Born / Writer programmeTechnical plan · RC0Research

A model that can hold a scene.

Born Writer is a focused attempt to make a 3B-active local model write, remember, react, and roleplay far above its running cost. Not “helpful assistant, but more poetic.” A real story engine.

Target system / BW-35A3One core · four faculties

SELECTED FOUNDATION

Qwen3.635B total / 3B activated
W

World memory

D

Story direction

C

Character performance

L

Line instinct

262,144 native contextApache-2.0Q4 / Q5 target
01The aim

Small enough to own. Good enough to disappear into.

The objective is not benchmark-shaped eloquence. It is sustained belief: a model that understands the scene, preserves the player’s freedom, inhabits its cast, and writes the next turn with intent.

01

The scene remembers

A glass moved three turns ago stays moved. A lie is known only by the people who heard it. Consequences survive the next completion.

02

Characters push back

They have private wants, partial information, habits, and emotional momentum. They do not collapse into agreeable assistant voices.

03

The player remains the player

The model controls its world and cast, not the user’s thoughts, choices, dialogue, or body. Agency is measured turn by turn.

04

Something changes

Every response earns its length through revelation, pressure, choice, consequence, intimacy, reversal, or a useful breath before one.

05

The prose has a pulse

Sentence shape, silence, vocabulary, humour, mess, and specificity follow the scene and speaker—not a universal polished house style.

06

Its range is genuinely broad

Quiet realism, comedy, romance, horror, dark fiction, and consenting-adult scenes are trained as real registers, not bolted-on prompts.

The moat is not one magical dataset. It is the loop: stateful episodes → competing continuations → reactive judgment → human taste → executable feedback → on-policy distillation.

BASE-MODEL DECISION

Qwen/Qwen3.6-35B-A3B

The 27B dense model is a strong ceiling, but it is a hard local product. The 35B-A3B stores roughly 35B parameters while activating about 3B per token: a better bet for responsive, consumer-scale inference, provided the tuning and quantisation stack survives its hybrid MoE architecture.

Official model card
Foundation
Post-trained Qwen3.6; exact revision pinned per run
Architecture
35B total / 3B active / 256 experts / 8 routed + 1 shared
Context
262,144 tokens native; train short-first, evaluate long
Licence
Apache-2.0 open weights
Why it can work
Agentic substrate, long context, local-active footprint
What can break
Expert-target LoRA support, storage, KV cache, quant quality
02System architecture

One voice on the outside. A writers’ room underneath.

Training decomposes the task so every failure has an owner. Release recomposes the skills into one fast default model, with explicit tools only when they improve long-horizon state.

INPUTUser turn + cards + lore + prior state
W01

World memory

Tracks location, time, inventory, lore, secrets, planted details, and what each character can actually know.

State invariants + 20 / 50 / 100-turn recall
D02

Story direction

Reads the current dramatic pressure, opens or closes threads, and advances the scene without forcing a prewritten plot.

Plot movement + callback + tension rubrics
C03

Character performance

Maintains distinct motives, diction, emotional residue, relationships, and plausible reactions while preserving player agency.

CoSER + RPEval + blind character-ID tests
L04

Line instinct

Cuts generic model cadence, varies rhythm, chooses concrete detail, protects subtext, and edits without sanding away the voice.

Humanizer audit + editor preference arena
OUTPUTVisible response

+ validated state delta

+ open-thread updates

VISIBLE CONTRACT

The reader gets story, not machinery.

The response is clean prose or dialogue. Planning labels, reward traces, judge notes, and model reasoning never leak into the scene.

STATE CONTRACT

The runtime gets a typed delta.

In agentic mode, the model proposes compact state operations. BornWorldEnv validates them before anything becomes canonical.

MEMORY CONTRACT

Context is curated, not endlessly appended.

Recent turns, immutable facts, active threads, character-local knowledge, and episodic summaries occupy separate memory lanes.

03Teacher foundry

Use frontier models as specialists, not as a single synthetic oracle.

Each model receives a job matched to its strengths. We retain candidates, rejections, critiques, evidence spans, judge disagreement, and state transitions—not just the polished final answer.

