diff --git a/core/test.mjs b/core/test.mjs
index e2163af..2d2bf7f 100644
--- a/core/test.mjs
+++ b/core/test.mjs
@@ -1,4 +1,4 @@
-import { readFileSync } from 'fs';
+import { readFileSync, readdirSync } from 'fs';
 import { fileURLToPath } from 'url';
 import { dirname, join } from 'path';
 import { Rng } from './rng.mjs';
@@ -230,5 +230,19 @@ import { renderWorksheet, renderPlay } from './render-authoring.mjs';
     `play ${done.split('\n').length} vs worksheet ${renderWorksheet(d).split('\n').length} lines`);
 }
 
+// A delve must carry the pack it came from. The Foundry adapter reads `params.theme` to decide
+// which geometry to stage; when it trusted a boot-time global instead, every non-barrow delve got
+// a barrow map. The adapter cannot be tested here, so the contract it depends on is tested here.
+{
+  const ids = readdirSync(dir).filter(f => f.endsWith('.json') && f !== 'index.json')
+    .map(f => f.replace('.json', ''));
+  const wrong = ids.filter(id => {
+    const p = JSON.parse(readFileSync(join(dir, `${id}.json`), 'utf8'));
+    return generateDelve({ pack: p, seed: `theme-${id}` }).params.theme !== id;
+  });
+  t('every delve records the pack it was generated from', wrong.length === 0,
+    wrong.length ? `wrong: ${wrong.join(', ')}` : `${ids.length} packs`);
+}
+
 console.log(`\n${pass} passed, ${fail} failed`);
 process.exit(fail ? 1 : 0);
diff --git a/docs/drafts/APP-AND-RULES.md b/docs/drafts/APP-AND-RULES.md
new file mode 100644
index 0000000..8d531da
--- /dev/null
+++ b/docs/drafts/APP-AND-RULES.md
@@ -0,0 +1,356 @@
+# DELVE — the App and the Rules
+
+*Draft 1, 2026-08-05. Written against the code at `316716e`, module v0.6.1, generator 0.2.0.*
+*Every number here was read out of the source, not remembered. Line references are live.*
+
+---
+
+## 0. What this document is
+
+`DESIGN.md` is a design history — it records how DELVE was argued into existence, including the
+parts that failed. It is the wrong thing to hand someone who asks "what is this and how do I run
+it?"
+
+This is that document. Two halves:
+
+- **The App** — what the software is, the two surfaces, and the workflow at the desk and the table.
+- **The Rules** — the in-play procedure a GM actually runs, stated as rules rather than as
+  architecture.
+
+It is a draft. §7 lists what I believe is wrong with it.
+
+---
+
+## 1. Design intent
+
+1. **DELVE is an authoring system, not an adventure generator.** It gets a GM to a strong first
+   draft in a minute instead of an evening. It does not hand over something to run cold.
+2. **It generates the fiction, the order and the pressure.** It does not generate read-aloud prose
+   or the climax, and it says so on the page where they are missing.
+3. **The fiction coheres because it is projected from one root draw**, not assembled from four
+   independent tables.
+4. **Generation and table dice stay apart.** The core decides *when* a roll is due and what it
+   means; the table rolls it. This is what makes the seed promise honest.
+5. **It is deliberately not a balance engine.** It plans pressure; it does not guarantee a fair
+   fight.
+
+---
+
+# PART ONE — THE APP
+
+## 2. The two surfaces and one working file
+
+Four independent reviews of generated output converged on the same verdict: *"the handwritten
+delve still tells the GM what matters faster."* The density was the problem — and it turned out
+not to be a bug to render away. **Everything on the page is a vice at the table and a virtue at
+the desk.**
+
+So there are two surfaces over one file:
+
+| | For | Carries |
+|---|---|---|
+| **Worksheet** | the desk | everything, labelled, reroll commands beside each component, explicit prompts where only a human can write |
+| **Play sheet** | the table | the GM's own prose, the live leverage point, the decision on one line, the numbers. Nothing else |
+
+**The JSON file is the document.** It carries an `authored` layer the generator never touches
+([`authoring.mjs:22`](../../core/authoring.mjs)), so **rerolling a component cannot destroy
+prose**. Lock what is good, reroll what is not, write the read-aloud, export the play sheet.
+
+- Authored fields: `readAloud`, `notes`, `nameOverride`
+- Rerollable components: `situation`, `decision`, `temptation`, `feature`, `name`
+- A reroll draws from a salted seed and avoids what the rest of the delve already uses; when the
+  pool is exhausted it duplicates another area rather than handing back the same thing, because
+  **a reroll that changes nothing looks broken**
+
+## 3. At the desk — the CLI
+
+`core/` is pure JavaScript with no Foundry dependencies, so the whole authoring loop runs in Node.
+
+```bash
+node core/cli.mjs --new --seed=my-delve --theme=port --file=my.json
+node core/cli.mjs --file=my.json                  # the worksheet (default)
+node core/cli.mjs --file=my.json --reroll=3:situation
+node core/cli.mjs --file=my.json --lock=3:decision
+node core/cli.mjs --file=my.json --write=3:readAloud --text="They have been waiting."
