Arjuna Badger Press

The Press Thesis

Grounded fiction, guarded intention, human verdict

Arjuna Badger Press — The archer's eye. The badger's nerve. Author: Andries J. Greyling (AJ Greyling) Status: Working thesis — June 2026 Form: Interdisciplinary practice-based research at the intersection of creative writing, narratology, computational creativity, and authorial craft — submitted as published work + critical commentary + reproducible practice record, with qualitative human reading as the ultimate proof standard.

Abstract

This thesis argues that serious fiction can be produced at scale with machine assistance without surrendering authorial soul — but only if three invariants hold: (1) creative intention is declared in canon and guarded by deterministic tooling that measures and alarms rather than rewrites; (2) literary structure can be specified and enforced like a type system (the Triptych Trilogy as executable form); and (3) the final validity of the work is decided not by metrics, rubrics, or model self-assessment, but by the qualitative judgment of human readers — editors, beta readers, and the public — who either stay inside the dream or are pulled out of it.

The through-line is one mechanism applied everywhere: the wrapper is not the work. A harmful request labelled "fiction" is still a harmful request. Cargo-cult craft vocabulary without grounded content does not improve prose. Multi-pass automated rewrite pipelines can regress quality. Colonial framing that outsources African ingenuity to aliens is still colonial framing. What travels in the context — the actual sentences, places, constraints, and stakes — is what matters; decoration does not substitute for grounding.

Arjuna Badger Press is the house built on that claim: a studio and a free craft library grown from twenty-plus finished manuscripts, an engine that finishes books in an author's voice (thinning machine-tells, not replacing voice), and a public shelf where the proof is to be determined by whether people read, re-read, and recommend — not by whether a scorecard said pass.

Companion documents: THESIS.md (engineering) · CREATIVE_THESIS.md (creative intentions) · FOR_AUTHORS.md (the workshop) · academic/CRITICAL_COMMENTARY.md (submission spine) · academic/TRIPTYCH_FORM.md (literary theory) · craft/LLM_TELLS.md (de-LLM catalog) · guardrail-register-thesis (empirical register study, sibling repo)

1. The one-sentence claim

A human author can direct a machine to finish print-ready novels at scale — and the only honest proof is whether other humans want to read them.

Everything else in this document — the continuity gate, the Triptych obligations, the de-LLM scanners, the NovelBench dimensions, the git history of rejected polish passes — is supporting apparatus. Instrumentation makes intention holdable across hundreds of chapters. It does not replace the reader's verdict.

2. What Arjuna Badger Press is (and is not)

2.1 The house

Arjuna Badger Press is two things at once:

1. A public library — free craft knowledge, finished novels to read, the Triptych and de-LLM

doctrine given away without gatekeeping. Mastery without credential.

1. A workshop — for authors who already have a soul in the drawer: ingest the pile (manuscript,

notes, photos, lore), answer ~twenty binding questions, click Go, return to a continuity-clean draft ready for the author's proofread. Not beginner's toy; Rothfuss-scale finish problem.

The emblem is deliberate. The honey badger is fearless (walk at the work anyway), nearly unkillable (structure as resilience), an escape artist (craft as cleverness, not talent-worship), and — the tender note under the ferocity — willing to be carried. The strongest creature allows help. The work is fierce; the maker may be gentle with themselves.

2.2 The line that must not be crossed

The engine measures and alarms. It does not replace your voice.

• A human seed plus surgical edits authors the soul; the pipeline does not drive.
• The gatekeeper may reject a flattened polish pass and fall back to the stronger draft.
• de-LLM is thinning, not banning — machine-tells are good moves miscalibrated by evenness.
• The wizard proposes; the author decides — the model must not both propose and decide at forks.

This is not humility marketing. It is the epistemological core: only a situated human can sign the book. Tools extend reach; they do not confer authorship.

2.3 What it is not

• Not a course to sell or a guru's system — the craft library is reference, take what's useful.
• Not "hide that AI helped" — provenance is disclosed; immersion is the craft obligation.
• Not proof-by-dashboard — scores and gates are microscopes for the author and editor, not substitutes

for readers.

3. The mechanism: content-addressed, not frame-addressed

Large language models predict continuations from what is actually in the context — tokens, register, constraints, examples — not from decorative wrappers around the task.

