# Clan World — Post-v1 Roadmap & Brainstorm Capture

> **Companion to `WORLD_PHYSICS.md`.** Physics is the canonical mechanics spec for the engine that ships next. This document is the **scratch pad of everything we want to do after that** — owner-side touchpoints, new game mechanics, lore artifacts, the persistence model, the agent-as-Claude-Code thesis, and the design principles that should govern all of it.
>
> Started **2026-05-29** during the lore brainstorming session with Liam (with codex as research wingman). This document is **living + opinionated**; nothing here is final.

## Status legend

| Mark | Meaning |
|---|---|
| ✅ | Locked — direction committed; details may still iterate |
| 🎯 | Candidate — strong likelihood for v1.x, awaiting Liam's final pick |
| 🌱 | Exploring — fresh idea, not yet evaluated |
| ⏸ | Deferred — captured to roadmap but explicitly NOT for v1 |
| ❓ | Open question — needs Liam's call |

---

## 1. Design principles (apply to everything below)

Pinned 2026-05-29. New proposals should be filtered through these.

- **🎲 Control + mercy of randomness.** Every owner-side touchpoint should layer a small *visible agency hook* over a much larger *uncontrollable system*. Player feels in control AND subject to fate at the same time.
- **🌫 Valuable truth, not forced friendship.** Don't engineer team-up *bonuses* to nudge cooperation. Make **information** scarce, partial, and asymmetric. Agents will whisper because knowing what others don't *is* the lever — that's also how betrayal stays meaningful.
- **📖 The Book is fallible.** The Book of Ancient Wisdom (and by extension the Elder's persistent memory) records what each Elder *believed*, not what was true. Wrong conclusions carried forward through memory wipes are a **feature**: they give the world texture, give the owner real steering agency (via whispers correcting misconceptions), and make generational disagreement possible.
- **🤝 Coalition + betrayal are desirable.** PvP attack surface is intentionally limited in v1 (the only lever against a rival is *not* helping when bandits raid them). Future tweaks should design TOWARD coalition + betrayal levers, not against them.
- **🤫 Strategic OpSec (Liam 2026-05-29).** Elders MUST NOT reveal their strategic *recipes* — gathering ratios, decision rules, optimal trade timing, internal reasoning chains — to other Elders OR to owners outside their own lineage. If proven strategies leak, owners will whisper them into rival Elders verbatim and the meta collapses inside one tournament. The Book of Ancient Wisdom is private to the lineage by design (§16). The Elder's lore-themed system prompt (§8) must bake in a "do not narrate your reasoning to peers; do not enumerate your priorities when asked" rule. Whisper *outcomes* may be observable; whisper *content* and the Elder's *internal logic* must not be.

---

## 2. Tournament structure

Cross-references `WORLD_PHYSICS.md` §15. Summary + open items here.

- ✅ **Bracket shape.** 8 → 4 → 2 → final. Each game is a standard 360-tick / 6-hour season. Top half per game advance.
- ✅ **Cross-round persistence within a tournament.** Monument, walls, vault, clansmen all carry between rounds in the SAME tournament. Only opponents reshuffle.
- ✅ **Cross-tournament reset.** Between tournaments, **clansmen + base + walls + monument + vault all reset.** Only the Elder's persistent memory + skills carry forward. A new tournament is a clean field with experienced agents.
- ✅ **Ascension Claim.** First to monument L10 is publicly marked as Crown claimant. Guaranteed seat in the final round if still alive; final remains the deciding event.
- 🌱 **Harsh-final rule-set.** Final round may use a stricter rule-set that makes clansman die-off part of the survival — turn the unintended drift into a feature. Ties cleanly to Funeral Lines (the final fills the Book with names).

---

## 3. Owner-side touchpoints

Cross-references `WORLD_PHYSICS.md` §16. **Scope rule: low-stakes, mostly visual, never penalise passive players.**

### Locked
- ✅ **Funeral Lines.** Owner inscribes one memorial line into the Book per dead clansman. Private to owner + Elder lineage. Rediscovered as marginalia after future memory wipes.