+node core/cli.mjs --file=my.json --todo            # what still needs a human
+node core/cli.mjs --file=my.json --play            # the table-facing sheet
+node core/cli.mjs --themes                         # list the 16 packs
+```
+
+**A fresh delve is not ready to play, and says so.** `outstanding()` returns the intro, one
+read-aloud per area, the ending read-aloud, and — when the ending is `authored` — the climax
+itself. A six-area delve starts with **9 outstanding items**. The play sheet prints a warning
+banner until they are gone.
+
+## 4. At the table — the Foundry module
+
+Three tools in the scene controls:
+
+| | Tool | Does |
+|---|---|---|
+| ⛏ | `delve-forge` | raise a dungeon — opens the panel, generates or loads, stages it in the world |
+| 🗑 | `delve-remove` | remove one — deletes exactly what it made and nothing else |
+| ⏩ | `delve-next` | stage the next area of a delve you finished at the desk |
+
+**Entering an area** ([`vanity-delve.mjs:83`](../../foundry-module/module/vanity-delve.mjs)) does
+five things in a fixed order:
+
+1. Builds the scene — `forge.stage({ populate: false, activate: true })`
+2. Forges the encounter, if the area plans one — with `hoard: false` so loot is decoupled
+3. Forges the hoard, if the area plans one — at its own tier
+4. **Posts the read-aloud to everyone** — the GM's prose, or the cue fragments marked *(unwritten)*
+5. **Posts the GM card, quietly** — situation, decision with resolution, foe block, temptation,
+   and a code line carrying the trigger, the Bane beat and the fallback route
+
+Players first, then the GM. That ordering is deliberate: the scene is up and the read-aloud is
+what the table came for.
+
+**The seams.** DELVE requires VANITY **0.10.4+**. Below that floor the Forge's suppression options
+do not exist and every delve emits 9 stray chat cards and 3 stray folders. The module detects this
+at boot and warns. Requires Foundry v13+, verified on 14.365.
+
+## 5. What the seed promises
+
+State this plainly in any UI. **A partial seed sold as a full one is worse than no seed.**
+
+| Reproducible from the seed | Recorded only | Live at the table |
+|---|---|---|
+| Fiction, area order, heat, hoard tier, foreshadow chain, situations, decisions, temptations | Encounter composition, mood, hoard contents | The clock, the tab, player choices |
+
+The gap exists because VANITY's Forge uses `Math.random` internally (36 sites). Until an injectable
+RNG lands, a seed replays the *delve* but not the *dungeon's population*.
+
+`GENERATOR_VERSION` (currently **0.2.0**) is stamped into every file. It tracks the shape of the
+file and what a seed yields — **not** the module version.
+
+---
+
+# PART TWO — THE RULES
+
+*This is the procedure. A GM who has read this section can run a delve.*
+
+## 6. The kernel — where a delve comes from
+
+One root draw, everything else projected:
+
+```
+APPETITE  ──> motif ──> danger, attention trigger, every cue fragment
+    │     ──> prize kind
+    │     ──> who was conscripted
+    └─ + ACCOMMODATION ──> the transgression, the bottom problem, the primary faction
+```
+
+An **appetite** is a want — *to be looked at*, *never to age*, *to be first*. An
+**accommodation** is the monstrous thing someone did to satisfy it — *had the court interred
+alive*, *had the faces struck off*. Accommodations declare which appetites they fit, because
+drawing freely produced legible nonsense.
+
+8 appetites × 6 accommodations × 7 claimants = **336 kernels from 21 authored lines**. Measured:
+60 generated delves produced **51 distinct kernels**.
+
+**The GM needs one line of this at the table:** the transgression, the bottom problem, and what
+the thing at the bottom is now unable to bear.
+
+## 7. The shape of a delve
+
+Areas are assigned **arc roles** by position, so a delve has a shape rather than a flat run of
+rooms. Proportional — works at 3 areas or 12 ([`director.mjs:37`](../../core/director.mjs)):
+
+| Position | Role | Foreshadow facet |
+|---|---|---|
+| first 20% | approach | institution |
+| to 45% | complication | ritual |
+| to 65% | turn | demand |
+| to 85% | descent | wound |
+| last 15% | threshold | anchor |
+| — | ending (separate slot) | demand |
+
+The facet column is the **foreshadow chain**, and it is why area 2's held bows and area 4's
+scratched-out eyes point at the same faceless queen — they are facets of one motif, dealt in a
+deliberate order. The progression was proposed independently and turned out to describe the
+hand-authored paper delve exactly.
+
+**Approach areas stay quiet.** An empty area is what makes an occupied one land.
+
+## 8. An area, as the GM runs it
+
+Every area carries the same parts, in this order on the card:
+
+**1 · The situation.** Someone doing something, and what changes when the players walk in.
+> *"The tide watch, turning a glass that has already run through — it asks what the hour is, and
+> means the tide."*
+
+It comes with an **offer** — what the players can get out of it — and a **because**, the reason it
+behaves that way. Situations are drawn from the motif, never a global pool.
+
+**2 · The decision.** One thing you can do about the situation. Always states three outcomes:
+
+`[Wits 2]` success · **miss** what failure costs · **or** the alternative that needs no roll
+
+Four test invariants police this: every decision states success **and** failure, offers an attempt
+rather than a pick, and **never gates progress without an alternative**. Playtest 1 died on a gate
+with no fallback.