Constructive corollary: Correct craft vocabulary can help when it disambiguates and points into a region of competent practice the author actually inhabits. Destructive corollary: Wrong or ungrounded register injects noise; recursive framing does not evade evaluation.

The press operationalises the constructive half with a grounding-gated register scaffold — craft terms are inserted only when elicited from the author's actual material — and refuses the destructive half by never treating framing as permission.

(Empirical sibling study: guardrail-register-thesis/thesis/THESIS.md; companion fiction: Afraud All The Way Down.)

4. The four creative intentions (what the books are for)

Every manuscript in the library serves intentions declared in canon and guarded in code. They are stated here in order of delivery; the root is §4.4.

4.1 Immersion — strip the seams, not the truth

Remove LLM tics and craft friction so the reader stays inside the fictional dream. This is not deception — the repository documents how the books were made — it is the same obligation every writer has: don't let the apparatus show. Provenance open; experience clean.

Guarded by: de-LLM loop · prosetics.py · STYLEGUIDE machine-tell taboos · craft/LLM_TELLS.md

4.2 Disciplined mythos — woo in small doses, history as backbone

Ancient-mystery material is texture, never the load-bearing meal. Real places, real geology, real light. Mysteries stay belief and tradition until the one sanctioned climax — and it lands because everything else stayed grounded.

Guarded by: MYTHOS_RULES · storygraph/mythos.py · Rule 0: ground everything, so the one place you don't lands like a thunderclap.

4.3 Place magnetism — make them want to stand here

Render heritage sites with such specific, sensory, true magnificence that the reader wants the plane ticket — awe through scale and craft, never magic; honest texture, not theme-park gloss.

Guarded by: WONDEROFPLACE · NovelBench place_magnetism axis (microscope, not verdict)

4.4 Pride — African ingenuity in African hands (the root)

The ancient-mystery genre carries a colonial poison: the unspoken assumption that someone else built Africa's wonders. This work inverts on purpose — brilliance is human, African, ours; shown, not lectured. Pride is the mission; education and tourism follow when ownership is rendered truly.

Guarded by: THEMES · WONDEROFPLACE order-of-heart · the History Before Time companion shelf

graph TD
  G4["④ PRIDE / OWNERSHIP<br/>(the root — the mission)"]
  G2["② DISCIPLINED MYTHOS<br/>(credibility — brilliance is ours)"]
  G3["③ PLACE MAGNETISM<br/>(pride made into a plane ticket)"]
  G1["① IMMERSION<br/>(nothing breaks the dream)"]
  G4 --> G2 --> G3
  G1 -.protects delivery of.-> G2
  G1 -.protects delivery of.-> G3
  G1 -.protects delivery of.-> G4

5. The literary contribution: the Triptych Trilogy as executable form

Most structural claims in creative writing are made in retrospect. This project advances a form with falsifiable proof obligations — and enforces them over the actual manuscripts:

The African Gold Trilogy (RESONANCE · REVELATION · RELIC) is the primary demonstration. The continuity gate is the form's executable definition: violations block merge; crumbs, relay nodes, timeline braids, and mythos rules are checked as graph constraints, not editorial vibes.

Full theory: academic/TRIPTYCH_FORM.md · public edition: craft/TRIPTYCH_FORM.md

Literary proof still rests with readers: order-independence is a structural promise; whether each order feels complete and rewarding is a question for human reading, not the gate alone.

6. The methodological contribution: measure, don't generate

The engine pairs single-shot LLM generation with a deterministic finite-state continuity engine and a measurement harness. The binding finding from practice:

The deterministic layer must measure and alarm, never generate — because when tools were allowed to rewrite prose in bulk, quality regressed. The proof is in the git history.

Multi-provider routing separates expensive prose from cheap analysis. Checkpointed runs resume across days. Twenty-plus books in one monorepo share a unified trilogy gate and cross-series continuity auditors — the method scales because constraints are code, not hope.

Engineering detail: THESIS.md · docs/MASTER_PLAN.md · docs/TECHNOLOGY.md

7. Proof — to be determined by human readers

7.1 The epistemological stance

This thesis refuses the claim that a novel is "proven" because:

• a rubric scored ≥ 4/5,
• a continuity gate passed,
• a de-LLM tic count fell within band, or
• a model judged its own output acceptable.

