a research devlog · est. 2026 · solo build

INSCRYPTION· KCM · AI

An RL agent for Kaycee's Mod, in build. Validated Rust simulator complete and surfacing actionable play patterns; warm-start training data generated; agent training next. The most transferable artefact so far: a framework for parallel agentic workflows.

Phase 1.6 — L6a four-boss validationPhase 1.8 warm-start training run · complete$0 / $80spent / cap

Algorithm: AlphaZero-style IS-MCTS + distilled neural network, fine-tuned with PPO self-play
Simulator: Rust + PyO3 bindings, validated against the live game via a BepInEx bridge mod
Repo: github.com/4abandoment/inscryption-kcm-ai

About the game

Why Inscryption — and specifically Kaycee's Mod — is the target.

Inscryption (2021) is a roguelike deck-building card game by Daniel Mullins Games, published by Devolver Digital. It won Game of the Year at the 2022 Game Developers Choice Awards and the Seumas McNally Grand Prize at the Independent Games Festival, selling over a million copies in its first three months.

The game presents itself as found footage — a mysterious floppy disk containing a tabletop card game played against a shadowy figure named Leshy in a dimly lit cabin. Act 1 is a self-contained roguelike: build a deck of creatures, navigate a procedurally generated map, and fight through three bosses before a climactic encounter with Leshy himself.

I.aHow the card game works — grid, scale, blood, bones, sigils

Combat takes place on a 3×4 grid — four slots per side. The goal is to tip a weighing scale by five damage points in your favour. Cards attack the opposing slot every turn; excess damage carries over to the scale.

Resource costs

Blood. Sacrifice your own creatures to pay the blood cost of a new card. A Squirrel (0/2, no abilities) is always free fodder.
Bones. Generated when your creatures die. Bone-cost cards reward aggressive play and death loops.

Sigils

Each card carries one or more sigils — passive or triggered abilities. Examples: Flying (only blocked by Reach), Bifurcated Strike (attacks two slots), Touch of Death (kills any card in one hit), Unkillable (returns to hand on death). Act 1 has 48 distinct sigils; their interactions — totems, evolved cards, cardmod stacks — are where most of the strategic depth lives.

I.bRun structure — branching map, candles, node types

Between combats players navigate a branching map across three regions. Death costs a candle; you start with two lives.

Combat. Standard encounter.
Card reward. Choose one of three cards to add to your deck.
Campfire. Upgrade a card's ATK or HP. Repeated attempts risk failure.
Trader / Trapper. Trade pelts (earned from kills) for cards.
Shop. Buy or sell cards and single-use items (Hammer, Pliers, Scissors).
Mycologists. Merge two cards into one.
Totem builder. Attach a sigil to an entire creature tribe for the rest of the run.

I.cKaycee's Mod — the scope target

Kaycee's Mod (KCM) is an official post-release expansion bundled free with Inscryption, named after the game's lead programmer. It strips away the metafictional narrative of Act 1 and turns the card game into a pure roguelike challenge mode — the game as a game, not as a horror story.

What KCM adds

Challenge modifiers. Unlocked progressively. Single Candle, Tipped Scales, No Hammer, Brutal Hunt, Bones Bummer, Boss Tougher, and more — stackable, with a difficulty score.
Full Act 1 card pool. All cards available in the reward pool, including rare/powerful cards uncommon in a single Act 1 playthrough.
All four bosses. Prospector, Angler, Trapper/Trader, and a multi-phase final encounter with Leshy (Moon card and starvation overrides).
Consistent rules. Unlike Act 1, narrative events do not override game logic. KCM is deterministic.

KCM with all challenges active simultaneously, full card pool, all four bosses, is the scope target for this project.

I.dWhy Inscryption is a strong research platform

Decompilable, auditable rules

Inscryption is a Unity game shipped with C# managed code. Game logic lives in Assembly-CSharp.dll and decompiles cleanly with ilspycmd / dnSpy. For this project, 2,098 classes were decompiled and 12 key behaviour classes were read in full. The simulator is built from the source of truth — not reverse-engineered from gameplay observations. Every hardcoded exception, every sigil-interaction edge case, every boss AI decision is documented with a DLL citation. That fidelity is unusual in game-AI research and eliminates an entire class of simulator bugs.