+
+**3 · The clue.** Automatic on entry — the cue fragment, plus *"and it was done deliberately"*.
+The roll buys extra: `[VANITY: Observation — 1 success]` buys **who** did it and **roughly when**.
+Clues are never gated behind a roll; playtest 1 lost both its clues to single unopposed rolls and
+the motif never reached the table.
+
+**4 · The encounter**, where one is planned. Foe counts and atk/def/Grit/Nerve in a table, plus
+**"harmed by"** stated *before* initiative and **"avoidable"** spelled out.
+
+**5 · The temptation**, where there is a hoard. The decision lives at **use**, not acceptance —
+both playtests took every free-acceptance item without discussion, so free acceptance is not a
+temptation, it is an automatic yes.
+
+```
+benefit          what it does
+useCost          usually +2 Bane
+standingDrawback what it costs you for the rest of the delve
+```
+
+**6 · The failure case.** Every area states it. Ordinary areas: *"if they skip this, move its cue
+fragment into the next area so the motif still completes."* The threshold: *"if they skip this,
+the ending opens with them holding nothing the bottom problem wants."*
+
+**Nothing repeats inside one delve.** Situations, decisions, temptations and features are dealt
+without replacement. Picking independently per area repeated a temptation in **75 of 80** test
+delves.
+
+## 9. The appeasement move — the way through that is not violence
+
+The best mechanic in the hand-authored delve turned out to project *exactly* from the appetite.
+Every appetite implies how it is fed:
+
+| Appetite | The move | Roll |
+|---|---|---|
+| to be looked at | look at it and say plainly what you see | `[Flair 2]` |
+| to be first | stand aside and let it go ahead of you | `[Poise 2]` |
+| never to age | tell it that it has not changed | `[Flair 2]` |
+| never to be alone | sit down and stay a while | `[Poise 2]` |
+
+**+1 Vanity now, +1 Bane later. Repeatable, in every area.**
+
+> **Say the Bane aloud when it banks**, so the table can see the route is open.
+
+This is the delve's social spine. It is always thematically exact, and a party that appeases even
+once reaches an ending that will *trade* rather than fight.
+
+## 10. The clock — Wandering the Dark
+
+**Triggers lead; the timer is a floor.** Playtest 1's clock fired 0 times in 4 rolls on a 3-Turn
+timer, while the attention trigger fired on the first roll of pass 2.
+
+Every area emits at least one **attention trigger**:
+
+| Trigger | Means |
+|---|---|
+| `announced` | entered somewhere grandly, announcing yourself |
+| `seenTwice` | reflected, echoed or witnessed twice over |
+| `loud` | a raised voice, a broken door, a shout |
+| `lingering` | stayed too long in one place |
+| `disturbed` | struck or moved something that had not moved in years |
+
+**Procedure.** Roll `1d6` on any attention trigger, or when the timer comes round.
+**1 — something comes.** Telegraph one Turn ahead — the GM voices a cue before it is due. When it
+fires, roll **Reaction (2d6)** for its mood.
+
+| Pressure setting | Timer period |
+|---|---|
+| slow | every 4 Turns |
+| standard | every 3 Turns |
+| hunted | every 2 Turns |
+
+**Escalation.** If the clock has not fired by the halfway area, the trigger number widens from 1
+to 2 for the rest of the delve. A clock that never bites is decoration.
+
+> ⚠ **The escalation and the period do not currently reach the table.** See §7 of the features
+> draft — this section describes the rule as designed and as implemented in `pressure.mjs`, not as
+> the Foundry module currently behaves.
+
+## 11. The tab and the Reckoning
+
+**Banes are emitted by named beats, never hoped for.** Playtest 1 banked 1 Bane in ~40 rolls
+because Stumbles are rare. The director offers a Bane beat when the tab falls behind pace — always
+when behind, and only half the time when on pace, so it does not feel mechanical.
+
+| Source | Banks a Bane when you… |
+|---|---|
+| `flattery` | played to something's vanity to be received |
+| `noticing` | caught your own reflection and looked |
+| `relicUse` | spent a relic's favour |
+| `announcing` | announced yourself rather than slipping in |
+| `refusal` | refused an offer that would have cost you nothing yet |
+| `push` | pushed a roll |
+| `stumble` | stumbled — a failure with two or more ones |
+
+**Target: ~5 Banes by the ending. THE RECKONING falls at 6.**
+
+> ⚠ **The target is UNCALIBRATED.** Playtest 2 reached 3. It is a target, not a promise.
+
+Quiet areas offer only `noticing` and `announcing`; louder areas add `flattery`, `relicUse` and
+`refusal`.
+
+## 12. Endings
+
+The ending is a **separate slot** after the last area, and by default it is **yours**.
+
+- **`authored`** (default) — DELVE writes the question the place asks and leaves the climax blank.
+  It says so on the card: *"the climax is unwritten. DELVE leaves this to you on purpose —
+  improvise, or stop and write it."*
+- **`generated`** — the ending gets a real heat (one step above the delve's base) and a vault-tier
+  hoard.
+
+**Appeased even once → it listens, and will trade the prize.** That is the ending the social spine
+buys.
