Repository path: framework/v0.3/README.md · rendered 2026-09-09

Framework v0.3 — the handbook, organised by translation problem

What this is. The project's guidance, rewritten from the record into entries a translator can open and use and a pipeline can execute — one entry per translation problem, each carrying both the human instruction and the pipeline step. Constituted 2026-09-04 (S244) on Tom's direction (PROJECT.md §11): a problem-indexed handbook, built for a working translator and for an autonomous pipeline equally. Written by W1 sessions (wiki/plan.md), one family per session.

What it replaces, and what it keeps. framework/v0.2/README.md — 52 numbered sections appended in session order, 380 KB, many later withdrawn or corrected inline — is frozen as the record: no new numbered section is added to it; every finding after S243 goes into an entry here. framework/v0.1/README.md (R1 and the fourteen candidates) is likewise the record. Nothing in either is deleted or restated whole; an entry consolidates them — what survived, what was withdrawn, on what evidence — and cites them.

Standing (at the close, 2026-09-09, S257). provisional: true. Tier D is NOT PASSED (config/models.md), so no entry may say one rendering is better than another on a jury's word: evidence classes X1a, X1b and X2 (framework/v0.1/README.md §1) are admissible; X3, panel-scored quality, is not until a Tier D verdict licenses a sense. An entry describes what published translators do, what the project's practice found, what options exist and what each costs; its guidance is tagged evidenced (with the pairs) or untested. W2's one authorized redesign FAILED (2026-09-06, S250, RS-20260906-tierD-verdict-v3); Tier D calibration is EXHAUSTED, permanently — X3 (panel-scored quality) stays inadmissible for good, not pending a future redesign, and HB-evaluating says so plainly when family M is written. This is an internal versioned artifact, not a publication (charter §7).

The two users (Tom, 2026-09-04: "both equally from the start")

Every entry's §5 has two halves, written to the same evidence:

• For a translator — numbered instructions a person can follow at the page, each saying what it decides and what it costs.
• For a pipeline — the same guidance as steps an autonomous run can execute (what to enumerate, what policy parameter to set, what to render, what to check), with the human entry points named: the decisions a person may take over, and what changes if nobody does (charter §1, human involvement as a designed, parameterised dimension).

Index — the families, their entries, and what each consumes

The mapping below was made from the section titles of v0.2 on 2026-09-04 and is provisional: the session that writes an entry verifies each section's content, adds sections the titles hid, and corrects this table. A section may be consumed by two entries. RS- pages are in wiki/findings/results/; craft reports in wiki/findings/essays/.

The template

entries/_template.md. Every entry has the same seven sections; the exemplar is entries/HB-change-of-language.md. Cap 25 KB per entry. Front matter: type: entry, id: HB-<slug>, status: draft | active, senses:, pairs:, provisional: true, and internal-judgment-only: true where the evidence is the lead's own coding rather than a catalogued anchor.

Procedure — how an H session writes an entry

1. Read this table's row and only the sections and pages it names (grep v0.2 by heading; do not read it whole). Note any withdrawal, correction or scope line inside them.
2. Sort every claim into: survived (never withdrawn; its evidence page still supports it), scoped (survived with a narrower scope), withdrawn (say by which page). Withdrawn claims appear in §7 as one line each, never in §2–§5.
3. Write §1–§4 from the surviving claims, numbers with what they count and on what material, every one cited to its page. Descriptive findings about hands are stated as what the hands did, not as what is better.
4. Write §5 twice — translator, pipeline — from §2–§4 only. Tag each item evidenced (pairs) or untested. A pipeline step no run has executed is untested by definition.
5. Write §6: what is not evidenced (pairs, periods, provenances, the instrument's limits) and what is Open — including any successor question the record names (the family rule).
6. Apply it (the session's translation limb): a fresh 300–800-word passage that presents the problem, rendered under §5's translator half, with a followability log filed with the translation and cited in §5. Where the pipeline half can be exercised for a few cents by a panel seat as a contrast subject, do that too and say what it did differently.
7. Set status: active, update this table's row, and record the entry in §Changelog below.