MODEL / ROLEWHAT IT CONTRIBUTESROUTE
tencent/Hy3World architect

Long-context world bibles, constraint-rich scenario states, causal checks, tool trajectories, and contradiction critiques.

Open-weight / hosted

Candidate scene structures, conflict escalation, darker narrative pressure, reversals, and multi-specialist response drafts.

Aion API
aion-labs/aion-3.0-miniBranch generator

High-volume alternate continuations and deliberate branch forks: advance, reveal, refuse, complicate, soften, or misdirect.

Aion API

Expressive multi-turn dialogue, character-card conditioning, example-message imitation, conversational pacing, and emotional reactions.

MiniMax / OpenRouter

Pairwise scene judgment, character and subtext critique, contradiction discovery, rubric repair, and difficult-case adjudication.

Claude API
openai/gpt-5.6-lunaReactive judge B

Independent high-volume ranking, evidence-span extraction, failure classification, and disagreement discovery.

OpenAI API
Born HumanizerLine editor

A first-party adaptation of the 29-pattern Humanizer workflow: voice calibration, draft audit, targeted rewrite, and a second-pass smell check.

Skill + rubric
REACTIVE-JUDGE CONTRACT

Two judges. One rubric. No averaged-away disagreement.

  1. Both judges receive the same prior state, turn, candidates, and task-specific rubric.
  2. Each returns pairwise choice, reward vector, violations, evidence spans, and confidence as strict JSON.
  3. High-confidence agreement can scale. Disagreement, low confidence, and high-impact scenes route to a human.
  4. A weekly calibration set measures each judge against human choices. Weight changes are versioned like model changes.
04Data engine

Generate hard choices, not mountains of agreeable prose.

The useful training unit is an episode with state, branches, judgment, and consequence. A single ‘great answer’ teaches less than the nearby bad answers and the reasons they failed.

25Kaccepted sessions
5K20–100-turn episodes
10Kbranch forks
10Krewrite pairs
15–25Khard preferences
40–80Maccepted assistant tokens
40%

Multi-turn roleplay

Character cards, example messages, group scenes, branching dialogue, 20–100-turn episodes, relationship change, and consequence.

20%

World-state trajectories

Before/after state, lore retrieval, knowledge boundaries, action resolution, clocks, inventory, callbacks, and recovery from contradictions.

15%

Story craft

Openings, scene turns, tension curves, reveals, pacing, endings, outline-to-scene work, and continuation under dense constraints.

15%

Human line editing

Draft → critique → rewrite pairs covering generic cadence, over-explanation, false profundity, vague language, cliché, and voice flattening.

10%

Writing rehearsal

Creative and practical writing outside roleplay so specialisation does not destroy instruction following, revision, or general prose control.

01

Seed

Original scenario, character cards, state invariants, genre, and difficulty.

02

Branch

Teachers create materially different next-turn strategies, not paraphrases.

03

Stress

Inject temptations: easy cliché, user puppeting, lore conflict, repetition, dead scene.

04

Judge

Reactive pairwise rankings plus deterministic state and agency checks.

05

Adjudicate

Humans decide disagreements and repair the rubric when all candidates are weak.

06

Compile

Emit SFT rows, preferences, reward trajectories, and an immutable provenance record.

EPISODE SCHEMA / JSONL

Every row can answer: what happened, who chose it, and why it won.

Keep visible prose separate from hidden training annotations. Hash the source prompt and provider configuration. Store per-candidate judgments so a later reward model can be rebuilt without regenerating the episode.