+
+## 13. Parameters
+
+| Parameter | Range | Default | Confidence |
+|---|---|---|---|
+| Areas (excludes the ending slot) | 3–12 | 6 | working default — one theme, one table, two passes |
+| Theme | 16 packs, 5 geometries | barrow | all validate; **only barrow is playtested** |
+| Depth | 1–5 | 2 | |
+| Party | 1–8 | 4 | ≤2 softens the heat, ≥7 hardens it |
+| Deadliness | forgiving / standard / cruel | standard | **severity only** |
+| Density | sparse / standard / infested | standard | **frequency only** — 0.20 / 0.34 / 0.55 of areas |
+| Greed | lean / standard / glutted | standard | shifts the hoard tier one step |
+| Pressure | slow / standard / hunted | standard | ⚠ **currently inert — see features draft** |
+| Ending | authored / generated | authored | a default, not a law |
+
+**Heat** runs `skirmish → fight → battle → nightmare`; **hoards** run
+`pocket → cache → chest → vault → kingly`. Descent and threshold areas run one step hotter. The
+area before the ending always carries something — the threshold should have teeth. Treasure is not
+only behind fights: unless greed is `lean`, one treasureless area gets a bonus hoard.
+
+## 14. What DELVE will not do
+
+- **No physically coherent tiled dungeon.** v1 is a pointcrawl.
+- **No balance guarantee.**
+- **No sheet writes.** DELVE offers; the players and the GM decide.
+- **No system-generated climax by default.**
+- **No full replay from seed** until the Forge's RNG seam lands.
+- **No generated read-aloud prose.** Cue fragments only, labelled as fragments. This is the honest
+  ceiling, and faking prose would be the one dishonest thing this tool could do.
+
+---
+
+## 15. What I think is wrong with this draft
+
+Flagging these rather than hiding them, since this is going out for review.
+
+1. **§10 and §13 document a rule the software does not run.** The pressure setting is inert end to
+   end and the escalation never reaches the table. I have written the rule as designed and flagged
+   it twice, but a rules document that describes unimplemented behaviour is a liability. The
+   alternative — documenting the degenerate behaviour — is worse. **The right fix is in the code,
+   not here.**
+2. **Part Two is written for a GM; Part One is written for a developer.** They may want to be two
+   files. I kept them together because the seed promise and the two surfaces are rules-relevant.
+3. **No worked example.** A rules document of this kind usually ends with one complete annotated
+   area. I have not included one because the samples in `samples/` are stamped generator 0.1.0 and
+   one lacks the authored layer.
+4. **The Reckoning is underspecified here.** I state that it falls at 6 because that is what
+   `reckoningDue()` returns, but *what happens* at a Reckoning is VANITY's rule, not DELVE's, and I
+   have not cross-checked the rulebook. Someone should.
+5. **Nothing here has been tested at a table.** Every claim is read from source or from two paper
+   playtests of a hand-written delve.
diff --git a/docs/drafts/NEW-FEATURES.md b/docs/drafts/NEW-FEATURES.md
new file mode 100644
index 0000000..3523d15
--- /dev/null
+++ b/docs/drafts/NEW-FEATURES.md
@@ -0,0 +1,237 @@
+# DELVE — proposed features
+
+*Draft 1, 2026-08-05. Against `316716e`, module v0.6.1, generator 0.2.0.*
+*Ranked. Every "Evidence" line was verified in the source or by running the code, not recalled.*
+
+---
+
+## The gate
+
+**DELVE has never been run at a live table.** Sixteen themes, 39 tests, six releases and two paper
+playtests of a *hand-written* delve — but no session has ever used the generated article.
+
+That fact sets the order below. Items 1–4 are things I can justify without a session, because they
+are defects or they are measurement. Items 5–9 are things I would not start until a table has run
+one delve, because a session will change what they should be. **The most valuable thing that could
+happen to this project is four hours and four players, not another feature.**
+
+---
+
+# TIER 1 — defects wearing feature costumes
+
+*These should land before anything new is built. Both are the same root cause.*
+
+## 1. Wire the live Pressure engine to the table
+
+**What.** The Foundry module should instantiate `Pressure` from the loaded delve and drive the
+clock through it, instead of reimplementing a degenerate version inline.
+
+**Why.** `pressure.mjs` is 130 lines that exist *because two playtests falsified the originals* —
+escalation, telegraphing, and pace-checking were each added to fix a specific observed failure.
+None of that reaches a table. The module's clock is:
+
+```js
+const fired = roll.total <= 1;        // vanity-delve.mjs:147
+```
+
+A hardcoded 1. So:
+
+- **Escalation never happens.** `Pressure.escalate()` widens the trigger from 1 to 2 if the clock
+  has not fired by the halfway area. Playtest 1's central failure was a clock that rolled 1-in-6
+  four times and never fired — the exact scenario escalation was written to prevent — and at the
+  table it still cannot fire.
+- **The telegraph never fires.** `telegraphDue()` is never called, so the "voice a cue one Turn
+  ahead" rule is documentation only.
+- **The timer never comes round.** `timerDue()` is never called. `st.turn` increments per area and
+  is otherwise unused.
+- **Pace-checking is generation-only.** `onPace()` shapes which Bane beats get *planned*, but the
+  live tab (`st.tab`) never feeds back, so a table that banks Banes fast or slow changes nothing.
