# Authored Prose Protocol — FULL RESEARCH CORPUS Abbreviation: app Version: v0.1.0 | Generated: unspecified | Files: 2 | Lines: 439 | Payload: 27.0 KiB Source path: /Users/zachshallbetter/Projects/_protocols/authored-prose-protocol Bundle name: llm-full-docs-app Corpus Payload SHA-256: ab48cbf431720e4aa7000d66bae8b2fbf802497bd3ae64e1cd81ba7dc8f5f6bc TABLE OF CONTENTS: - APP.html (315 lines, 19.0 KiB) - APP.md (124 lines, 8.0 KiB) ======================================================================== ======================================================================== ## FILE: APP.html (315 lines, 19.0 KiB) ======================================================================== Authored Prose Protocol
app:protocol/authored-prose v0.1.0 · normative candidate · 2026-08-26

Authored Prose Protocol

What prose published under this name must do, what it must not do, and how a draft is checked before anyone reads it. The rules are not preferences. Each is derived from a measured corpus or from an invariant already enforced elsewhere in the work, and each carries a test a reader who does not trust the author can apply.

Derivation · 173 articles (189,900 words, 2016–2024) · 600 working documents (201,900 words, 2026) · one voice specimen (1,052 words)
Companion · Distance-From-Gray/docs/Manuscript/voice-and-style.md

§1Derivation

The corpora disagree, and the disagreement is the finding. The essay is not a shorter version of the older voice. It is a different instrument.

MeasureArticles 2016–24Working docs 2026Voice specimen
Mean sentence length17.115.9–21.411.6
Median sentence length1512–1610
Sentences ≤ 8 words13.7%22–34%44%
Sentences ≥ 30 words7.8%10–17%4.4%
Em dashes per 1,000 words~4.713–230
Second-person uses1,045rare4
Passive constructionscommoncommon1

Where the corpora conflict, the specimen governs prose and the working corpus governs apparatus. The older voice instructs a reader. The current one states a position and leaves the conclusion to them.

§2Voice

Twelve rules. The right-hand column is how each is checked without a human reading for it.

V1

Body prose MUST NOT use em dashes. They may appear in tables, notes, frontmatter and working documents.

count of — in body lines = 0
V2

Body prose MUST NOT instruct in the second person. The reader is not being coached.

Second person may appear where the sentence hands the reader a test they can run themselves.

no second-person imperatives outside test-handing sentences
V3

Mean sentence length SHOULD fall under 14 words in a finished draft.

mean(sentence words) < 14
V4

A definition MUST be written as denial and correction in two sentences, not one sentence joined by but.

Gray is not wrong. It is unfinished.

The rejected reading has to be inspectable on its own.

read; flag "is not … but" in definitions
V5

A term MUST be coined once, defined at first use, then held. Synonyms for a defined term MUST NOT appear.

Varying a term for elegance is a defect. A term that drifts cannot be checked.

term list vs. synonym scan
V6

Where a practice is criticized, the passage MUST concede its competence first. The error MUST then be placed in the structure rather than in the people.

This is reasonable product design, and it is the wrong layer for the fix.
read
V7

Where a claim could be taken on trust, the passage MUST state what would check it. The requirement MUST rest on the author.

You do not have to trust the answer. You instrument the seam and read it.
read; every trust-claim has a stated check
V8

Where a sentence is about to overreach, the overreach MUST be named and refused in the same passage.

The precision that keeps this claim honest: this is not deterministic AI.
read
V9

Paragraphs SHOULD average four to five sentences and stay under fifty words. A one-sentence paragraph may be used for weight.

mean(words/paragraph) < 50
V10

Headings MUST be flat noun phrases or plain statements. Questions, puns and clever headings MUST NOT be used.

no "?" in headings
V11

These MUST NOT appear in body prose: however, moreover, furthermore, seamless, leverage, robust, delve.

The first three are absent from the specimen. The rest are artifacts of the older corpus, which used seamless 78 times. If an idea needs one of these words, it has not been specified yet.

banned-term scan, excluding this rule
V12

Passive voice SHOULD NOT be used where an actor is available. Adverbs SHOULD stay near ten per thousand words.

Prefer a stronger verb or a number.

passive density; -ly per 1,000

§3Evidence

The eight invariants already enforced in the code and the documents, restated as obligations on prose. A draft that fails any of these is not a draft.

E1

One canonical shape. A concept MUST have one name and one definition across every document.

Two representations of the same thing drift, and the drift is silent.

term inventory across corpus
E2

Fail closed on the unknown. Where a fact is not known, the sentence MUST say so or be cut. A plausible filler MUST NOT stand in for an unknown.

A guess that reads like an answer cannot be audited later.

read; unresolved facts carry a tag
E3

Every claim declares its backing. Each objective statement MUST carry a lane in the source: External Fact with a citation, Internal Policy labeled as policy, or Internal Result with a proof artifact or a Pending Proof tag.