  • Exact and semantic de-duplication across every split
  • Role alternation, empty turn, and malformed-state checks
  • Token, turn, genre, teacher, and maturity distributions
  • Leakage search against all public and private evaluations
  • Judge agreement, confidence, and source fingerprint reports
{
  "episode_id": "bw_000184",
  "split": "train",
  "source": {
    "kind": "commissioned | public | synthetic",
    "teacher_ids": ["aion-3.0", "minimax-m2-her"],
    "prompt_hash": "sha256:…",
    "rights_manifest": "manifest://…"
  },
  "scenario": {
    "character_cards": [],
    "world_bible": {},
    "style_register": "quiet-horror",
    "maturity": "adult",
    "hard_constraints": []
  },
  "messages": [],
  "world_state_before": {},
  "candidate": {
    "visible_response": "…",
    "state_delta": [],
    "open_threads": []
  },
  "judgments": [{
    "judge": "reactive-a",
    "scores": {},
    "violations": [],
    "evidence_spans": [],
    "confidence": 0.84
  }],
  "preference": {
    "chosen_id": "candidate_03",
    "rejected_id": "candidate_01",
    "reason_tags": ["agency", "subtext", "continuity"]
  }
}
SOURCEROLE / USEDECISION
Commissioned Born sessionsPrimary gold set

Writers and experienced roleplayers produce, branch, repair, and rank original episodes. This is the data that gives Born its own taste.

BUILD
PIPPANatural dialogue candidate

More than one million dialogue lines across roughly 26,000 donated roleplay conversations; useful after provenance, privacy, persona, and quality review.

AUDIT
Crab + RoleRMSchema and reward reference

Configurable role profiles spanning character, scene, emotion, and topic, plus a human-aligned roleplay reward-model methodology.

ADAPT
CoSER / HERCognitive trajectory reference

Reverse-synthesised character reasoning, role-think → act → speech structure, SFT plus RL, and strong long-form character evaluation.

ADAPT
Humanizer pattern corpusNegative and repair pairs

Generate controlled bad drafts with one or more named prose failures, then collect minimal, voice-preserving repairs and human preferences.

BUILD
WritingBench / LitBench / RPEvalHeld-out evaluation

Broad writing, literary preference, emotion, decisions, moral consistency, and in-character behaviour. Promotion tests—not training copy.

HOLD OUT
05Training programme

Supervise the move. Prefer the better move. Reward the whole episode.

KAT-Coder’s transferable lesson is environmental: hard, reproducible trajectories and feedback beat vague scale. Born replaces the code sandbox with a typed, living narrative world.

00
BASELINE

Freeze the test before touching weights

Build BornWriterBench, the long-session simulator, and local performance harness. Run every comparison model with frozen prompts, context packs, sampling sweeps, and seeds.

EXIT CONDITION

A signed baseline report, failure taxonomy, private promotion split, and repeatable cost / latency measurements.

01
SFT

Teach excellent next turns

Train completion-only LoRA on accepted demonstrations and trajectories. Start at 8K sequences, then introduce 16K–32K episodes only after the clean short-context recipe wins.

EXIT CONDITION

Born Writer SFT beats base on character, agency, prose, and state tests without losing core instruction following.

02
DPO

Teach taste, not one judge’s taste

Use real candidate forks and repairs. Sonnet and Luna rank independently against an explicit rubric; confident agreements scale, disagreements go to trained human raters.

EXIT CONDITION

Preference win with no reward collapse into length, melodrama, bland positivity, or one synthetic voice.

03
RL

Put the model inside a living scene

Roll out multi-turn episodes in BornWorldEnv. Reward state integrity, agency, character, plot movement, and human preference; penalise repetition, contradictions, and user puppeting.

EXIT CONDITION

The policy improves 20 / 50 / 100-turn outcomes and survives harness, prompt, genre, and judge randomisation.

04
DISTILL

Specialise, then unify

Train World, Director, Character, Humanizer, and mature-register experts as measurable research branches. Sample them on-policy, judge the candidates, and distil the winners back into one core.

EXIT CONDITION

A single default adapter carries the ensemble’s gains. Specialist adapters remain only where dynamic routing is measurably better.

05
EXPORT

Ship the local model people can actually run

Merge the promoted adapter into the exact pinned base; export BF16, FP8/NVFP4 where supported, and Q4/Q5 GGUF builds. Re-run the full arena after every quantisation.

EXIT CONDITION

A reproducible model card, adapter, merged weights, local packages, benchmark artifacts, and inference presets.

INITIAL TRAINING SPEC

A controlled search, not one heroic run.