Pair coverage (family Z — written at the close, S257)

Computed from the pairs: front matter of the fourteen entries on 2026-09-09 (D-20260724-04: an entry is untested on any pair it does not declare). Thirty pairs appear; English is the target in twenty-five of them. Only four pairs — FA→EN, RU→EN, JA→EN, IT→EN — are declared by six or seven entries; sixteen pairs are declared by a single entry. Every entry is untested on every pair it does not list, and no entry is evidenced on any pair with English as source except HB-emphasis-and-typography (EN→FR, EN→JA, EN→RU) and HB-footing-and-address (EN→JA).

Closing statement (family Z — written at the close, S257)

What v0.3 recommends. Fourteen entries, five active (A, G, I, H, K — each applied to a fresh passage with a followability log) and nine draft (F, E, J, L, M, B, C, D, P — consolidated at the project's close on 2026-09-09 from the v0.2 record and the result pages, independently verified against their sources, but never applied to a fresh passage, which is why they stay draft). Every entry gives guidance twice, for a translator and for a pipeline, with the human entry points named and a default for the no-human case. The instruction the handbook repeats in every family is the first item of most entries: enumerate the sites of the problem from the source, under a written rule, before opening any other translation. After that the guidance is mostly subtractive — do not invent an archaic pronoun; do not treat respelling as placeless; do not expect the free slot; do not supply a sound the source has none of unless the reader is to take it as the author's; do not add a mark the source did not license; do not add the note until the draft has been checked; do not re-render for comparison and call it independent — and it prices what it does recommend: what a choice keeps, what it costs, on which pairs it was measured.

What it marks untested. Every pipeline step in every entry (no run has executed one autonomously end to end); every instruction on every pair the entry does not declare (§Pair coverage above); every selection among policy options (the evidence prices options and ranks none); and everything about readers, since every "reader" in the record is a model seat and no human-subject data was ever collected. The regime comparison the charter calls the workshop's basic experiment (R06 vs R04 vs R03 on prose, W2 step 5) was never run, and R03 was never specified.

What it withdrew from v0.2. Each entry's §7 lists its withdrawals with the withdrawing page. The pattern across the record: v0.1's one rule (R1, displaced marking) kept its text but lost its headline prediction (retired at v0.2, replaced by S1: the loss R1 repairs is uncommon); v0.2's positive instructions about sound, rhyme and radif were withdrawn or narrowed one by one (§7.12 → §7.14; §7.15 → §7.17; §7.16 → §7.20; §7.19 → §7.21; §7.24 → §7.25 → §7.29; §7.36.1 and §7.40 refuted by Payne 1901 and Leaf 1898; §7.44 → §7.49.2; §7.49.1's "factor of eleven" → about 5.5 at RS-20260908-chime-slot); the general register recommendation was refused four times and replaced by four sharper refusals; and the placeless-baseline claim of §7.3 was withdrawn and replaced with a number. Nothing in the handbook rests on a jury score: Tier D calibration is EXHAUSTED (config/models.md), X3 is inadmissible for good, and every page carries provisional: true.

What a successor would do first. Apply the nine draft entries to fresh passages with followability logs (the procedure's step 6), one per session, and set them active or amend them; then the end-to-end application to one whole prose work (W3 step 6) that the W1 finish line requires and the close did not reach.