Active modding community

Inscryption has a large, technically capable modding community centred on InscryptionAPI — a BepInEx-based framework that exposes hooks into the game's event loop. This project uses the same infrastructure: a BepInEx validation mod reads live game state mid-run, exports it to JSON, and is used to verify that the Rust simulator produces identical results to the real game. CommunicationMod served as the template for the validation bridge.

Tractable complexity at the right scale

Not too simple — 126 cards, 48 sigils, stochastic draws, branching map, multi-phase bosses, 10+ challenge combinations, and tightly coupled deck-building / combat decisions. No known AI plays it at strong human level. Not too expensive — a full run simulates in milliseconds on laptop CPU; MCTS data generation needs zero cloud spend, GPU fine-tuning is estimated at $30–50 total. Coupled decisions — the key unsolved problem in Inscryption is that card-reward and combat decisions are bidirectionally coupled. Slay-the-Spire RL work reportedly failed to discover this coupling when treating the two separately. This project models them jointly via MCTS over full runs with a single shared-trunk neural network.

The approach

MCTS over full runs as the data engine; a small distilled NN as the policy; PPO self-play as the polish.

I'm building this because nobody has done it. There's no published RL agent for Kaycee's Mod, and the game's coupling of meta-game and combat decisions broke the only large-budget Slay-the-Spire attempt I'm aware of.

The wager: MCTS over full runs on a laptop CPU is free; it generates clean training signal that handles the coupling correctly. Distill into a small NN, fine-tune with PPO self-play on a single spot GPU. Total budget cap: $80.

Goal isn't "a bot that wins." Goal is novel play patterns and calibrated coaching — a tool that says "play Wolf in slot 2, opponent queues Stoat opposite, scale tips +2" and is right 70% of the time when it says 70%.

Algorithm

AlphaZero-style

IS-MCTS over full runs — deck-building and combat as one decision space — distilled into a small shared-trunk NN, then fine-tuned with PPO self-play.

Simulator

Rust + PyO3

Built from the decompiled DLL, not from gameplay observation. A BepInEx bridge mod validates Rust against the live game on every level transition.

Budget

$80 hard cap

$0 through Phase 3 (laptop CPU). $30–50 of GPU spot time for PPO. $2–10 for warm-start LLM calls. Buffer covers overruns.

III

Project phases

Six phases · $0 → $80 · roughly twenty weeks. Click any phase for detail; click a sub-phase for its key work and acceptance gate.

P0Mechanics Audit P1Simulator P1.8Warm-Start BC P2MCTS Data P3Distillation P4Self-Play P5Coaching

Phase 0

Mechanics Audit & Architecture

$0complete · 2026-04-20L0complete

Establish ground-truth rule documentation and lock irreversible architecture decisions before any code is written.

Detail · key work, gate, outputs

Hardware

Laptop

Outputs

docs/rule-reference.md — 1,247 lines, 48 sigils with DLL citations
docs/MECHANICS-AUDIT.md, MECHANICS-COVERAGE-MATRIX.md, MECHANICS-VALIDATION-SUMMARY.md
docs/TRAINING-ARCHITECTURE-DECISIONS.md (8 ADRs)
sim/ (L0 simulator scaffold)

Phase gate. L0 acceptance ≥ 80%. BepInEx bridge validated vs live game. All [VERIFY] markers in rule-reference.md resolved.

Sub-phases

Phase 0.1DLL Decompilationcomplete

Full decompile of Assembly-CSharp.dll (2,098 classes) via ilspycmd. Grep passes for hardcoded exceptions, persistent state mutations, KCM-specific code paths.

Phase 0.2Mechanics Coverage Auditcomplete

Systematic audit of all 52 sigils (48 documented with DLL citations; 4 data-driven). 23 hardcoded exceptions identified and mapped to simulator rules.

