vanity-delve/DESIGN.md

# DELVE — a dungeon layer for VANITY

*Working name `vanity-delve`. **Draft 6**, 2026-08-05.*
*First draft describing a tool that **exists**. The core is built, tested and generating.*

---

## 0. Status

| Build step | State |
|---|---|
| 1. Handwrite a delve, run it twice on paper | ✅ complete — [`paper/`](paper/) |
| 2. Seam spike | ✅ complete — [`spike/SPIKE-REPORT.md`](spike/SPIKE-REPORT.md) |
| 3. Land the Forge seams | ✅ **branch ready for review** — `~/projects/Vanity` branch `delve-seams`, committed, **not pushed** |
| 4. Core + Barrow content | ✅ built — [`core/`](core/), 11/11 tests passing |
| 5. Curated overlays | not started |
| 6. Foundry vertical slice | not started |
| 7. Growing Delve Map | not started |
| 8. More themes | not started |

**What runs today:** `node core/cli.mjs --seed=gilded-court-404` emits a complete, coherent,
GM-readable delve. Samples in [`samples/`](samples/).

---

## 1. What changed since draft 5

| Change | Source |
|---|---|
| **Seven seams → six → four landed** | Spike: `populate:false` already existed; I listed work that was already done |
| **Seam 4 narrowed** | `forgeEncounter` already returned a good object; only `forgeHoard` returned nothing |
| **`post:false` promoted to the decisive seam** | Spike: 3 chat cards per area, 9 per delve, unsuppressable |
| **Browser-only constraint stated as a fact** | Spike: `forgeStage` needs `Image`/canvas/`XMLSerializer` — the CLI is map-free by necessity |
| **6 areas: "validated" → "working default"** | Review: one Barrow, one simulated table, two passes, one heat |
| **Bane target marked UNCALIBRATED** | Review: pass 2 reached 3, not 4–6 |
| **Survivability becomes a concrete test** | Review: "can the weakest PC survive one clean hit and two rounds of focus?" not a vibe tag |
| **The fiction model exists** | New — §3 |

---

## 2. What DELVE is — an authoring system

**This changed in draft 7, after four readings of generated output.**

DELVE is the dungeon-scale layer over VANITY's Forge — it generates the fiction, the order and
the pressure, and sequences the Forge without letting it talk over DELVE. But it is **not** a tool
that hands a GM a delve to run. Four independent reviews converged:

> *"The handwritten delve still tells the GM what matters faster."*
> *"Would you run this at a real table tomorrow, as-is?"* — **No.**

Three renderer passes each fixed what the last review named and each introduced a new problem
(blocks → menu-like; subtraction → buried). The density is not a bug to render away: **everything
on the page is a vice at the table and a virtue at the desk.**

So a delve has two surfaces and one working 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. An `authored` layer holds what the GM writes and the generator
never touches it, so **rerolling a component cannot destroy prose**. Lock what is good, reroll
what is not, write the read-aloud, export the play sheet.

DELVE does not write the read-aloud and does not write the climax. It gets a GM to a strong first
draft in a minute instead of an evening, and then gets out of the way.

