>/<n>
+- packet: batches/packets/<YYYY-MM-DD>-<slot>.html
+- review: batches/reviews/<YYYY-MM-DD>-<slot>.md
 - review-minutes-actual: <n>
 - notes: |
     What actually happened. Anything that stalled, and why.
@@ -49,6 +53,18 @@ standing one, which doctrine voids.
 is a seed from the sizing faculty; this is what the slot really cost. Without it
 the estimate never improves, and the whole batch size rests on it.
 
+**`packet:` is the page the human actually reviewed, archived.** One
+self-contained HTML file per slot under `batches/packets/` (see that folder's
+`AGENTS.md`), written at dispatch with PENDING members and refreshed at
+collect. It is the historical record of what the human was shown; a batch
+whose packet was never archived cannot be audited against what they ruled on.
+
+**`review:` is what the human said, verbatim.** The markdown their packet-page
+submission produced (or the orchestrator transcribed), one file per slot under
+`batches/reviews/`. The orchestrator parses it at close-out — merges, rulings,
+queued follow-ups all trace back to a line in this file, not to a memory of a
+conversation.
+
 **`review-at:` is the shift, and it is the human's to declare.** There is no
 schedule: a slot is whenever they come in, so at dispatch they state when they
 expect to be back and **the shift is dispatch → `review-at:`**. The batch's
@@ -58,3 +74,11 @@ measured rather than promised — the pair is what the human's own estimates are
 calibrated against, exactly as `review-minutes-actual:` calibrates the cost. A
 `review-at:` that passes with nobody there dispatches nothing; the grant just
 expires.
+
+**`expected-effects:` is the only licence for "generated by an earlier member".**
+Leg 4 lets a new YELLOW reason pass as *generated by an earlier member of this
+batch* — but only where it matches an `expected-effects:` line the **human**
+wrote here at dispatch. A run classifying a new reason as its sibling's doing on
+its own authority is an autonomous acknowledgement, which the hard invariants
+forbid (`AUTONOMY.md`, "Leg 4 under a batch launch", 2026-08-31 amendment). No
+line, no match, the run parks.