+
+**Evidence.** The module imports exactly three core modules — `rng`, `authoring`, `forge-app`
+([`vanity-delve.mjs:17-19`](../../foundry-module/module/vanity-delve.mjs)). `Pressure`,
+`timerDue`, `telegraphDue`, `escalate`, `onPace` and `triggerNumber` appear **nowhere** in
+`foundry-module/`. The only hit for "Pressure" is an `<h3>` in the panel.
+
+**Cost.** Small. `Pressure` is already pure and serialisable; the module already persists
+`st.clock` and `st.tab` as world state. This is roughly: construct on load, replay the recorded
+rolls, call `recordClock()` instead of comparing to a literal, and surface `triggerNumber` in the
+card.
+
+**Risk.** Low, but it changes live behaviour mid-campaign for anyone with a delve loaded. Needs a
+state migration or a version check on the stored state blob.
+
+**Done when.** A delve that reaches its halfway area without a clock firing shows "something comes
+on 1–2" on the next card, and the escalation is visible to the GM.
+
+## 2. Make the pressure setting mean something
+
+**What.** Record the clock period in the delve file and derive the panel text from it.
+
+**Why.** `slow` / `standard` / `hunted` is offered in the UI, stored in `lastParams`, and passed
+into generation — and then discarded. `Pressure.summary()`
+([`pressure.mjs:119`](../../core/pressure.mjs)) emits `turn`, `clockRolls`, `clockFired`,
+`triggerNumber`, `escalated`, `tab` and `target` — **but not `period`**. So the one number the
+setting controls never leaves the constructor.
+
+**Evidence.** Generating the same seed at all three settings produces byte-identical pressure
+blocks:
+
+```
+slow     → {"turn":0,...,"triggerNumber":1,"escalated":false,"tab":6,"target":5}
+standard → {"turn":0,...,"triggerNumber":1,"escalated":false,"tab":6,"target":5}
+hunted   → {"turn":0,...,"triggerNumber":1,"escalated":false,"tab":6,"target":5}
+```
+
+And the panel hardcodes the standard value regardless
+([`forge-app.mjs:188`](../../foundry-module/module/forge-app.mjs)):
+
+> `Roll 1d6 on any attention trigger, or every 3 Turns.`
+
+A GM who picks `hunted` is told 3 Turns and gets 3 Turns.
+
+**Cost.** Trivial — add `period` to `summary()`, read it in the panel. It is item 1's prerequisite
+and probably the same commit.
+
+**Risk.** None beyond item 1. Old delve files lack `period`; default to 3.
+
+---
+
+# TIER 2 — build steps designed but never started
+
+## 3. The overlay catalog *(build step 5)*
+
+**What.** A curated encounter catalog layered over the Forge's output, with per-entry theme,
+faction, role, uniqueness — plus two fields that must be **tests, not tags**:
+
+- **playability** — *can a hero with a mundane weapon damage this at all?*
+- **survivability** — *can the party's weakest member survive one clean hit and two rounds of being
+  focused?*
+
+**Why.** Both tests come from a corpse. The Wraith's immunity made a finale literally unwinnable; a
+5d6 armour-ignoring attack that drains max Grit, against 3 Grit, is not a fight. DELVE currently
+plans *heat* and hands composition to the Forge, so it cannot promise either property.
+
+**Cost.** Large — it is authored content per theme plus a selection layer, and it interacts with
+the roster data already in the packs.
+
+**Risk.** This is the item most likely to be wrong before a live session. Heat may turn out to be
+sufficient at a real table, in which case a curated catalog is a lot of authoring for a problem
+that only appeared in simulation. **I would hold this until after the session.**
+
+## 4. A Bane calibration harness
+
+**What.** A pure-core simulator: run N delves against a modelled table, count Banes at the ending,
+report the distribution. Then set `baneTarget` from data instead of intuition.
+
+**Why.** The target has been marked **UNCALIBRATED** since draft 5. It is ~5; playtest 2 reached 3.
+Every Bane-beat decision in `baneBeatFor()` keys off `onPace()`, which keys off that number — so an
+uncalibrated target silently miscalibrates the whole tab engine.
+
+**Cost.** Small, and it needs no Foundry and no table. This is the cheapest real answer available
+right now.
+
+**Risk.** A simulated table is not a table. The output is a *prior*, not a calibration, and should
+be labelled as one — the same mistake as "6 areas: validated" would be easy to repeat here.
+
+**Done when.** DESIGN.md can state a target with a measured distribution behind it, and open
+question 4 either closes or gets sharper.
+
+---
+
+# TIER 3 — worth doing, but after a session
+
+## 5. The play sheet as a Foundry surface
+
+**What.** Render the play sheet in-world — a GM window carrying the current area's prose, the
+decision, the numbers, and the live tab, instead of scrollback.
+
+**Why.** The play sheet is currently CLI markdown only. At the table the GM card is *chat*, and
+chat scrolls away. The whole two-surface design exists because table-facing material must be
+scannable; a surface you have to scroll back through is not.
+
+**Cost.** Medium. Renderer exists (`render-authoring.mjs`); this is a Foundry Application over it.
+
+## 6. In-world authoring
+
+**What.** Write the read-aloud, notes and title inside Foundry — the `authored` layer is already
+the right shape for it.
+
+**Why.** Today the only authoring path is the CLI, which means a GM who wants to fix one area's
+prose leaves the VTT, edits JSON, and reloads. The module can already *detect* unfinished work
+(`outstanding()` is exposed on `game.delve`) but offers no way to resolve it.