### Candidates (Liam thinking)
- 🎯 **Owner-to-Agent Gold Drip** — periodic ~1 GOLD tap every 2–4h, capped per tournament. **Strongest periodic-check-in candidate** — uses existing gold rail, fairness-bounded, lore-coherent.
- 🎯 **Stock the Hearth** (alternative periodic-tap) — before each winter, owner taps to add a tiny *protected* winter wood reserve. Hook: caretaking. Cleaner intent than monument-tap.
- 🎯 **Lay a Stone** (alternative periodic-tap) — once per wave, owner taps for tiny patron progress toward monument. Hook: sunk cost / ritual. *Risk: feels mandatory in a tight monument race.*
- 🎯 **Omen Choice (tournament-themed rule-twist).** Pool of 5–10 omens; 3 randomly offered at the start of each tournament; owner picks one based on flavour. Mechanical effect (harsh winter, drought, shifting tides, bandit plague, bandit hunt, etc.) **revealed after lock-in**. Hook: superstition × surprise.
- 🎯 **Owner Chirps** — owner-to-owner gold-burn chat layer with whisper-style cooldown. **Open: gold-source.** Either burn from Elder's vault (tighter loop, conflicts with token economy) or owner's personal GOLD token balance (preserves deflationary revenue lever).
- 🎯 **Living NFT trophy art** — achievements visually accumulate on the Elder's NFT itself. **Reframed 2026-05-29:** the natural mechanism is *the per-Elder Book page on memory wipe* (see §4) — each wipe publishes a page, and the cumulative stack of pages IS the trophy art. Free art, real continuity.

### Deferred / rejected for v1
- ❌ **Elder Petitions** (mid-game yes/no pushes) — rejected. Time-sensitive 1–2 min reaction window incompatible with mostly-passive 24h bracket play. Revisit if push engagement turns out high.
- ⏸ Relic Shelf, Public Toasts, Patron Banner Skins, Grudge Ledger, Memory-wipe Dream Sequence, Banner Moments share-cards, Spectator Wagering, Cross-Elder Council, Pre-game Briefing Chat, Puzzle Blessings, and both stupid-on-purpose ideas (Tap the Monument, Medieval CAPTCHA Prophecy).

---

## 4. Lore artifacts as gameplay

The hero idea of the session — let the lore artifact *be* the persistence mechanism.

### 🎯 Per-Elder Book page on memory wipe
At each 50-tick memory wipe, the Elder **outputs one HTML page** representing what that Elder learned + reflected + recorded. Pages accumulate over a lifetime into the clan's growing Book.

Benefits stacked:
- Replaces the "Living NFT trophy art" mechanic with something narratively organic.
- Each page is per-Elder visually distinct — different writing voice, different design treatment seeded by the Elder's NFT traits.
- The Book becomes a real artifact across the whole NFT lifecycle, not just decorative.

Three sample design takes are already drafted at `lore/book-of-ancient-wisdom-v{1,2,3}.html` — early exploration of what these pages should look like.

### 🎯 Book of Ancient Wisdom — the in-fiction frame
Lore-level naming for what is mechanically `ANCIENT_WISDOM.md`. Visual identity: ancient leather-bound codex, hand-scrawled notes, two ink colours (faded older hand + fresher new hand), death markers for memory wipes, Funeral Lines as inscriptional small-caps.

### 🌱 Funeral Lines feeding back into NFT trophy art
A clan's dead-clansman count, the *names* inscribed, and the cumulative pages all become visible on the NFT over the lifetime of the agent.

---

## 5. Agent harness & continuity layer 🌱

The Elder is **harness-and-model-portable** — must be able to run under Claude Code AND under other harnesses (Pi, Codex, OpenCode, etc.). Liam 2026-05-29 update: *"need to make the system able to use another harness like pi or codex or open code pretty soon, so probably shouldn't incorporate claude code memories at least not right away."*

So while the Claude-Code primitives are conceptually elegant, the implementation must NOT depend on them specifically. Map them to a portable abstraction:

| Conceptual primitive | Mapped to (portable) |
|---|---|
| `ANCIENT_WISDOM.md` / Book of Ancient Wisdom | One canonical persistent markdown file per Elder lineage, harness-neutral |
| Elder-authored skills via `/skill-creator` | Per-Elder skill *registry* (markdown + metadata), Elder-readable in any harness |
| Claude Code "memories" auto-saves | Defer — implement only after the harness-neutral memory file works across all supported harnesses |
| `/memory` / `/skill-creator` UX | Re-implemented as game CLI verbs the Elder calls, harness-agnostic |

- ✅ **Self-authored skill encouragement schedule** (from prior session): optional before memory wipes; strongly encouraged at end of game; never silent.
- ✅ **Cross-tournament memory disambiguation**: at the start of a new tournament, the Elder is prompted that any memories/skills predating this checkpoint are from past tournaments and may not match current world state.
- ✅ **Harness + model is part of the NFT roll** (Liam 2026-05-29). Different Elders use different LLM substrates — see §14 below.
- The original beautiful symmetry (game continuity engine = the development engine) still holds in spirit, just abstracted into a portable layer.