Changelog

• v0.3.5 (2026-09-09, S257 — the close-out) — the nine remaining families written as consolidations of the record by parallel agents following §Procedure steps 1–5 and 7, each entry then adversarially verified against its cited sources by an independent agent and repaired: F HB-mimetics, E HB-sound-in-prose, J HB-emphasis-and-typography, L HB-long-works, M HB-evaluating, B HB-rhymed-prose, C HB-radif-and-refrain, D HB-meter-and-line, P HB-pipeline. All nine carry status: draft because step 6 (a fresh 300–800-word passage translated under the entry with a followability log) was not done; their Standing paragraph says so and their §5 Application reads "not applied". Family Z written: the §Pair coverage table and the §Closing statement above. Index rows updated. Shared consumption noted by the entries themselves: HB-mimetics also consumes v0.2 §7.13's mimetic-class finding (listed under H); HB-radif-and-refrain also consumes §7.32's struck sentence, §7.43.1/§7.43.3, §7.47.2 and §7.49.1's radif stratum; HB-rhymed-prose takes only the craft finding of §7.34 (family K's) and shares §7.39 with K. The project was wound up the same day (wiki/close-out.md).
• v0.3.4 (2026-09-08, S255) — HB-telling-the-reader (family K) written and set active: consolidates v0.2 §7.34, §7.37, §7.38, §7.39, §7.43.2, §7.50.6 into four kinds of printed translator (or author) statement that behave differently — a refusal, checkable and twice confirmed on the page; a promise, unchecked and found empty at every enumerated locus in the one case measured (though not proven empty everywhere); a comparative numerical self-report (Leaf on his own rhyme-word frequency), which held against a registered prediction that it would fail; and a disclosure of a source fact, which moves reader preference toward a source-carrying rendering where showing the source itself does not. An author's own printed language policy (Dante's) binds no translator, including the one who reproduces it. §7.34's own both-orders statistic is explicitly flagged as withdrawn (not replicated at RS-20260827b), while its direction stands — the entry states this rather than reusing the old number. Translation limb: T-chumon-R04-v1, a fresh, compressed extent of Miyazawa Kenji's 注文の多い料理店 (JA→EN, the project's first Miyazawa Kenji translation limb), applying the entry's guidance to strike a drafted footnote once the drafted English was checked and found to carry the title's pun unaided, with a followability log.
• v0.3.3 (2026-09-07, S252) — HB-register (family H) written and set active: consolidates v0.2 §7 (Q-e), §7.1–§7.3, both §7.4 headings (a numbering defect — two unrelated S156/S178 entries shared one number) and §7.11, §7.13 into the finding that a general register-carriage recommendation was refused four times running, each refusal replaced by a sharper reason, ending in: no lowering device is placeless (nonstandard spelling reads as a specific region; located idiom is barely taken even under permission); no published hand approaches an explicit rule's source-ward register rate; a light, targeted elevation policy reads as raising the source's own low points, a wholesale one raises everywhere; and a source-ward device's fidelity payoff to a reader depends on whether the reader can see the source. Two RS pages (RS-20260809g-device-cross, RS-20260813g-register-quadrants) were missing from this family's row in the index above and are added. Translation limb: T-malavoglia-i-conscription-R04-v1, a fresh Verga excerpt (IT→EN, a new disjoint extent of I Malavoglia ch. I) applying the entry's no-eye-dialect, discrete-marker guidance, with a followability log.
• v0.3.2 (2026-09-05, S249) — HB-realia (family I) written and set active: consolidates v0.2 §7.5, §7.6, §7.9, §7.10 (four machine-scored experiments, RU/PL/JA/ZH/SR/CS/EL→EN) plus two Tier 1 precedent anchors (A-shaw-spider-thread, A-garnett-vanka) into the finding that native-sounding prose and a preserved sense of the source's setting are two different channels no measured strategy buys at once, with the option set splitting on proper name vs. common noun. Translation limb: T-max-havelaar-i-c-R04-v1, a fresh Multatuli excerpt (NL→EN, the project's first Dutch translation limb) applying the entry's own guidance, with a followability log; a new source manifest (workshop/canon/max-havelaar-i-c/) and a third disjoint extent of the book's chapter I.
• v0.3.1 (2026-09-04, S245) — HB-footing-and-address (family G) written and set active: consolidates v0.1 whole (R1, its four device-category findings, the traceability sites) and v0.2 §2–§6, §10 (S1, ARM-carrier, ARM-footing, ARM-honorific-hands, ARM-supplied-footing) plus thirteen RS- result pages. Translation limb: T-hirurgiya-R04-v1, a fresh Chekhov excerpt (RU→EN) applying the entry's own guidance, with a followability log.
• v0.3.0 (2026-09-04, S244) — constituted on Tom's direction; the index, the template, the procedure, and the exemplar entry HB-change-of-language (family A) written; v0.2 frozen as the record at v0.2.52.