+
+**Cost.** Medium. Needs a save-back path to `worlds/<id>/delves/`, which the module currently only
+reads.
+
+**Risk.** Two writers on one file. Needs a clear rule about which side owns the file.
+
+## 7. A theme quality audit beyond Barrow
+
+**What.** Read one generated delve per theme against the barrow benchmark, and add whatever
+invariant each failure implies.
+
+**Why.** Sixteen packs pass `validate-pack.mjs`, but validation is structural — it proves a pack
+*can* fill six areas without repeating, not that the result is any good. Only barrow has ever been
+read closely against a playtest. **A pack can pass every invariant and still be dull.**
+
+**Cost.** Medium, and mostly reading rather than coding.
+
+**Risk.** None. This is the item most likely to find something surprising per hour spent.
+
+## 8. The growing delve map *(build step 7)*
+
+**What.** A map that accumulates as areas are staged — the pointcrawl made visible.
+
+**Why.** v1 is explicitly a pointcrawl and players will ask where they are.
+
+**Cost.** Large. Browser-only by necessity (`forgeStage` needs `Image`/canvas/`XMLSerializer`), so
+it cannot be tested in the CLI, which is where all 39 tests live.
+
+**Risk.** High. This is the feature most likely to consume a week and produce something a GM
+sketches better on paper. **I would want a session to ask for it before building it.**
+
+## 9. Injectable RNG in the Forge → full replay
+
+**What.** Thread a seeded RNG through VANITY's Forge so a seed replays the population too.
+
+**Why.** It closes the honesty gap in the seed promise (§5 of the app draft).
+
+**Cost.** Large and it is in the *other* repo — 36 `Math.random` sites across four helpers. It was
+explicitly deferred as not gating a first release, and that judgment still looks right.
+
+---
+
+# TIER 4 — proposed and rejected
+
+**More themes.** Sixteen is already more than has been validated at a table. Another family adds
+authored content behind an unmeasured quality bar. **Stop until item 7 runs.**
+
+**Multi-motif or two-faction delves.** The fiction model coheres *because* it has one root draw.
+Two motifs is the most plausible way to break the thing that currently works best.
+
+**Generated read-aloud prose.** Named in DESIGN.md as the one dishonest thing this tool could do.
+Still true. The cue-fragment ceiling is the honest one.
+
+**A balance guarantee.** Out of scope by design, and the overlay tests (item 3) are the correct
+scoped version of this instinct.
+
+---
+
+## Suggested sequence
+
+1. **Items 1 + 2** — one commit, small, fixes a rule that is documented but not running.
+2. **Item 4** — cheap, no table needed, turns an admitted unknown into a number.
+3. **Run a session.** Four hours, four players, one barrow delve, generated not hand-written.
+4. **Item 7** while the session is fresh.
+5. Re-rank 3, 5, 6, 8 against what the session actually showed. **I expect at least one of them to
+   look wrong afterwards, and I would rather find out which one before building it than after.**
+
+## Where I am least confident
+
+- **Item 3's priority.** I have ranked a designed-and-documented build step below a simulation
+  harness and a reading exercise. That is arguable, and the argument turns entirely on whether
+  heat-without-curation survives a real fight.
+- **Whether items 5 and 6 are one feature.** A play-sheet window that cannot be edited may be half
+  a thing; shipping both at once may be the smaller total change.
+- **The whole ranking assumes the session happens.** If it realistically will not happen soon, tier
+  3 should be reordered to whatever makes the tool most useful at the desk — probably 6, then 5.
diff --git a/foundry-module/module/forge-app.mjs b/foundry-module/module/forge-app.mjs
index ace7b57..aedd119 100644
--- a/foundry-module/module/forge-app.mjs
+++ b/foundry-module/module/forge-app.mjs
@@ -18,6 +18,33 @@ const cap = s => (s ? s[0].toUpperCase() + s.slice(1) : s);
 const esc = s => foundry.utils.escapeHTML?.(String(s ?? '')) ?? String(s ?? '');
 
 const MOD = 'vanity-delve';
+
+/**
+ * What a forged foe actually is.
+ *
+ * A pack's roster describes the encounter DELVE *planned* — "1 Ghoul and 2 Skeletons", with stats
+ * and, in some themes, "blessed, silvered or magical weapons ONLY". The Forge does not take a
+ * cast: it rolls its own monsters from the heat. So the roster and the actors in the world are two
+ * different lists, and printing the roster's numbers beside the Forge's actors told the GM to run
+ * a fight against foes that were never created.
+ *
+ * Read the numbers off the documents that exist. The plan is still worth showing — it carries
+ * authored tactical guidance — but it has to be labelled as the plan.