Phase 0.3Architecture Decision Recordscomplete

Lock irreversible decisions: simulator language (Rust), algorithm (IS-MCTS), card encoding (feature-vector only), policy structure (hierarchical), warm-start strategy, GPU provisioning.

Phase 0.4L0 Simulatorcomplete

Fixed deck, scripted opponent, ~8 sigils, PyO3 bindings. Prove the Rust/PyO3 stack before full implementation.

Phase 1

Simulator — L1 → L6b

$06–10 weeks totalL1 → L6bin progress

Complete, bridge-validated Rust simulator from L1 through L6b — full Kaycee's Mod with all challenges active.

Detail · key work, gate, outputs

Hardware

Laptop

Outputs

sim/kcm-sim/src/ (complete Rust simulator)
sim/kcm-sim/src/card_registry.json (126 cards)
acceptance gate scripts (l1_acceptance_gate.py, l4_acceptance_gate.py, l6b_acceptance_gate.py)

Phase gate. All Phase 1.7 sim-completeness criteria met (see L6b gate). Hard block on Phase 2.

Sub-phases

Phase 1.1L1Single combat with draws1–2 weekscomplete

Real deck draws, heuristic opponent, 6 new sigils — makes isolated combat stochastic and meaningful.

Key work

init_l1_game() and draw_phase()
Heuristic opponent
Sigils: SharpQuills, LooseTail, Unkillable, Strafe, Sniper, CorpseEater
l1_acceptance_gate.py

Gate. All Rust unit tests pass; each L1 sigil has passing test; greedy 30–80% win rate vs heuristic; bridge ≥ 5 live states; throughput ≥ 200 eps/sec; L0 regression ≥ 80%.

Phase 1.2L2Multi-fight runs1–2 weekscomplete

Persistent deck state, candle/lives system, side deck — models a complete multi-fight run.

Key work

RunState struct (deck, candles, bones, items, seed)
Combat-chain persistence
CardModificationInfo stack
Side deck (unlimited Squirrel)
Starvation mechanic
2 → 1 → 0 candle death sequence

Gate. Greedy agent completes 10-fight run > 50% of attempts.

Phase 1.3L3Card rewards & linear progression~1 weekcomplete

Deck growth via card rewards and evolution — enables a complete linear run.

Key work

CardChoiceNode (3 cards from pool, pick 1)
Full card-registry.json for reward-eligible cards
Evolve mechanic (Stoat → Stinkbug)
Ice Cube spawn on death
Linear run skeleton (fight → reward → fight)

Gate. Agent completes L3 linear run end-to-end without crash.

Phase 1.4aL4aBranching map · combat nodes~1 weekcomplete

Seeded map generation with branching node selection — run-path decisions become real.

Key work

Map generation (3 regions × ~14 nodes, branching, seeded RNG)
ChooseMapNode action type
completed_nodes BitSet on RunState
Node traversal state machine

Gate. Agent navigates map to each region's boss node without crash.

Phase 1.4bL4bNon-combat nodes~1 weekcomplete

Campfire, shop, and trader nodes with full decision surfaces.

Key work

CampfireNode (ATK/HP upgrade with failure risk)
ShopNode (buy/sell cards and items)
TraderNode (pelt trading for cards)
ActivateItem action in combat

Gate. Agent can reach each boss node via map and enter combat from each node type.

Phase 1.4cL4cMeta systems · bones, items, totems~1 weekcomplete

Bones economy, usable items in combat, totem builder node — largest action-space expansion in Phase 1.

Key work

Combat items: Hammer, Pliers, Scissors, Moose Antlers, Steak Dagger, remainder
Totem builder node (tribe + sigil → whole-run buff)
Bones wiring across run scope

Gate. Agent can navigate map, activate items in combat, and reach each boss.

Phase 1.5aL5aCore card subset · ~30 cards, ~10 sigils~4–5 dayscomplete

Validate the full pipeline with a representative card subset before full-pool complexity.

Key work

Mycologists merge node
Bone Lord node
Totem builder functional with subset
Hardcoded exceptions for subset cards

Gate. 5,000 random-action runs without crash.