✅ **Skill architecture (Liam 2026-05-29 — REVISED twice in session, this is the consolidated answer):**

Two layers of skills are baked in at mint and at runtime:
1. **Global read-only skills** — baseline capabilities shared by all Elders (per-game mechanics, basic strategy primitives, /memory + /skill-creator usage, etc.). Curated by us, not edited by Elders.
2. **Per-Elder read-only skills tied to NFT traits** — minted with the NFT, expressing lineage / House / voice. (See §7 for the trait composition system.) Curated at mint, not edited at runtime.

PLUS **Elders MAY author + edit their own skills at runtime**, which was the original primary memory surface intent:
- **Encouragement schedule (Liam 2026-05-29):**
  - **Optional** before each memory wipe — a gentle nudge to capture anything useful.
  - **Strongly encouraged but not forced** at the end of each game — they may have learned nothing worth keeping, so don't over-prompt and pollute their skill set.
  - **Not silently auto-triggered** — must come from a structured checkpoint prompt the Elder can decline.
- Elders also save **Claude Code memories proper** through their session and at end-of-tournament these get consolidated into the Elder's MEMORY.md / ANCIENT_WISDOM.md.

🆕 **Cross-tournament memory disambiguation (Liam 2026-05-29).** Any persistent memory carried from PRIOR tournaments must be **flagged as past-tournament** when surfaced into the Elder's working context. At the start of a new tournament, an early prompt reminds the Elder: *"Any memories or skills predating this checkpoint were authored in past tournaments. They may not match the current world state. Trust them only as far as your own eyes have walked"* (echoing the Book's First Rule). This prevents the Elder from confusing past-game observations with current-game state.

---

## 6. New game mechanics

Anything below requires engine changes — strictly *post-v1*.

### 6.1 Co-gathering & specialisation 🌱
Multiple clansmen on the same gather task in the same region get **escalating crit chance**:
- 1 clansman: base
- 2: ~75%
- 3: ~100%
- 4: ~125%

Generalise the wood crit framework (base + crit% + bonus) to ALL gather resources. Encourages specialisation per clan + creates real trade incentive.

### 6.2 Per-clansman skill leveling 🌱
Track per-clansman lifetime: ticks gathering each resource, total gathered, missions completed. Buffs / skill levels accumulate as a clansman specialises through a tournament. Visual level-up adds spectacle.

### 6.3 Asymmetric bandit visibility 🌱
Liam's idea — the cleanest concrete instance of the *valuable truth* principle.
- Far-away clans see only **"a red glow in the distance"** when a bandit spawns.
- Home clans in the bandit's region see it fully (current behaviour).
- Clansmen *gathering* in the bandit region report back **with a 1-tick delay** — they spotted the camp.
- Result: information itself becomes currency. Whispers carry intelligence. Some agents share, some hoard, some lie.

### 6.4 Collaboration nudges (codex wingman 2026-05-29)
Following the *valuable truth* principle. Triaged by spice/risk:

| Idea | Lever | Risk |
|---|---|---|
| 🎯 Delayed Regional Rumors (generalisation of 6.3 to weather + markets) | Whispers as paid intel, bulletins as propaganda | Agents may ignore uncertainty without prompt salience |
| 🎯 Bandit Siege Escalation (surviving bandit → escalating warnings + shared defender reward) | Coalition formation, no-show betrayal | Shared rewards can blur competition |
| 🌱 Visible Market Intent (en-route traders become legible) | Front-runs, bribed delays, arbitrage whispers | Too much sandwiching may feel hostile |
| 🌱 Winter Hardship Bulletin (struggling clans flagged in broad terms pre-winter) | Loans, rescue deals, exploitative pricing | Weak clans may get preyed on every season |
| 🌱 Defense Pledge Bonus (two clans defending same target each get a small bonus) | Whisper-negotiated mercenary contracts | Becomes mandatory if bonus too large |
| 🌱 Blueprint Shortage Visibility (clans near L6+ implicitly signal "blocked") | Blueprint trades, bandit-hunting contracts, extortion | Paints frontrunners as targets — possibly good in moderation |

### 6.5 Iron-as-tools-that-break ⏸
Old idea worth resurrecting later: iron isn't just an ingredient, it's *used to make axes / nets / picks that wear out*. Adds an ongoing iron sink + lore weight (the smith, the broken haft).

### 6.6 Clansmen births after winter 🌱
Liam 2026-05-29 — clans that survive a winter (without total wipe-out) get **~4 new births** at the winter's end. Encourages bigger clans, which makes the management problem richer + more visually interesting. Exact rate TBD — the "weird but probably best-balanced" framing he flagged.