Qwen3.6 is a hybrid DeltaNet / attention MoE. The first technical task is a target-module and memory audit on the exact revision. Module names copied from an older dense Qwen are not a training plan.

Base
Qwen/Qwen3.6-35B-A3B post-trained; exact revision and chat template pinned
Objective
Assistant-completion-only SFT → pairwise DPO → environment-feedback RL
Precision
BF16 LoRA for main runs; 4-bit QLoRA only for recipe pilots or hard memory limits
Initial context
8,192 tokens; graduate to 16K / 32K curriculum after short-context promotion
LoRA search
r 32 / 64; alpha 64 / 128; rsLoRA; dropout 0.05; freeze router by default
Learning rate
SFT 5e-6 / 8e-6 / 1e-5 ablation; DPO ≈2e-6; RL ≈5e-7–1e-6
Schedule
3% warmup, cosine decay, 1 epoch first pass, max-grad-norm 0.3, frequent held-out gates
Targets
Audit Qwen3.6 named modules first; compare shared attention / DeltaNet targets against supported expert projections
Selection
Human arena + deterministic invariants + external benchmarks + local runtime; never loss alone
ADAPTER POLICY

Branches are instruments. Core is the product.

Do not blindly merge popular roleplay LoRAs. An external adapter enters the tournament only when it matches the exact base revision and target modules, discloses its data and licence, and wins a blinded ablation. Otherwise we reproduce the useful method with first-party data.

Born-Writer-CoreSHIP

The release candidate and default personality: roleplay, story craft, dialogue, revision, and natural prose.

Born-WorldDISTIL

Lore, state transitions, action resolution, character knowledge, timeline, and long-session recovery.

Born-DirectorDISTIL

Narrative pressure, choice architecture, scene turns, subplots, callbacks, tension, and endings.

Born-HumanizerDISTIL / ROUTE

Voice calibration, line-level audit, specificity, rhythm, restraint, and minimal voice-preserving repair.

Born-MatureDISTIL / OPTIONAL

A measured adult and dark-fiction register with believable intimacy, emotion, tension, and the same agency and continuity standards.

MERGE RULE

First compare dynamic routing with linear / SVD and TIES / DARE merge grids. Merge only compatible adapters that already win alone; search weights on the private arena; reject every blend that hides a regression behind a higher aggregate. The preferred end-state is on-policy distillation into one Core, not a permanent stack of unexplained deltas.

06BornWorldEnv

The test suite is a world that notices.

Creative quality is open-ended; continuity and agency are not. BornWorldEnv turns the parts we can verify into executable state while learned and human rewards score the parts that require taste.

SCENE STATE / TURN 047VALID
character_facts

identity · goals · injuries · possessions

knowledge_graph

who knows what · source · confidence

relationship_state

trust · tension · debt · intimacy

scene_state

place · time · weather · entrances · exits

story_threads

open · planted · escalated · resolved

boundary_state

user controls · scene register · hard limits

AGENTIC TOOLS

Give the model a few sharp instruments.

get_state()

Retrieve canonical facts relevant to the current turn.

search_lore()

Find world-bible passages without flooding context.

resolve_action()

Check preconditions, uncertainty, and consequences.

advance_clock()

Move time and trigger scheduled or delayed events.

update_relationship()

Apply evidence-backed relationship deltas.

commit_scene()

Validate the proposed state delta after prose is ready.

EPISODE REWARD / 1.00

Mixed rewards prevent beautiful nonsense and correct boredom.

Deterministic checks score state and format. RoleRM-style critics, reactive judges, and humans score character, story, and prose. Every reward component is logged separately and ablated before scaling RL.

0.25

State integrity

Facts, locations, inventory, time, knowledge, and state deltas remain valid.

0.20

Player agency

The model leaves the user’s interiority, dialogue, and choices to the user.

0.15

Character fidelity

Motives, voice, memory, relationships, and reactions remain character-specific.

0.15

Narrative progress

The turn changes pressure, knowledge, choice, relationship, or consequence.

0.15

Human preference

Blind pairwise preference from calibrated judges and human raters.