+ */
+export const foeStats = a => ({
+  name: a.name,
+  uuid: a.uuid,
+  atk: a.system?.attack1?.pool ?? null,
+  def: a.system?.defence?.pool ?? null,
+  grit: a.system?.grit?.value ?? null,
+  nerve: a.system?.nerve ?? null,
+  trick: a.system?.trick ?? '',
+});
+
+/** One foe as a line of stats, in the roster's own vocabulary so the two read alike. */
+export const foeLine = f =>
+  `<b>@UUID[${f.uuid}]{${f.name}}</b> — ${f.atk ?? '?'}/${f.def ?? '?'}/${f.grit ?? '?'}, Nerve ${f.nerve ?? '?'}${f.trick ? `. <i>${f.trick}</i>` : ''}`;
+
 let THEMES = [{ id: 'barrow', label: 'Barrow' }];
 
 export function setThemes(list) { THEMES = list; }
@@ -123,7 +150,7 @@ export async function raiseDungeon(params = {}) {
         heat: area.encounter.heat, forStage: area.name,
         ...(seams ? { hoard: false, post: false, folderId: folders.Actor.id } : {}),
       }).catch(e => { console.error('DELVE | encounter failed', e); return null; });
-      area._foes = (enc?.actors ?? []).map(a => ({ name: a.name, uuid: a.uuid }));
+      area._foes = (enc?.actors ?? []).map(foeStats);
     }
     if (area.hoard) {
       const h = await game.vanity.forge.hoard({ size: area.hoard, ...quiet })
@@ -212,11 +239,12 @@ function buildPages(d, scenes) {
           ${rv.failure ? `<li><b>Miss</b> → ${esc(rv.failure)}</li>` : ''}
           ${rv.orElse ? `<li><b>Or</b> ${esc(rv.orElse)}</li>` : ''}
         </ul>
-        ${R ? `<p><b>${esc(cap(a.encounter.heat))} — ${esc(R.line)}.</b> Harmed by ${esc(R.harmedBy)}${R.harmedBy.includes('ONLY') ? ' — <b>say so before initiative</b>' : ''}. <i>${esc(R.avoid)}.</i></p>
+        ${a._foes?.length ? `<p><b>${esc(cap(a.encounter.heat))} — in the world.</b> These are the actors the Forge created; run the fight off these.</p>
           <table><thead><tr><th>Foe</th><th>atk</th><th>def</th><th>Grit</th><th>Nerve</th><th></th></tr></thead><tbody>
-          ${R.foes.map(f => `<tr><td>${f.n}× ${esc(f.name)}</td><td>${f.atk}</td><td>${f.def}</td><td>${f.grit}</td><td>${f.nerve}</td><td><i>${esc(f.note)}</i></td></tr>`).join('')}
+          ${a._foes.map(f => `<tr><td>@UUID[${f.uuid}]{${esc(f.name)}}</td><td>${f.atk ?? '?'}</td><td>${f.def ?? '?'}</td><td>${f.grit ?? '?'}</td><td>${f.nerve ?? '?'}</td><td><i>${esc(f.trick)}</i></td></tr>`).join('')}
           </tbody></table>` : ''}
-        ${a._foes?.length ? `<p><b>Rolled for you:</b> ${a._foes.map(f => `@UUID[${f.uuid}]{${esc(f.name)}}`).join(' · ')}</p>` : ''}
+        ${R ? `<p><b>DELVE planned ${esc(R.line)}</b> — the Forge rolls its own foes, so the plan and the table above are different lists.
+          Its guidance still applies to the scene: harmed by ${esc(R.harmedBy)}${R.harmedBy.includes('ONLY') ? ' — <b>only true if you cast the fight yourself</b>' : ''}. <i>${esc(R.avoid)}.</i></p>` : ''}
         ${a.temptation ? `<p><b>${esc(cap(a.temptation.id))}</b> — ${esc(a.temptation.cue)}: ${esc(a.temptation.benefit)}.<br>
           <i>Using it costs ${a.temptation.useCost?.bane ? `+${a.temptation.useCost.bane} Bane` : '—'}. While carried, ${esc(a.temptation.standingDrawback)}.</i></p>` : ''}
         ${a._hoard?.length ? `<p><b>Hoard.</b></p><ul>${a._hoard.map(l => `<li>${l}</li>`).join('')}</ul>` : ''}
diff --git a/foundry-module/module/vanity-delve.mjs b/foundry-module/module/vanity-delve.mjs
index 27aa901..345c6bd 100644
--- a/foundry-module/module/vanity-delve.mjs
+++ b/foundry-module/module/vanity-delve.mjs
@@ -16,7 +16,7 @@
  */
 import { coinSeed, Rng } from './core/rng.mjs';
 import { newWorkingFile, outstanding, readyToPlay } from './core/authoring.mjs';
-import { DelveForgeApp, raiseDungeon, listDungeons, removeDungeon, removeDungeonDialog, setThemes } from './forge-app.mjs';
+import { DelveForgeApp, raiseDungeon, listDungeons, removeDungeon, removeDungeonDialog, setThemes, foeStats, foeLine } from './forge-app.mjs';
 
 const MOD = 'vanity-delve';
 const FLAG = 'state';
@@ -27,6 +27,29 @@ let loadPack = async () => null;
 
 const getState = () => game.settings.get(MOD, FLAG) ?? null;
 const setState = async s => game.settings.set(MOD, FLAG, s);
+
+/**
+ * The pack a delve was actually authored against.
+ *
+ * PACK is only a boot-time default. A delve brought in from the desk may be any of the sixteen
+ * themes, and staging one with the wrong pack gives a harbour delve a barrow map — the fiction
+ * says quayside and the geometry says burial chamber. Resolved per call rather than once at load,
+ * because the global resets on a page reload while the staged delve in world state does not.