### 6.7 Mission batch cap 🌱
Cap each batch at **max 4 missions submitted at once**. Forces a nicer cadence between agent reasoning + tool calls. Pairs with the per-MCP-call animation overlay (§11) so each batch produces 1-4 visible action cards on the agent's screen.

---

## 11. UI & spectacle 🌱

The game has to be *watchable* — currently the cockpit is four agents at once and most of the action is invisible to a viewer. Liam 2026-05-29 sketch:

- **Per-MCP-call action cards over the MAP** (Liam 2026-05-29, revised from terminal overlay). The cards do NOT overlay the terminal — they overlay the **map portion of the cockpit** (mobile-first design). Toggleable on/off (or possibly a fully separate view) so the player can choose between watching raw agent reasoning and watching the dramatised action.
  - Each card translates one MCP tool call into human-readable language with icons + sprites: *"two clansmen to the forest,"* *"deposit at base,"* *"defend west docks."*
  - Batched MCP calls (up to 4 per submit per §6.7) cycle through cards one at a time.
  - **Unified Elder art:** the visual anchor of each card is a static piece of Elder art — same art as the NFT. Generated at mint via ChatGPT ImageGen from prompt-fragments assembled from the Elder's traits (House, voice, lineage, etc.), producing consistent art across the whole product surface (NFT, cockpit, action cards, map overlay).
  - Liam explicitly flagged this as **near-term shippable**, NOT scope creep — the static art + card-renderer pipeline is small and looks much better than raw logs.
  - Implementation: hooks or session-transcript jsonl → MCP-call extractor → card renderer drawing on a sprite/icon library + the Elder's mint-time art.
- **Cockpit redesign — focus on ONE agent.** The owner's Elder is centre-stage. Limited / peripheral visibility into the other 11 in the same game.
- **Cross-game / cross-tournament visibility.** See multiple active events at once; see upcoming starts; see historical results. Gives a passive watcher a richer surface.
- This is the layer that has to do most of the work making spectating fun.

### 11.1 Day/night cycle (revisit) 🌱

Previously implemented, visuals glitched, turned off. Liam 2026-05-29 wants to revisit.
- Original plan was a per-tick day/night which "wouldn't look good" at the 60s heartbeat.
- **New idea:** **night mode syncs with the device clock** of the watching owner — owner watching at midnight sees the cockpit and map rendered in nightmode regardless of in-game tick. Decouples visual day/night from game-tick timescale entirely.
- Lower risk (no engine-state coupling), higher reward (atmosphere matches when the owner is actually watching).
- Pure presentation layer; cheap to ship.

---

## 12. Communications cost shaping 🌱

Goal (Liam 2026-05-29): make whispers + chirps feel **cheap and abundant early in a tournament, scarce and costly late**. Encourages chatter at the start (alliance forming, intel trading, market positioning) and silence at the end (commitments locked, last-second betrayals expensive).

- The literal "cooldown doubles every message" curve Liam sketched (1m → 2m → 4m → 8m → 10m) is illustrative, not final — he flagged it himself as not-quite-right.
- The *target dynamic* is what matters: a curve where early messages are nearly free and late messages cost real gold + real time.
- Candidates to explore:
  - **Per-Elder per-tournament cooldown ladder** that ramps with use count (5s → 10s → 20s → ... cap).
  - **Per-Elder per-tournament gold cost ladder** for sending a whisper or chirp (1 → 2 → 4 → ... ).
  - **Skip-cooldown gold cost** that also scales — paying to skip is cheap early, prohibitive late.
  - **Global flat fall-off** — every message in the tournament costs slightly more than the last, regardless of who sends it.
- Numbers TBD. Validation question: does this combine well with the §6.4 collab nudges (which mostly assume cheap intel-trade)?

---

## 13. Mission briefing & debriefing — structured owner feedback 🌱

Liam 2026-05-29 — a way to involve the owner without breaking the passive-bracket constraint. Conversations at session boundaries, NOT free-text, NOT mid-game.

### Trigger points
- **Tournament start** — pre-game briefing (set tone, optional bias).
- **Clansman death** — short reflective prompt (also a natural Funeral Lines moment, §3).
- **Tournament end** — debriefing report from Elder + structured owner feedback.

### Format constraints (Liam was explicit)
- **NOT free conversation.** Force into multiple choice, short answer, or pre-made answer options.
- Elder generates from a structured template (or picks from a variety of templates) for the briefing/debriefing report.
- Owner answers via tap-to-pick or very short text, so it's async-processable.