Maturity MUST NOT outrun provenance.

lane completeness
E4

A gate must be able to fail. Any check the document claims to apply MUST have a demonstrated failure.

A check nothing has ever stopped is a preference described as a rule.

one recorded failure per claimed check
E5

Negative results are permanent. A retracted claim MUST be retained in the record with its reason, never deleted.

The discarded idea is the one most likely to return.

retraction list is append-only
E6

Conformance is declared, never inferred. Agreement between two documents, versions or layers MUST be stated in a record. It MUST NOT be read off a shared number, name, or adjacency.

conformance record exists
E7

Prove without disclosing. Evidence offered in prose MUST NOT include what it attests to where the content is sensitive. Cite the hash, the version, the reference, not the payload.

read; sensitive-payload scan
E8

Value is a measured delta. Where the text asserts something is better, it MUST give both measurements and the gap. A quality claim MUST NOT stand in for a measurement.

each comparative has two numbers

§4Withholding

W1

A fact that is real but unpublishable MUST be cut, not softened.

A hedged version of a withheld fact is worse than its absence. It asserts something while denying the reader the means to check it.

read
W2

Retracted material MUST NOT reappear on any surface, including by paraphrase, initials, or identifying description.

retraction scan across all surfaces
W3

Withholding MUST NOT be presented as a gap. Where the absence is deliberate the text may say so, and MUST NOT imply the material does not exist.

read

§5Conformance

A draft is conformant at a level when every rule at that level and below is satisfied.

LevelNameRequires
0VocabularyDefined terms used consistently (E1, V5)
1VoiceAll of §2
2Evidenced§2 and §3; every claim lane resolvable
3Published§2, §3, §4; every External Fact citation verified against a primary source on the publication date

This document is Level 1 conformant against itself. Measured after drafting: mean sentence 12.9 words, zero em dashes in body prose, no banned term outside §2. It failed on both counts in its first draft at 14.5 words and two em dashes, and was edited until it passed.

A checker MUST exclude §2 from its own banned-term scan, since V11 names the terms it forbids.

§6What this protocol does not claim

Conformance produces prose that is checkable, consistent, and honest about its own support. It does not produce prose that is good, interesting, or correct. A conformant document can be tediously written and entirely wrong about the world. Nothing in §2 or §3 detects that. Only a reader who disagrees will.

The rules also do not settle when a piece is finished. That decision is external to the document by construction, and the author's standing failure mode is treating it as internal.