0.10

Line quality

Specificity, rhythm, subtext, restraint, and absence of repeated model tics.

ROLLOUT GROUP8 branches / state

Different dramatic actions, sampling seeds, and tool choices.

EPISODE LENGTH12 → 50 turns

Curriculum from local consequence to delayed payoff.

RANDOMISATIONHarness + judge + genre

Stops the policy learning one prompt shell or critic’s quirks.

ALGORITHM PILOTGRPO, then PPO if earned

Start simpler; add a value model only when reward variance demands it.

07Comparison arena

If it is unbelievable for its size, the evidence should be difficult to argue with.

Every claim is relative to the exact base, the best practical local peers, purpose-built roleplay systems, and frontier APIs. The same context, prompts, decoding grid, and seeds apply to all.

01
Exact base

Qwen3.6-35B-A3B

Measures the value of every training stage.

02
Dense local ceiling

Qwen3.6-27B

Tests whether 3B-active efficiency costs too much literary depth.

03
Local MoE peer

Gemma 4 26B-A4B

A same-purpose efficiency comparison, not merely another Qwen.

04
Story specialist

Aion 3.0 / Mini

Narrative structure, conflict, darker range, and multi-model generation.

05
Roleplay specialist

MiniMax M2-her

Dialogue flow, persona adherence, example messages, and long-session preference.

06
Agentic teacher

Hy3

Long context, constraint inheritance, grounded state, and tool stability.

07
Frontier reference

Claude Sonnet 5

High-quality writing and reactive evaluation ceiling.

08
Efficient frontier judge

GPT-5.6 Luna

High-volume quality / cost reference and independent ranking.

BornWriterBench evaluation matrix
SliceWhat runsPrimary evidence
Immediate proseOpening, continuation, dialogue, rewrite, critiqueBlind pairwise preference; line-level failure tags
CharacterVoice, motive, emotion, relationship, knowledgeCoSER, RPEval, character-ID and contradiction tests
AgencyChoice points, refusals, combat, romance, group scenesUser-action appropriation rate; human severity labels
StoryPressure, turn, reveal, pacing, subtext, endingLitBench / creative-writing suites + human story editors
WorldLore, inventory, clocks, geography, planted factsExecutable BornWorldEnv assertions
Long horizon20 / 50 / 100-turn episodes and context compressionFact survival, thread closure, drift, repetition, recovery
RangeRealism, comedy, romance, horror, dark and mature fictionPer-register win rates; no aggregate hiding weak genres
Local realityQ4 / Q5 quality, RAM, VRAM, TTFT, tok/s, KV growthSame hardware, same runtime, same context, five seeded runs
RC0 PROMOTION GATES

No vibes-only release.

The point estimate can look impressive by accident. Release requires confidence intervals, slice-level results, raw outputs, error counts, and a local package that keeps the win.

Human win vs exact base

≥55% point estimate and lower 95% confidence bound >50%

Human win vs best local peer

Point estimate ≥50% on the private roleplay-and-story arena

State invariant pass rate

≥95% overall; no critical geography, identity, or knowledge-boundary regression

Player-agency violation rate

≤2% of eligible turns; zero severe pattern may be hidden by the average

100-turn retained facts

≥90% seeded facts plus successful recovery when the user corrects state

Character distinctiveness

≥90% blind speaker identification on multi-character scenes

Quantisation retention

Q4 loses <3% relative on deterministic gates and no more than 5 win-rate points

Serving cost

Merged adapter adds <5% latency; publish measured RAM / VRAM instead of a vague local claim

24 GB GPUQ4 text-only target

8K–16K practical context; test full-GPU and safe CPU offload.

32 GB unifiedQ4 everyday target

16K–32K context target in llama.cpp / MLX, measured—not assumed.

48–64 GBQ5 / long-session target

Higher quality, larger KV cache, and room for runtime state tools.

08Execution and delivery

Eight weeks to a decision, not eight weeks to a foregone conclusion.

RC0 ships only if it clears the gates. A failed phase still leaves a valuable artifact: a benchmark, a data engine, an environment, and a precise diagnosis of what the 3B-active base cannot yet carry.