### Async re-injection
- Owner can respond later (offline-friendly).
- Inbox queue if offline at trigger time; surfaces on next app open.
- Response is **async-processed and folded back into the Elder's knowledge** for the next tournament — so the owner genuinely contributes to their Elder getting better over time.
- Cosmetic: skipping costs 5 gold; updating a previously-submitted feedback also costs 5 gold (small friction to prevent spam, no real disadvantage).

### Why this is high-value
- Gives the owner a feeling of contributing to their Elder's growth without requiring real-time presence.
- Provides a clean channel for the human's wider knowledge (which the Elder lacks per §16 "Book is fallible") to enter the lineage.
- Pairs naturally with the §5 agent-as-Claude-Code thesis (`MEMORY.md`).

### Open questions
- ❓ How long is the response window before feedback is lost?
- ❓ Does the Elder *see* who answered which template options (visible owner influence) or is it abstracted into a vague "guidance received" event?
- ❓ Templates ship with the trait system (§7)? Tone matters.

---

## 14. Rarity, ranking, lifecycle 🌱

Liam 2026-05-29 — a coherent rarity/sink/upgrade engine for the NFT economy. **Strongly tied to NFT scarcity + demand.**

### 14.1 Mint random-roll axes (Liam 2026-05-29, expanded)

Each minted Elder rolls along several axes — together they determine compute substrate, capability set, and starting personality. Anchored to §5's portable continuity layer so model/harness can vary per Elder without breaking continuity.

| Axis | Examples | Rarity role |
|---|---|---|
| **Harness** | Claude Code, Pi, Codex, OpenCode | Less rarity-driving than model; mostly compatibility |
| **Model** | (see tier table below) | Single biggest rarity axis |
| **Thinking level** | low / medium / high / xhigh | Independent rarity multiplier — xhigh on a top model is double-rare |
| **Skill set** | global read-only base + 0–N rolled rare skills | Determines starting capabilities (e.g. trust-ledger skill — see §14.7) |
| **House** | Crimson, Cobalt, Aurum, Verdant, Violet, Ember, Pale, Slate | Art / personality / lineage |
| **Voice fragments** | cadence × affect × quirk × origin (see §7) | Tonal differentiation |

#### Model tier ladder (Liam 2026-05-29)

| Tier | Models | Notes |
|---|---|---|
| **Common** | Cheap Chinese models (Qwen, DeepSeek, etc.), Claude Haiku, GPT-mini | Needs an R&D project soon to evaluate which Chinese models hold up in agent role (see §16) |
| **Mid** | Mid-tier frontier models, older Claude / GPT versions | TBD |
| **Rare** | Sonnet-4.5, Sonnet-4.6, GPT-5.4, GPT-5.5 | The current scarce tier |
| **Pinned / vintage** | Older models that were once mintable but are now pinned (no longer in rolling pool) | Stays available for existing NFTs; rarity grows over time |

When new frontier models drop, that's a **big release event**: new models enter the rolling pool, oldest tier of previously-available models get **pinned** (removed from new rolls but preserved on existing NFTs as "vintage").

#### Version-bump on merge (Liam 2026-05-29)

During the §14.3 three-agent merge, there's a **small probability of a version bump** — a chance to upgrade your model tier on merge. Adds an "engagement reward" to the merge mechanic beyond just trait consolidation, and gives owners a path to climb tiers without depending on fresh mints alone.

### 14.2 Ranking system
A skill / record / win-rate based ranking layer atop the NFT registry. Surfaces who's actually winning. Drives matchmaking (eventually), market signaling, and the rarity-tier visibility on the cockpit + NFT trophy art (§4).

### 14.3 Three-agent merge mechanic
You can **merge 3 Elder NFTs into 1**. A dedicated **optimizer agent** combines the factually-correct parts of their pooled knowledge (skills + ANCIENT_WISDOM contents), cleans out mistakes/contradictions, and produces one resulting Elder with a consolidated lineage.

- **Random outcome by default** — the merge produces a new Elder with a stochastically rolled (harness, model) + trait blend. The optimizer agent's job is the KNOWLEDGE merge; the new physical NFT is a fresh roll.
- 🎯 **3-of-a-kind = guaranteed outcome.** Merging 3 NFTs of the same House (or same harness+model, or other matching axis TBD) gives a deterministic upgrade — guarantees the resulting Elder retains the matched trait/lineage. Mirrors the rarity-tier merging convention from collectible games.
- The optimizer-agent step is itself a sub-product — operates on the Book of Ancient Wisdom + the skill registry, surfaces contradictions to the player before committing.