Phase 1.5bL5bFull card pool · 126 cards, 48 sigils1–2 weekscomplete

All 126 cards and all 48 sigils implemented, tested, and interaction-validated.

Key work

Full card_registry.json loaded
All 48 sigils with unit tests
Stat icons: Ant count, Hand size, Bell proximity, Bones held, Mirror
Hardcoded exceptions: Daus slot pattern, HydraEgg transform, Ijiraq repulsive, Shapeshifter, Cat counter
All 27 CardModificationInfo fields handled

Gate. All 48 sigil unit tests pass; 5,000 random-action full runs without crash.

Phase 1.6L6aAll four bosses1–2 weeksin progress

All four boss encounters with phase transitions and boss-specific AI. Currently validating.

Key work

Prospector — Phase 1 standard; Phase 2 PickAxe + Pelt reward
Angler — Phase 1 standard; Phase 2 BaitBucket + Hook theft
Trapper/Trader — Phase 1 standard; Phase 2 trader mid-fight
Leshy — Phase 1 standard; Phase 2 Death card; Phase 3 Moon + starvation
Boss AI scripted per phase
Integration tests for all boss phase transitions

Gate. Greedy agent completes a full L6a run (all bosses) > 20% of attempts; bridge ≥ 5 live states per boss phase.

Runs in parallel with Phase 1.8.

Phase 1.7L6bChallenge modifiers — full KCM~1 weekfuture

Full challenge modifier system — all KCM challenges active simultaneously. This is the hard block on Phase 2.

Key work

ChallengeFlags bitmask on RunState
Per-challenge implementations: Tipped Scales, Single Candle, No Hammer, No Sniper, Brutal Hunt, Tougher Hunt, Bones Bummer, Boss Tougher, Limited Lives, remainder
apply_challenges() PyO3 binding
20 curated challenge combination tests
Zobrist state hashing
Throughput verification ≥ 100 full runs/sec

Hard block on Phase 2 — all of the following must pass:

All challenges active; greedy runs complete without crash
All 48 sigil unit tests pass
Bridge round-trip ≥ 20 distinct live states
Full-run determinism: 1,000/1,000 same-seed runs produce identical outcomes
Heuristic vs heuristic at L6b: win rate 30–70%
All prior acceptance gates pass (L0, L1, L4, L6b)
All boss phases end-to-end
20 curated challenge combinations tested
Throughput ≥ 100 full runs/sec
10,000 random-action fuzz: zero panics, zero infinite loops

P1.8

Phase 1.8 · optional · parallel: 1.6, 2

Warm-Start Behavioural Cloning

$2–103–5 days · run completeL6bcomplete

Pre-generate an initial policy baseline via cheap LLM API to reduce Phase 2 MCTS warm-up time. The training run is complete.

Detail · key work, gate, outputs

Hardware

Laptop + OpenRouter API

Acceptance gate. Cheap-model policy accuracy > 40% on held-out states (baseline greedy ~30%); value MSE < 0.20; dataset importable into Phase 2.

Outputs

guard.py with resolve_provider() routing to OpenRouter (Mistral/Llama, ~$0.14 / 1M tokens)
Synthetic rollout pipeline (random/greedy games → LLM action ranking → soft distribution)
Dataset target: 500k positions minimum / 5M positions primary

Phase 2

MCTS Data Generation

$02–4 weeks of overnight runsL6bfuture

Generate a large dataset of (game state, action distribution, win probability) triples via pure MCTS on laptop CPU.

Detail · key work, gate, outputs

Hardware

Laptop CPU (rayon, 4–8 threads)

Prerequisite

All Phase 1.7 / L6b sim-completeness criteria.

Acceptance gate. MCTS (200 sims) beats greedy 70%+ in 500 combat episodes; full-run win rate > 25% over 500 runs; value estimates calibrated.