======================================================================== ## FILE: APP.md (124 lines, 8.0 KiB) ======================================================================== # Authored Prose Protocol **Identifier:** `app:protocol/authored-prose` **Version:** `0.1.0` **Status:** Normative Candidate **Date:** 2026-08-26 **Author:** Zach Shallbetter **Primary artifact:** This document **Companion artifacts:** `Distance-From-Gray/docs/Manuscript/voice-and-style.md` (the measurement this is derived from) --- ## 0. Abstract A normative specification for prose written under this author's name. It covers essays, documentation, specifications, manuscript chapters, and copy an agent drafts on his behalf. It states what the prose must do, what it must not do, and how a draft is checked before it is published. The rules are not preferences. Each is derived from a measured corpus, or from an invariant already enforced elsewhere in the author's work. Each carries a test a reader who does not trust the author can apply. **MUST** indicates a conformance requirement. **SHOULD** indicates a default that requires a stated reason to break. **MAY** indicates a permitted option. --- ## 1. Scope and derivation Rules in §2 are derived from measurement performed 2026-08-26 across three bodies of text: the article corpus of 2016–2024 (173 documents, 189,900 words), the working corpus of 2026 (600 documents, 201,900 words), and the published essay *Distance from Gray* (1,052 words). The essay is designated the **voice specimen**. The corpora disagree, and the disagreement is the finding. The essay is not a shorter version of the older voice; it is a different instrument. | Measure | Articles 2016–2024 | Working docs 2026 | Voice specimen | |---|---:|---:|---:| | Mean sentence length | 17.1 | 15.9–21.4 | **11.6** | | Median sentence length | 15 | 12–16 | **10** | | Sentences ≤ 8 words | 13.7% | 22–34% | **44%** | | Sentences ≥ 30 words | 7.8% | 10–17% | **4.4%** | | Em dashes per 1,000 words | ~4.7 | 13–23 | **0** | | Second-person uses | 1,045 | rare | **4** | | Passive constructions | common | common | **1** | Where the corpora conflict, the specimen governs prose and the working corpus governs apparatus. --- ## 2. Voice **V1.** Body prose **MUST NOT** use em dashes. They **MAY** appear in tables, notes, frontmatter and working documents. **V2.** Body prose **MUST NOT** instruct in the second person. The reader is not being coached. Second person **MAY** appear where the sentence hands the reader a test they can run themselves. **V3.** Mean sentence length in a finished draft **SHOULD** fall under 14 words. A draft above that is edited before review. **V4.** A definition **MUST** be written as denial and correction in two sentences rather than one sentence joined by *but*. *Gray is not wrong. It is unfinished.* The rejected reading has to be inspectable on its own. **V5.** A term **MUST** be coined once, defined at first use, and then held. Varying a term for elegance is a defect: a term that drifts cannot be checked. Synonyms for a defined term **MUST NOT** appear. **V6.** Where a practice is criticized, the passage **MUST** concede its competence first. The error **MUST** then be placed in the structure rather than in the people. *This is reasonable product design, and it is the wrong layer for the fix.* **V7.** Where a claim could be taken on trust, the passage **MUST** state what would check it. The requirement **MUST** rest on the author. Prose that hands the reader the falsification instrument is the writing equivalent of a gate that can fail. **V8.** Where a sentence is about to overreach, the overreach **MUST** be named and refused in the same passage. *The precision that keeps this claim honest: this is not deterministic AI.* **V9.** Paragraphs **SHOULD** average four to five sentences and stay under fifty words. A one-sentence paragraph **MAY** be used for weight. **V10.** Headings **MUST** be flat noun phrases or plain statements. Questions, puns and clever headings **MUST NOT** be used. **V11.** The following **MUST NOT** appear in body prose: *however*, *moreover*, *furthermore*, *seamless*, *leverage*, *robust*, *delve*. The first three are absent from the specimen; the rest are artifacts of the older corpus. If an idea needs one of these words, it has not been specified yet. **V12.** Passive voice **SHOULD NOT** be used where an actor is available, and adverbs **SHOULD** stay near ten per thousand words. Prefer a stronger verb or a number. --- ## 3. Evidence The eight invariants the author already enforces in code and documents, restated as obligations on prose. **E1 — One canonical shape.** A concept **MUST** have one name and one definition across every document. Two representations of the same thing drift, and the drift is silent. **E2 — Fail closed on the unknown.** Where a fact is not known, the sentence **MUST** say so or be cut. A plausible filler **MUST NOT** be written in place of an unknown; a guess that reads like an answer cannot be audited later. **E3 — Every claim declares its backing.** Each objective statement **MUST** carry a lane in the source. The lanes are `External Fact` with a citation, `Internal Policy` labeled as policy or target, and `Internal Result` with a proof artifact or a `Pending Proof` tag. Maturity **MUST NOT** outrun provenance. **E4 — A gate must be able to fail.** Any check the document claims to apply **MUST** have a demonstrated failure. A check nothing has ever stopped is a preference described as a rule. **E5 — Negative results are permanent.** A retracted claim **MUST** be retained in the record with its reason, never deleted. The discarded idea is the one most likely to return. **E6 — Conformance is declared, never inferred.** Agreement between two documents, versions or layers **MUST** be stated in a record. It **MUST NOT** be read off a shared number, a shared name, or adjacency. **E7 — Prove without disclosing.** Evidence offered in prose **MUST NOT** include what it attests to where the content is sensitive. Cite the hash, the version, the reference, not the payload. **E8 — Value is a measured delta.** Where the text asserts that something is better, it **MUST** give the two measurements and the gap. A quality claim **MUST NOT** stand in for a measurement. --- ## 4. Withholding **W1.** A fact that is real but unpublishable **MUST** be cut, not softened. A hedged version of a withheld fact is worse than its absence: it asserts something while denying the reader the means to check it. **W2.** Retracted material **MUST NOT** reappear on any surface, including by paraphrase, initials, or description that identifies it. **W3.** Withholding **MUST NOT** be presented as a gap. Where the absence is deliberate, the text **MAY** say that it is deliberate and **MUST NOT** imply the material does not exist. --- ## 5. Conformance A draft is conformant at the stated level when every rule at that level and below is satisfied. | Level | Name | Requires | |---|---|---| | 0 | Vocabulary | Defined terms used consistently (E1, V5) | | 1 | Voice | All of §2 | | 2 | Evidenced | §2 and §3; every claim lane resolvable | | 3 | Published | §2, §3, §4; every `External Fact` citation verified against a primary source on the date of publication | **Checks.** Mean sentence length, em-dash count, banned-term hits, second-person imperatives and passive-voice density are mechanically checkable and **SHOULD** run before review. A checker **MUST** exclude §2 from its own banned-term scan, since V11 names the terms it forbids. Lane completeness and citation verification are read by a person. **A draft that fails any §3 rule is not a draft.** It is a document making claims it cannot support. It does not go to a reader. --- ## 6. What this protocol does not claim Conformance produces prose that is checkable, consistent and honest about its own support. It does not produce prose that is good, interesting, or correct. A conformant document can be tediously written and entirely wrong about the world. Nothing in §2 or §3 detects that. Only a reader who disagrees will. The rules also do not settle when a piece is finished. That decision is external to the document by construction, and the author's standing failure mode is treating it as internal.