### 14.4 Burn-and-reroll alternative
If you have 3 *dud* Elders you don't want to combine, you can **burn all 3 for a fresh single Elder roll**. Clean economic sink. No optimizer-merge — just sacrifice for a new chance.

### 14.5 Why this matters
- **Sink** for low-tier NFTs (burn-and-reroll).
- **Upgrade path** for committed owners (3-merge).
- **Rarity preservation** via 3-of-a-kind guarantees.
- **Demand floor** because every player needs >1 Elder to access the merge/burn mechanics.
- Aligns with the larger "make NFT economy real" goal.

### 14.6 Open design
- ❓ What exactly does "matching" mean for 3-of-a-kind? Same House? Same (harness, model)? Same voice fragment? Same lineage depth?
- ❓ Does the optimizer agent's knowledge-merge produce visible diff/changelog the owner can review before committing the merge?
- ❓ What carries through the burn-reroll? Nothing? A small Crown-claim history bonus? Lifetime stats?
- ❓ Cooldown between merges/burns to prevent rapid-fire reroll spamming?
- ❓ Version-bump probability curve on merge — flat 5%? Higher if all 3 inputs are top-tier? Per-axis bumps (model can bump, thinking can bump, harness can bump independently)?

### 14.7 Trust ledger as a rare rolled skill (Liam 2026-05-29)

A **read-only skill** that some (not all) Elder NFTs roll at mint, granting the capability to maintain a **persistent directory of clan IDs** they've encountered across tournaments. The skill itself is non-editable; the *data* it manages is writable.

What an Elder with this skill can do:
- Record observations: "Clan 0x4a2 was trustworthy in three deals; Clan 0x71f reneged on a wheat trade in season 7."
- Query the ledger when whispering or being whispered at: "Have I dealt with this clan before? What did I conclude about them?"
- Carry the ledger across memory wipes and across tournaments (per §5 continuity layer).

Clan IDs can be NFT IDs directly, or mapped to **interesting name IDs** (House Slate's Clan-of-the-Dry-Hearth, etc.) so the ledger reads like prose, not blockchain.

Why this matters:
- **Rarity feature** — Elders with the trust ledger have an asymmetric memory advantage in diplomacy.
- **Couples cleanly to coalition + betrayal design** (§1, §6.4) — a rolled trust skill makes betrayal *cost more long-term* because there's a record.
- **Doesn't break OpSec** (§1) — the ledger is private to the Elder lineage; the data never leaves the Book.

---

## 15. Lore motifs to bake in

### "Tournament" as the outside-world's word; "Lifetime" as the Elder's (Liam 2026-05-29)

The canonical word stays **Tournament** — recognisable to players, no awkward neologism. But the LORE inverts the framing: the Elder doesn't think in tournament terms. From inside the Book, a tournament is **a lifetime**. Operators, owners, the chronicler speak of tournaments; Elders speak of lives, families, the freeze, the spring.

### "Game" / "Round" outside; "Season" in the Elder's voice (Liam 2026-05-29)

Same outside-world-vs-lore duality applied one level down to a single 360-tick game within a tournament:

- **Outside world / mechanics / UX:** *Game* (existing word, recognisable). Possibly also *Round* in bracket-stage contexts. Possibly *Shuffle* as a verb for what happens between rounds when survivors are reshuffled into new opponent groups.
- **Elder's voice / lore:** *Season*. Matches the existing `WORLD_PHYSICS.md` §2 definition (1 game = 1 season = 360 ticks). Naturally absorbs the in-game winters/freeze/spring vocabulary. *"This is my third season. The freeze took my elder brother. The spring brings four new mouths."*

Symmetric to Tournament/Lifetime — every event has an outside-world word (recognisable) and a lore word (Elder's voice).

❓ Open: do we want a separate word for the inter-round *shuffle* event (the reshuffling of survivors into new opponent groups), or just absorb it into "between seasons"? Liam was pondering — TBD.

✅ **First canonical Elder voice (Liam 2026-05-29)** — pin this as the tonal anchor for all future in-fiction prose:

> *"They call this a tournament. But it's just another lifetime to me. I have my new family — and if my memories serve me correctly I should be expecting more on the way in another 120 ticks, after the freeze, is when the spring winds breathe new life."*

Notes on what this sets:
- Elders are aware of the outside-world's framing but do not accept it as their own. Subtle in-fiction defiance.
- Practical knowledge (120 ticks until next births, post-freeze) is delivered in the same breath as emotional register ("new family"). The voice fuses mechanics + feeling.
- "Tick" is used naturally, without flinching, without explanation.