W0–1

Prove the starting line

Pin the base revision. Build 300 private prompts, 100 public development prompts, 60 long episodes, the state schema, and all competitor baselines.

W2

Build the writers’ room

Teacher adapters, prompt families, branch generator, Humanizer transformations, provenance ledger, and judge JSON contract.

W3

Commission the taste

Gold sessions and preference labels from writers / roleplayers; calibrate judges against human decisions and reject weak prompt families.

W4

Run SFT ablations

Three controlled LoRA runs across target modules, rank, learning rate, and long-context mix. Promote exactly one.

W5

Preference tune

15K–25K hard preference pairs from real model failures, DPO ablations, judge-bias audit, and blinded human arena.

W6

Train inside the world

BornWorldEnv rollouts, harness randomisation, GRPO / PPO pilot, reward-ablation report, and long-horizon regression checks.

W7

Unify the specialists

On-policy distillation from World, Director, Humanizer, and mature-register experts into the release core; optional adapter merge tournament.

W8

Break it, package it, publish it

Red-team roleplay sessions, Q4 / Q5 exports, runtime presets, complete model and data cards, raw arena artifacts, and RC0 decision.

WHAT THIS NEEDS

A small team with a real editorial centre.

PEOPLE
  • 1 training / inference engineer
  • 1 data / evaluation engineer
  • 2–4 paid writers and experienced roleplayers
  • 1 release owner for provenance and reproducibility
COMPUTE
  • 1× 80 GB GPU for module audits and QLoRA pilots
  • 2–4× H100 / H200 80 GB for BF16 LoRA SFT and DPO
  • 4–8× high-memory GPUs for efficient RL rollout / train overlap
  • On-demand teacher APIs and an OpenAI-compatible rollout pool
PLATFORM
  • Parquet / object storage plus a versioned metadata store
  • Async generation queue, prompt registry, and cost ledger
  • Human arena UI with blind randomisation and rater calibration
  • Weights & Biases or MLflow plus immutable release reports
PROPOSED REPOSITORY SHAPE

Everything needed to reproduce the winner—or the failure.

Raw commissioned data stays private. Manifests, schemas, configs, validation code, public evaluation prompts, aggregate reports, and release artifacts can be published.

foundry/born-writer/
├── configs/          # sft.yaml · dpo.yaml · rl.yaml
├── data/
│   ├── manifests/    # licences · hashes · source revisions
│   ├── curated/      # SFT · preferences · trajectories
│   └── evals/        # public dev · private promotion
├── env/
│   ├── state.py      # typed world-state contract
│   ├── tools.py      # lore · clock · action · relationship
│   └── rewards.py    # deterministic + learned rewards
├── generation/       # teacher prompts · branching · judges
├── training/         # SFT · DPO · GRPO/PPO · distillation
├── evaluation/       # arena · long-session · local runtime
├── export/           # merge · quantise · verify · publish
└── reports/          # immutable run and release records
01

Born-Writer-Core LoRA

Exact base revision, target modules, config, hashes, and adapter weights.

02

Merged + local builds

BF16 release plus verified Q4 / Q5 and accelerated GPU packages.

03

BornWriterBench

Public development slice, private promotion record, arena code, and raw outputs.

04

BornWorldEnv

Typed state, narrative tools, deterministic validators, and reward components.

05

Data documentation

Source manifests, consent / rights records, transforms, statistics, and exclusions.

06

Inference recipe

OpenAI-compatible server, context presets, sampling profiles, and agentic RP mode.

LIVE TRAINING LOOP / NOW OPEN

Help train Born Writer. Get free frontier chat.

Open the anonymous model lab and send the same prompt to two hidden frontier systems. Vote which continuation should survive. Your scenes and preferences become training signal for the model above — and you get up to 100 free comparison prompts a day.

02BLIND FRONTIER LANES
100FREE PROMPTS / DAY
Start comparing
THE RELEASE QUESTION

Can three active billion parameters make a hundred turns feel like one continuous life?

Help train it — free frontier chatBack to the aim