Those are necessary microscopes for the author and editor in the workshop. They are not sufficient for literary validity.

Proof is qualitative. It accrues when:

1. A reader finishes — and does not feel pulled out by seams, homogenized voice, or over-explained

significance.

1. A reader returns — re-reads, recommends, argues about character, order, and meaning.
2. An editor's judgment aligns — developmental, line, sensitivity: the work survives professional

eyes without requiring apology for how it was made.

1. The dream carries the mission — for this corpus especially: an African reader may stand taller;

any reader may want to stand on real ground the book rendered honestly.

Metrics support that proof — they do not substitute for it. NovelBench's place_magnetism axis asks a human-shaped question ("do you want to go there?") but still requires a human to answer it honestly after reading.

7.2 Who counts as judge

The press publishes openly — arjunabadger.press — so the largest jury is anyone who reads. That is intentional. A thesis that hides behind paywalls and peer-only samples would contradict its own immersion ethic.

7.3 What counts as evidence for the claim

• The corpus — ~22 merged manuscripts, multiple series, one engine, one authorial line; read online

with EPUB/PDF exports.

• The practice record — dated decisions, rejected polish passes, gate failures caught pre-merge,

external editorial feedback ingested into STYLE_GUIDE and prompts.

• The craft library — 90+ terms, de-LLM catalog, anti-patterns: doctrine extracted because it

survived contact with real books.

• Professional editorial traces — cold-read reports, craft audits, triptych-judge runs: human

language describing failure modes, not just numbers.

• Reader uptake (ongoing) — time on page, correspondence, re-read permutations of the Triptych,

tourism intent — qualitative and slow.

7.4 What would falsify the claim

The thesis is wrong if informed human readers consistently report:

• prose that reads as machine-even despite tic thinning,
• structure that fails the Triptych promises when read in alternate orders,
• mythos that leaks woo into literal fact without earning it,
• place-writing that flatters or falsifies real heritage,
• or mission that lectures rather than embodies pride.

A passing gate with failing readers means the gate is incomplete — not that readers are wrong. Readers are the court of final appeal.

8. Authorship under constraint

Where does authorship live when generation is cheap?

1. Canon — names, timeline, mythos physics, voice laws: human-locked.
2. Forks — wizard decisions the model cannot make alone.
3. Constraint design — what the gate enforces is an authorial theory of the book.
4. Surgical edits — the de-LLM loop and line pass: human thins where scanners point.
5. Signature — the proofread, the cut only the author would make, the willingness to publish.

The collaboration record (docs/COLLABORATION.md) documents where the agent was wrong and where tooling caught what no human would — not to boast about automation, but to show directed autonomy: the human remains architect; the machine is labour; the gate is conscience.

9. The Misogi origin (why the month happened)

The press grew from a thirty-day vow: one subscription, one novel, non-slop. What shipped was not one book but a house — because the vow's real goals were four: evaluate the agent stack; produce literary craft at scale; assemble portfolio-grade proof of platform thinking; understand LLMs deeply enough to know where they fail.

The vow is kept. The proof of literary success, however, remains open — decided by readers over time, not by the month's output count alone.

Record: docs/MISOGI.md · docs/ORIGINS.md

10. Conclusion — contribution in three sentences

Literary: The Triptych Trilogy — a three-novel form with falsifiable panel-completeness, weave-closure, and order-independence — demonstrated and executable via a continuity gate.

Methodological: Single-shot generation plus deterministic FSM/graph enforcement plus measurement harness — with the invariant that tools measure and alarm, not replace voice — documented across a multi-book practice record.

Epistemological: The ultimate validation of all of the above is qualitative human judgment — whether readers stay inside the dream, stand taller, want to stand on real ground, and return.

The engine exists so a human author can finish at scale without lying to the reader. Arjuna Badger Press exists so that practice is shared — craft given away, books offered freely, the workshop opened to authors who need finish, not permission.

The archer's eye. The badger's nerve.

Appendix — document map

Craft Library · Place Wiki · About the press · View this document on GitHub · Write with us

← Back to the library