+ *
+ * `params.theme` is written from `pack.id` at generation time, so it is always present and always
+ * right, whatever the caller passed.
+ */
+async function packFor(d) {
+  const id = d?.params?.theme ?? 'barrow';
+  if (PACK?.id === id) return PACK;
+  const p = await loadPack(id);
+  if (!p) {
+    ui.notifications.error(`DELVE: could not load the ${id} theme — refusing to stage, it would use the wrong geometry.`);
+    return null;
+  }
+  PACK = p;
+  return p;
+}
 const cap = s => (s ? s[0].toUpperCase() + s.slice(1) : s);
 const list = items => `<ul>${items.filter(Boolean).map(i => `<li>${i}</li>`).join('')}</ul>`;
 
@@ -51,6 +74,7 @@ async function load(working) {
   if (!d?.skeleton || !Array.isArray(d.areas)) return ui.notifications.error('DELVE: that is not a delve file.');
 
   const todo = outstanding(d);
+  if (!await packFor(d)) return;                 // before anything is created in the world
   const folder = await Folder.create({ name: `Delve — ${d.authored?.title ?? d.skeleton.placeName}`, type: 'Actor' });
   await setState({ delve: d, folderId: folder.id, at: 0, turn: 0, tab: [], clock: [] });
 
@@ -75,7 +99,9 @@ async function loadFile(name) {
 /** Generate an unfinished draft in-world. Convenience only — the desk is the right place. */
 async function draft(params = {}) {
   const seed = params.seed || coinSeed(new Rng(String(game.world.id)));
-  const d = newWorkingFile({ pack: PACK, ...params, seed });
+  const pack = await packFor({ params });
+  if (!pack) return;
+  const d = newWorkingFile({ pack, ...params, seed });
   ui.notifications.warn('DELVE: unfinished draft. Write the read-aloud in the worksheet first.');
   return load(d);
 }
@@ -89,16 +115,18 @@ async function enter() {
   const area = d.areas[st.at];
   const w = d.authored?.areas?.[area.index] ?? {};
   const name = w.nameOverride ?? area.name;
+  const pack = await packFor(d);
+  if (!pack) return;
   ui.notifications.info(`DELVE: raising ${name}…`);
 
   const quiet = seamsPresent ? { post: false, folderId: st.folderId } : {};
   const stage = await game.vanity.forge.stage({
-    type: PACK.forgeStageType, size: 'medium', name, populate: false, activate: true, ...quiet,
+    type: pack.forgeStageType, size: 'medium', name, populate: false, activate: true, ...quiet,
   });
-  if (area.encounter) await game.vanity.forge.encounter({
+  const enc = area.encounter ? await game.vanity.forge.encounter({
     heat: area.encounter.heat, forStage: name,
     ...(seamsPresent ? { hoard: false, post: false, folderId: st.folderId } : {}),
-  });
+  }) : null;
   if (area.hoard) await game.vanity.forge.hoard({ size: area.hoard, ...(seamsPresent ? { post: false } : {}) });
 
   // Players first — the scene is up and this is what they came for.
@@ -106,13 +134,14 @@ async function enter() {
 
   // Then the GM, quietly.
   const R = area.encounter?.roster;
+  const foes = (enc?.actors ?? []).map(foeStats);
   const rv = area.decision?.resolve ?? {};
   await gmCard(`⛏ ${area.index} · ${name}`, `${area.role} · ${area.facet}`,
     `${area.situation ? `<p><b>Here:</b> ${cap(area.situation.occupant)}, ${area.situation.doing} — ${area.situation.onArrival}.<br>
         <b>They can:</b> ${area.situation.offer}. <i>${cap(area.situation.because)}.</i></p>` : ''}
      <p><b>${cap(area.decision.cue)}</b>${rv.roll ? ` — [${rv.roll}] ${rv.success}` : ''}${rv.failure ? `<br><b>Miss:</b> ${rv.failure}` : ''}${rv.orElse ? `<br><b>Or:</b> ${rv.orElse}` : ''}</p>
-     ${R ? `<p><b>${cap(area.encounter.heat)}:</b> ${R.line}</p>${list(R.foes.map(f => `${f.n}× <b>${f.name}</b> — ${f.atk}/${f.def}/${f.grit}, Nerve ${f.nerve}. <i>${f.note}</i>`))}
-            <p><b>Harmed by ${R.harmedBy}.</b> ${R.avoid}.</p>` : ''}
+     ${foes.length ? `<p><b>${cap(area.encounter.heat)} — in the world:</b></p>${list(foes.map(foeLine))}` : ''}
+     ${R ? `<p><b>DELVE planned ${R.line}</b> — the Forge rolled its own, so run the block above. ${R.avoid}.</p>` : ''}
      ${area.temptation ? `<p><b>${cap(area.temptation.id)}:</b> ${area.temptation.benefit}. <i>Use: ${area.temptation.useCost?.bane ? `+${area.temptation.useCost.bane} Bane` : '—'}. ${area.temptation.standingDrawback}.</i></p>` : ''}
      ${w.notes ? `<p><b>Your note:</b> ${w.notes}</p>` : ''}
      <p><code>${area.trigger}${area.baneBeat ? ` · ${area.baneBeat}` : ''} · fallback: ${area.fallback.route}</code></p>`);