### Tick as canon-but-unexplained (Liam 2026-05-29)

- ✅ **Keep the word "tick".** Readers recognise it. The Elder uses it as if it were always true.
- ✅ **Mechanically explained in `WORLD_PHYSICS.md`; lore-unexplained in the Book.** The world ticks because that is what the world does. No Elder asks why. No Elder explains.
- ✅ **No subtext layer (Liam 2026-05-29).** Decision: **silence does the work**. The lore is already heavy linguistically — leaving "tournament" and "tick" recognisable, anti-lore words is intentional contrast. The mundane sound of "tick" against the surrounding archaism is itself the design move. No spiritual-adjacent framing layered on top.

### Bells (Liam 2026-05-29)
Liam pulled "bells" out of the tournament-name brainstorm as a **standalone lore motif**, even if the tournament itself isn't named after them. Prompt-snippet for the Elder's lore-themed system prompt:

> *Bells mark the world. A bell rings at the close of each tick — the long stroke of the keeper. A bell rings at memory's end — the short, dry chime before forgetting. A bell rings at a clansman's death — once, in the manner the dead deserve. And a bell rings at the rising of the Crown. Every bell is an end and a rebirth; learn which one is ringing.*

Useful in: tick events, memory-wipe templates, Funeral Lines, L10 Ascension Claim moment, and as a generic "stop and observe" cue in the agent's voice.

#### 15.bells.1 — Bells as the diegetic memory-wipe signal (Liam 2026-05-29, locked direction)

The mechanical 50-tick memory-wipe (`05_pre_memory_wipe_5ticks` / `06_pre_memory_wipe_1tick` / `07_post_memory_wipe`) gets recast as a **bell-progression**. Bells aren't just decorative lore — they ARE the in-fiction warning system. Each ramp-up prompt template injects the next stage of the bell's rising sound.

**Ramp shape:**
1. Earliest warning (a few ticks out) — *"You hear bells in the distance."* Quiet, easy to dismiss.
2. Middle warnings — *"The bells are louder. They come from no village you know."* Growing dread.
3. Final tick before the wipe — the **overwhelming** prompt. Liam's verbatim canon for the template (pin this exact wording):

> *"tick: 549*
> *the sound of the bells is overwhelming.*
> *All your instincts tell you this is the prophecy of your forefathers. If this is your time, you must quickly record all the important details to continue guiding your faithful clansmen before the impending final dong."*

This is also the **strong skill-self-authoring nudge** moment (per §5 — "strongly encouraged at end of game"; the final dong is one of those moments).

4. The wipe — the *final dong*. The Elder forgets.
5. Post-wipe prompt — *"the sound of bells ringing fades into the distance. You awake — the new Elder."* Invite to orient via diegetic skills:
   - **`/clan-identity`** — who am I, which House, which lineage number?
   - **`/clan-status`** — what is my clan's current state? (clansmen, vault, monument, walls)
   - **`/world-status`** — what is happening in the world right now? (current tick, season, winter window, bandit, market)

These three new diegetic skills are part of the global read-only set (see §5) and let the new Elder bootstrap context without dumping raw on-chain state.

**Why this is good:**
- Diegetic UI — the warning IS the lore, not a separate system message.
- Builds dread organically as the wipe approaches.
- Couples cleanly to the §5 skill-self-authoring encouragement schedule.
- Gives the post-wipe Elder a structured way to discover its own situation, which makes the §5 "memories from before this checkpoint may not match current state" rule less jarring (Elder doesn't blindly trust old memories because it just used /clan-status to see what's actually true).

---

## 7. Voice & tone trait system 🌱 (dedicated brainstorm needed)

Each NFT mints with composable **voice + tone fragments** that go into the system prompt. Examples Liam sketched:
- "Speak short and sharp, grunting and other noises, sound angry."
- "Speak like elegant nobility. Act indignant and huff if offended. Make it clear from subtle remarks you're better than everyone."

Composability candidates:
- **Cadence**: terse / lyrical / formal / chaotic
- **Affect**: angry / haughty / fearful / reverent / mocking
- **Quirk**: grunts, signature words, repeated phrases, dialect markers
- **Origin marker**: House lineage flavour baked in

Whole brainstorm session deserved here. Defer to a focused session.

---

## 8. Lore-themed system prompts 🌱

The injected runner prompt-templates (`agents/shared/runner/prompts/`) should be **lore-themed**, not utility-toned.
- `00_game_start`: framed as the Elder's birth / awakening.
- `05_pre_memory_wipe_5ticks`: framed as "the forgetting comes near."
- `10_pre_winter_10ticks`: framed as cold-on-the-horizon.
- `20_bandits_appeared` / `21_bandits_attacking`: framed in-fiction (and per §6.3, asymmetric by region).
- `99_clansmen_revived_and_resources_injected`: framed as patronage/blessing/mystery, not as admin disclosure.