Configuration

Rollout policy

heuristic-greedy

Sims / move

200–800

Selection

PUCT + progressive widening

Transposition table

Zobrist hashing

IS determinization k

Data targets

Tier	Positions	Timeline
minimum viable	1M	1–2 nights
primary target	10M	1–2 weeks
extended (only if value head struggles)	50M+	extended

Phase 3

Distillation — small NN from MCTS labels

$0–153–5 daysL6bfuture

Train a small neural network to approximate the MCTS action distribution and win probability.

Detail · key work, gate, outputs

Hardware

Laptop or modest GPU

Acceptance gate. Value MSE < 0.05; policy top-1 accuracy > 50%; live win rate vs heuristic > 30%; if not met after 5 epochs return to Phase 2.

NN architecture

Input

~500–800 floats

Trunk

3-layer MLP or small transformer, ~300K–500K parameters

Policy head

softmax over legal actions

Value head

scalar win probability

Encoding

feature-vector only — attack/health/cost/sigil bitmask/tribe one-hot · never card-ID lookup

Training

Loss

CE(policy) + MSE(value)

Optimiser

Adam · lr 1e-3 cosine decay

Batch size

512–2,048

Phase 4

Self-Play Fine-Tuning · PPO

$30–502–3 days wall-clockL6bfuture

Improve the distilled model via PPO self-play to discover patterns beyond heuristic-rollout MCTS.

Detail · key work, gate, outputs

Hardware

GPU spot · Vast.ai / RunPod RTX 4090 (~$0.35/hr)

Cost cap

$80 total including buffer

Algorithm

PPO + past-checkpoint opponent pool + entropy bonus 0.01 + 128-unit LSTM

Acceptance gate. Win rate > 60% vs heuristic at L6b over 500 eval runs; OR $80 budget cap; OR 3 consecutive plateau evaluations.

Curriculum

Stage	Configuration	Condition	Target
4.1	L6a (no challenges)	baseline	win rate > 30%
4.2	+Tipped Scales	after 4.1 gate	win rate > 25%
4.3	+Single Candle	after 4.2 gate	win rate > 20%
4.4	All challenges (L6b)	after 4.3 gate	budget exhausted or plateau

Operational

Checkpoint interval

15 min

Eval interval

1 hr

Eval episodes

500

Phase 5

Inference & Coaching Deployment

$0~1 weekL6bfuture

Deploy trained agent as an in-game coaching tool with real-time win probability, best-move suggestions, and on-demand deep analysis.

Detail · key work, gate, outputs

Hardware

Laptop

Integration

BepInEx bridge mod → JSON export → Python inference script → CLI or HUD injection.

Acceptance gate. Win probability calibrated within ±10% of actual win rate on 200 held-out live positions; deep analysis latency < 5s on laptop.

Inference modes

Mode	Mechanism	Latency
Real-time overlay	NN value head	< 10 ms
Best-move suggestion	NN policy top-3	< 100 ms
Deep analysis	MCTS · 200–800 sims	1–5 s
Counterfactual	NN value on hypothetical	< 100 ms

LLM explanation layer (MCTS trace → natural language) — out of scope but architecturally enabled.

Scope ladder · L0 → L6b

Each rung is its own acceptance gate. Lit candles are passed; the flickering candle is L6a — currently validating.