---

## 3. The fiction model — solved

Draft 5 asserted DELVE generates "prize, bottom problem, factions, motif, foreshadow" and never
said how. Four independent draws produce nonsense. The model that works has **one root draw**:

```
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*). Everything else is **projected**, never drawn — which is why the fiction coheres.

**Content scales additively:** 8 appetites × 6 accommodations × 7 claimants = **336 kernels from
21 authored lines**. Measured: 60 generated delves produced **51 distinct kernels**.

### Cue fragments and facets

Fragments hang off the **motif**, tagged by **facet** — `institution · ritual · demand · wound ·
anchor` — so two delves sharing a motif draw different imagery. 15 fragments per motif, 8 motifs.

The **foreshadow chain** is a fixed facet progression mapped onto the arc:

| Arc role | Facet |
|---|---|
| approach | institution |
| complication | ritual |
| turn | demand |
| descent | wound |
| threshold | anchor |
| ending | demand |

This is why area 2's held bows and area 4's scratched-out eyes both point at the same faceless
queen: they are facets of one motif, drawn in a deliberate order. The progression was proposed
independently and **turned out to describe the hand-authored paper delve exactly** — which had
arrived at it by instinct. That convergence is the strongest evidence the model is right.

Fragments are dealt without replacement, borrowing from a neighbouring facet before repeating.

**The furniture is motif-scoped too.** Codex's review found that only the motif cue fragment was
projected — decisions and temptations came from global pools, making the output "a consistent
wrapper around generic room furniture". Both now live under `motifs.<id>`: **6 decisions and 5
temptations per motif**, 88 authored entries across 8 motifs. A `keeping` delve offers a flower
that has not wilted, a lamp on three-century-old oil, and a room you may search or leave exactly
as you found it. A `precedence` delve offers a corridor that will not let two walk abreast.

Two rules the tests enforce:
- **Nothing repeats inside one delve.** Decisions and temptations are dealt without replacement,
  not picked per area — picking independently repeated a temptation in 75 of 80 test delves.
- **Motif purity beats variety for temptations.** Decisions may top up from the generic pool if a
  long delve outruns its motif; temptations may not. A generic relic breaks the fiction, while a
  repeated one only looks thin.

**Remaining limit:** the second cue fragment still comes from a global `features` pool. That is
the last piece of generic furniture.

### The appeasement move

The best mechanic in the hand-authored delve — bow back, be received as a guest, **+1 Vanity now,
+1 Bane later, repeatable** — 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 |

This is the delve's social spine, it is always thematically exact, and it is the reason the
generated delve has a way through that is not violence.

**Accommodations are filtered by appetite.** Drawing freely produced "had the faces struck from
the young, so that nothing would ever age" — legible but not folklore. Each accommodation now
declares which appetites it fits.

---

## 4. Architecture (built)

```
core/                  pure JS, seeded, zero Foundry globals, Node-testable
  rng.mjs              derived streams — seed::area::3::encounter. Order-independent.
  skeleton.mjs         the fiction kernel (§3)
  director.mjs         arc roles, heat, hoard tier, density, Bane beats
  beat.mjs             an area as a beat: Cue + Truth
  pressure.mjs         the clock and the tab — never rolls a die
  delve.mjs            composition
  render.mjs           markdown
  cli.mjs              --seed --areas --depth --density --greed --json
  content/barrow.json  the theme pack
  test.mjs             11 invariants
```

**The core never sees an Actor, a UUID, HTML or a roll.** `pressure.mjs` decides *when* a clock
roll is due and what it means; the adapter rolls it at the table. That separation is what makes
the seed promise honest.

---

## 5. The Forge seams — landed

Branch `delve-seams` in a clone of `slaguru666/Vanity`, **committed, unpushed, awaiting review**.
All default to current behaviour, so nothing changes for existing callers.

| Seam | Change | Verified |
|---|---|---|
| 1 | `forgeEncounter({ hoard: false })` | loot decoupled from combat |
| 3 | `{ post: false }` on all three | **9 chat cards → 0** across three areas |
| 4 | `forgeHoard` returns `{size,label,goods,relics,lines,card}` | was `undefined` |
| 5 | `forgeEncounter({ folderId })` | **3 folders → 0** |
| — | guard two unchecked `rnd()` calls | latent crash on an empty pack |

**Deferred:** injectable RNG (36 `Math.random` sites, four helpers) and live Reaction `Roll`.
Neither gates a first release.

---

## 6. Parameters

| Parameter | Range | Default | Confidence |
|---|---|---|---|
| **Areas** (excludes ending slot) | 3–12 | **6** | *working default* — one theme, one table, two passes |
| Theme | Barrow (+4 planned) | Barrow | only Barrow exists |
| Depth | 1–5 | 2 | |
| Party | 1–8 | 4 | |
| Deadliness | forgiving/standard/cruel | standard | **severity only** |
| Density | sparse/standard/infested | standard | **frequency only** — 2 fights in 6 |
| Greed | lean/standard/glutted | standard | |
| Pressure | slow/standard/hunted | standard | |
| Ending | authored/generated | authored | a default, not a law |
| Skeleton seed | string | random | **not a replay seed** |

---

## 7. Encounters and the overlay catalog

Heat is a coarse shape dial. Authored overrides stay for the bottom problem and any named foe.

The overlay needs six fields. The last two came from two dead mages, and the review is right that
they must be **tests, not tags**:

- theme · faction · role · uniqueness
- **playability** — *can a hero with a mundane weapon damage this at all?* (the Wraith's immunity
  made a finale unwinnable)
- **survivability** — *can the party's weakest member survive one clean hit and two rounds of
  being focused?* (a 5d6 armour-ignoring attack that drains max Grit, against 3 Grit, cannot)

Not built yet — step 5.

---

## 8. Temptation — at the point of use

Both playtests: every free-acceptance item was taken without discussion. Free acceptance is not a
temptation, it is an automatic yes. The decision lives at **use**.