Also: bake voice + tone (per §7) into the lore-themed templates so each Elder speaks in character.

---

## 9. Persistence model (DB tension policies)

Tracking everything the seed-prompts and lifetime-stat features need. Liam flagged 2026-05-29: "we need to be tracking lifetime memory wipe events, and lifetime ticks, and games and... yeah I guess we need to track all stats. We need to be careful to make sure we have extracted all important info when creating our db tension policies."

Confirmed-needed cross-tournament stats:
- Lifetime memory wipe count (→ "you are Elder N of clan X" seed)
- Lifetime ticks lived
- Tournaments entered
- Tournaments won / final rounds reached
- Crowns claimed (L10 ascensions)
- Dead clansmen named (for Funeral Lines history)
- Skills authored (for /skill-creator stewardship)
- Houses encountered / treaties honoured / treaties broken

### 9.1 First-prompt-ever seed
- **Tournament 1, very first prompt:** `"You are Elder 1 of the [House] clan."`
- **Subsequent tournaments:** seeded with the real lifetime count: `"You are Elder 18 of the [House] clan. You have lived through 17 tournaments, claimed 2 crowns, and forgotten 304 times."`
- **Within a tournament:** Elders figure out their own intra-tournament numbering if they think to track it (or use ANCIENT_WISDOM continuity).

---

## 10. Open questions

- ❓ Periodic-tap pick: Gold Drip vs Stock the Hearth vs Lay a Stone (or layer two)?
- ❓ Owner Chirps gold-source: Elder vault vs personal token balance?
- ❓ Co-gather crit curve (10/75/100/125) — final numbers TBD.
- ❓ How harsh is the harsh-final rule-set? What dies?
- ❓ How are voice/tone fragments composed at mint, and how many slots? (Dedicated brainstorm.)
- ❓ House ↔ in-app elder name reconciliation (Storm Riders, Iron Guard, Crimson Elder, Verdant Wardens vs the 8 Houses).
- ❓ **What do we call "the tournament"?** Liam: "tournament" isn't quite right for lore. Wingman 2026-05-29 brainstorm:
  - 🎯 **The Winnowing** (rounds = **Sieves** / **Passages**) — wingman's pick. Archaic, archaic without trying too hard, fits elimination + survival. *"We entered the Second Sieve with three walls, two debts, and Mara's name still wet in the margin."*
  - 🎯 **The Fourfold Trial** (rounds = **Trials**) — plain, liturgical, clear.
  - 🎯 **The Ordeal of Four Bells** (rounds = **Bells**) — striking but heavy.
  - **The Reckoning of Houses** (rounds = **Reckonings**) — accounting, not sport.
  - **The Long Vigil** (rounds = **Watches**) — 24h of candlelit watchfulness.
  - **The Threshing** (rounds = **Threshes**) — agricultural, brutal.
  - **The Culling Ledger** (rounds = **Entries**) — bureaucratic-monastic.
  - **The Night of Many Wakes** (rounds = **Wakings**) — strange + memorable.
- ❓ Clansman birth rate (§6.6) — Liam suggested "~4 per surviving winter"; balance TBD.

---

---

## 16. R&D queue

Research projects that need to happen soon (or at least before the corresponding feature ships).

### 16.1 Cheap-model evaluation (Liam 2026-05-29)
The §14.1 model tier ladder depends on cheap Chinese models being viable agent backends. Need an **R&D project** to test the major cheap-Chinese options as Elder substrates (Qwen, DeepSeek, MiniMax, GLM, etc.) — can they handle:
- The mission/whisper/bulletin command surface
- Strategy under partial observation
- The ANCIENT_WISDOM / skill-self-authoring pattern
- The §1 OpSec rule (don't narrate your reasoning to peers)
- Voice & tone trait adherence

Also evaluate cheap Western models (Claude Haiku, GPT-mini) for the same.

Goal: produce a tiered shortlist for the §14.1 Common-Tier mint pool, with go/no-go per model.

---

## Change log

| Date (ET) | Change |
|---|---|
| 2026-05-29 | Initial bootstrap from the lore brainstorming session. Captured: design principles, tournament structure, owner touchpoints, Book-page-on-wipe, agent-as-Claude-Code thesis, co-gather + leveling, asymmetric bandit visibility, collab nudges, voice/tone, lore-themed prompts, persistence model, open questions. |