Single combat · fixed deck

Real draws · heuristic opponent

P1.1

Multi-fight runs · candles · side deck

P1.2

Card rewards · evolve · Ice Cube

P1.3

L4a

Branching map · combat nodes

P1.4a

L4b

Campfire · shop · trader

P1.4b

L4c

Bones · items · totems

P1.4c

L5a

~30 cards · ~10 sigils subset

P1.5a

L5b

Full pool · 126 cards · 48 sigils

P1.5b

L6a

All four bosses · multi-phase

P1.6

L6b

All challenges stacked

P1.7

Single combat · fixed deckPhase 0

Gate. 100% win-rate · 100/100 episodes

complete

Real draws · heuristic opponentPhase 1.1

Gate. Greedy 30–80% vs heuristic

complete

Multi-fight runs · candles · side deckPhase 1.2

Gate. 10-fight run > 50%

complete

Card rewards · evolve · Ice CubePhase 1.3

Gate. Linear run end-to-end

complete

L4a

Branching map · combat nodesPhase 1.4a

Gate. Reach each region's boss

complete

L4b

Campfire · shop · traderPhase 1.4b

Gate. Each node type → combat

complete

L4c

Bones · items · totemsPhase 1.4c

Gate. Items active in combat

complete

L5a

~30 cards · ~10 sigils subsetPhase 1.5a

Gate. 5k random-action runs · no crash

complete

L5b

Full pool · 126 cards · 48 sigilsPhase 1.5b

Gate. All sigil tests · 5k fuzz

complete

L6a

All four bosses · multi-phasePhase 1.6

Gate. Greedy full run > 20%

in progress

L6b

All challenges stackedPhase 1.7

Gate. Calibrated 30–70% h-vs-h · hard block on Phase 2

future

Architecture

Sim ↔ MCTS ↔ Distill ↔ Self-play. The bridge mod keeps the simulator honest; warm-start BC seeds the policy.

Acceptance gates

Every level has a gate. The L6b composite gate is the hard block on Phase 2.

L0 acceptance

100/100 eps · 24 turns deterministic · ~410 eps/sec

passing

L1 acceptance

Loose Tail · Strafe · Sniper · 30–80% win-rate band hit

passing

L2 acceptance

10-fight greedy run > 50%

passing

L3 acceptance

Linear run end-to-end · no crash

passing

L4 acceptance (a/b/c)

Map + non-combat + items integrated

passing

L5 acceptance (a/b)

All 48 sigil tests pass · 5k fuzz clean

passing

BepInEx bridge round-trip

≥ 5 live states / boss phase under L6a · in progress

in progress

L6a

L6a acceptance

All 4 bosses · greedy full run > 20% target

in progress

WARM

Warm-Start BC dataset

Training run complete · accuracy > 40% on held-out · value MSE < 0.20

passing

FUZZ

10k random-action fuzz

L6b gate · zero panics · zero infinite loops

todo

ZOB

Zobrist state hashing

Required for MCTS transposition · L6b gate

todo

DET

Full-run determinism

1,000/1,000 same-seed identical outcomes · L6b gate

todo

VII

Mechanics coverage

47 of 48 sigils implemented and unit-tested. Audited line-by-line against Assembly-CSharp.dll.

Combat

14 of 14 implemented

100%

Bifurcated Strike · Trifurcated · Sharp Quills · Sniper · Mighty Leap

Spawn

7 of 7 implemented

100%

Loose Tail · Bone King · Hoarder · Many Lives · Fledgling

Stat icons

5 of 5 implemented

100%

Ant count · Hand size · Bell prox. · Bones held · Mirror

Tribe

6 of 6 implemented

100%

Bee · Avian · Reptile · Canine · Insect

Hardcoded

6 of 7 implemented

86%

Hodag · Ijiraq · HydraEgg · Daus · Shapeshifter

Passive

9 of 9 implemented

100%

Buff Neighbours · Cultist · Flying · Reach · Worthy Sacrifice

VIII

Budget

Hard cap $80. Spent so far: $0. Most of the budget is reserved for a single GPU spot run during PPO self-play.

Cloud spend · running total$0 / $80

Phase 0 — Mechanics audit

$0–$0

Phase 1 — Simulator (1.1 → 1.7)

$0–$0

Phase 1.8 — Warm-start BC

$2–$10

Phase 2 — MCTS data gen

$0–$0

Phase 3 — Distillation

$0–$15

Phase 4 — Self-play PPO (GPU)

$30–$50

Phase 5 — Coaching

$0–$0

Buffer

$5–$15

Up next & references

Active work this week and the documents this devlog draws from.

Up next · this week

Leshy phase 3 — Moon + starvation override — boss/leshy.rs · 2d
Bridge round-trip · Prospector P2 (PickAxe) — ValidationBridge · 1d
Bridge round-trip · Angler P2 (Hook theft) — ValidationBridge · 0.5d
L6a integration tests — all phase transitions — tests/l6a_bosses.rs · 2d
Gate script — l6a_acceptance_gate.py — scripts/ · 1d

Reading list