Honest critique from LLM personas, despite the training.

A Python pipeline for getting alpha-reader feedback on a fiction manuscript that pushes past RLHF sycophancy. Anti-sycophancy bindings, two-pass design, multi-model dispatch, reflexive Opus self-critique. ~$8-25 per full novel.

— start here

Read the methodology first.

"v1 produced 7/7 keep-reading verdicts on a manuscript that needed serious rework. v0.4 rebuilds the cohort to push past that."

Read METHODOLOGY.md for the four mechanisms (anti-sycophancy bindings, two-pass design, reflexive self-critique, multi-model dispatch). Then the personas snapshot for the 9-persona cohort and the 10 universal anti-sycophancy bindings. Then run new-splinter.py on a test manuscript.

— I

The Pipeline — scaffolder + orchestrator + analyzer

Three Python scripts plus a scaffolder that turns "I have a manuscript" into "ready to fire" in ~2 minutes.

1.1

new-splinter.py

Scaffolder · ~2 minutes from manuscript to ready

One command: converts manuscript (.docx → .md), generates book_config.yaml + CLAUDE.md + kickoff, copies orchestrator + analyzer + auth_probe + personas, optionally sources API key, runs readability analyzer.

Start here

1.2

run-alpha-reader.py

Canonical orchestrator · 1072 lines · reads book_config.yaml

Runs personas across passes with anti-sycophancy bindings, multi-model dispatch (Sonnet for most, Opus for adversarial + audit), reflexive self-critique, atomic state for resume.

Core

1.3

readability_analyzer.py

textstat metrics panel

Per-chapter and whole-book readability stats (FK, FRE, sentence/word/syllable averages). Output feeds into per-persona symmetric FK guard binding.

Analysis

1.4

auth_probe.py

Verify API key pre-fire

~30-line script that does a single low-cost API call to confirm the loaded key works. Run before firing 9 personas across 2 passes.

Safety

1.5

install-into-existing-splinter.sh

Refresh toolkit in existing project (.sh + .ps1)

Brings an existing project up to current toolkit version. Replaces orchestrator/personas without touching manuscript/runs/book_config.

Maintenance

— II

Methodology — how the pipeline pushes past sycophancy

Four overlapping mechanisms. Each is partial alone; together they produce noticeably more honest output.

2.1

METHODOLOGY.md

The full read

All four mechanisms (anti-sycophancy bindings, two-pass design, reflexive self-critique, multi-model dispatch). When to use, when not to. Honest limits.

Foundation

2.2

9-persona snapshot

Anti-sycophancy preamble + persona profiles

The 10 universal bindings (rating discipline, no hedging, DNF honesty, no conversion arcs, no warmth padding, forced negative generation Q0, mandatory 1-star Q9, comp ranking, distribution prediction, authority-relationship inversion). 7 lane-specific personas + 2 adversarial (Veronica, Emma).

Load-bearing

— III

Templates — per-project scaffolds

The scaffolder generates these for new projects. Reference copies for existing projects or manual setup.

3.1

book_config.yaml

Per-project config schema

Load-bearing — orchestrator reads from this. Defines book metadata, manuscript path, genre context, cost cap, model defaults, optional Pass B reference files.

Required

3.2

CLAUDE.md template

Per-project Claude Code context

Project-scoped configuration: cost discipline, fire commands, persona roster, hard rules. Generated by scaffolder; customize per project.

Template

3.3

Kickoff template

Project kickoff document

Authorization-gate document with pipeline architecture, fire commands, pre-fire checklist, post-fire deliverables list.

Template

3.4

.env.example

API key template

Note: the .env loader uses python-dotenv with override=True to bypass session-injected Claude Code tokens that fail on direct API calls.

Template

3.5

.gitignore template

Defensive gitignore

Excludes .env, runs/ (sensitive: contains critique of unpublished manuscript), Python build artifacts.

Template

— for non-github readers

Send a quick note.

Used the toolkit on your own manuscript? Hit a problem with the pipeline? Have findings on calibration vs human alpha readers? This form goes straight to the maintainer.

If you have a GitHub account, opening an issue is preferred. This form is the path for everyone else.