```json
{ "id": "ledger", "cue": "a book of names, one line left blank",
  "benefit": "write a name and the Host obeys it once",
  "acceptanceCost": { "vice": "Pride" },
  "useCost": { "bane": 2 },
  "standingDrawback": "your own name is now in it" }
```

A minority carry real acceptance friction so "do we take it?" is sometimes a genuine question.
Vanity is **fuel, not a cost** (`vanity.mjs:1224`). No sheet writes in v1 — DELVE offers.

---

## 9. Pressure

**Clock: triggers lead, the timer is a floor.** Every area emits an attention trigger. Escalates
if it has not fired by the halfway area. Telegraphed one Turn ahead. **The core never rolls it.**

**The tab: Banes are emitted by named beats**, never hoped for — playtest 1 banked 1 in ~40 rolls
because Stumbles are rare. The director offers a Bane beat when the tab falls behind pace.

**Target: ~5 Banes by the ending. UNCALIBRATED** — playtest 2 reached 3. Needs another pass.

---

## 10. The skeleton seed

| Reproducible | Recorded only | Live |
|---|---|---|
| Fiction, area order, heat, hoard tier, foreshadow chain | Encounter composition, mood, hoard contents | The clock, player choices |

The UI must say this plainly. A partial seed sold as a full one is worse than no seed.

---

## 11. What v1 will not do

- No physically coherent tiled dungeon — **v1 is a pointcrawl**.
- No balance guarantee.
- No sheet writes.
- **No system-generated climax by default** — the ending slot is the GM's.
- No full replay from seed until the 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.

---

## 12. The quality verdict — and what it changed

Codex read three generated delves against the playtested hand-written one:

> *"The paper delve has procedures. The generated delves mostly have prompts."*
> *"Not yet... still a very good random-table document with a coherence engine and CLI in front."*

That was right, and specific. The cue-fragment gap was **not** the problem — fragments are fine
if the rest of the packet is operational. It wasn't. Fixed since:

- **The appeasement move** (above) — the social route, projected, repeatable, priced
- **Concrete encounter blocks** — foe counts and atk/def/Grit/Nerve in a table, plus
  **"harmed by"** stated before initiative and **"avoidable"** spelled out
- **Bane beats state who, when, and whether they repeat** — not just a tag
- **A trigger glossary**, so `seenTwice` is not shorthand
- **Accommodation/appetite compatibility**, killing the equation-output kernels

Still missing, and still the gap to a document: per-area read-aloud, exact route costs and
failure outcomes, concrete clue answers, and a full ending packet with negotiation terms.

## 13. Open questions

1. **Does the module beat a document?** The spike answers the mechanical half — unmodified, the
   Forge emits 9 cards and 3 folders per delve that no document could intercept, and the seams
   take that to 0. The *table* half is still open until a live session.
2. Is `Ending: Authored` right, or right only for its author?
3. Is 6 areas right beyond one theme and one simulated table?
4. Is the ~5-Bane target reachable without feeling mechanical?
5. Do cue *fragments* work at a live table, or does a GM need prose?

---

*Private VANITY development material — not for sale or distribution.*