Machinery for constrained homeosyntactic transposition: hold a text's syntactic form at a chosen degree while inventing entirely new content, then measure — with an automatic gate — which of a literary critic's formal claims actually survive the move.
The worked case is the incipit of Manzoni's I promessi sposi (1840), transposed into the border/migration landscape of the central Mediterranean. It is the proof-of-concept for the "Inverse Queneau" method (ACH 2026, Marchesini & Franchi): where Queneau holds content and varies form, we hold form and vary content, to reveal how form naturalizes meaning.
Talk & slides. Marchesini & Franchi, The Proof is in the Pudding: Structural Experimentation as Close Reading Strategy in the Age of LLMs (ACH 2026). Deposited open access (CC BY-SA 4.0) on Humanities Commons CORE: https://doi.org/10.17613/1p8bw-67y06.
Marchesini & Franchi, Keeping It in Shape: Semantic Transpositions of a Form-Bound Text
(Computational Humanities Research 2027) draws on two parts of this repository: the
Manzoni stress test (§5, Appendix A, and the discourse-type control) here in the root,
and the Queneau ladder (§4, Appendices C and D, Tables 2–6) in queneau/,
which has its own README with a table-by-table map and a run-by-run PROVENANCE.md.
Notation. The paper names the degrees D₀ (ordered tree), D₁ (order-free dependencies),
D₂ (clause nesting), D₃ (clause inventory) and D₄ (depth). The code below predates those
names: here D2 is the clause architecture taken whole (inventory and nesting) and D3 the
hypotaxis profile, and in the Queneau code the paper's D₂ is D2c and its D₃ is D2.
queneau/README.md gives the full table.
Every run is a point in a 2-D space:
-
Degree axis — how much form is held. A single dependency parse is projected to successively coarser targets — a partial order, not a chain:
D0— the full ordered dependency tree: POS + linear order + relations. The top.S— the surface: the ordered part-of-speech sequence (the tree's leaves). Forgets the arcs. The calque / Oulipian end.D1— the dependency relations, order-free. Forgets the order.D2— the clause architecture (inventory and nesting). The operative rung.D3— the hypotaxis profile only (clause count + max depth). The loosest.
SandD1are two independent forgetful projections ofD0— one drops the arcs, the other drops the order — so they are incomparable: neither is "more form" than the other. (Hence the warning against writingS ⊃ D0 ⊃ D1as a line: that would smuggle the kind of structure into the amount of structure and collapse the two axes into one.)D2andD3coarsenD1further. The metrics mirror the rungs:s0_pos↔S,d1_triple↔D1,d2_clause↔D2,depth↔D3. -
Constraint-load axis — which critical claims are additionally imposed: from
0(pure homomorphism — form only) up tofull(the eight critic-derived patterns + the threefold criterion = the constrained homeomorphism). This axis is orthogonal to the degree axis: you can demand any pattern set at any degree.
Corners of the space: (D0 × 0) ≈ a near-copy of the syntactic skeleton (iso-syntacticism);
(S × full) ≈ the calque (surface locked, patterns can't rescue coherence — see the
results); (D2 × full) = the critical instrument used here; (S × low) = the
Borgesian-Oulipian generator.
A transposition is accepted only if all three hold together:
- Form preserved at the chosen degree —
clause-F1high (for D2). - Divergent — a genuinely new scene, no lexical calque —
s0_pos,lemma_jacclow,frame_divhigh. - Coherent — one continuous, sensible description —
coherencehigh.
score_production.py computes all five axes per candidate and the combined SCORE.
docs/Phase_1_Workbook_Manzoni.md ← critic's claims, provenance (the rubric's source)
│ build_rubric.py ← edit + (re)generate + validate the rubric
▼
spec/{patterns,constraints,constraints_en}.json
│
│ spec/brief_<lang>.json ← the transposition DOMAIN (docs/Target_Domain_Brief.md)
│ │
▼ ▼
build_specs.py --lang it --degree D2 ← project form + fold in the brief
▼
spec/specs_it_D2.json
│ run_homeo.py --degree D2 --backend … --model …
▼
results/production_<label>_<tag>_<model>.json ← the generations
│ score_production.py
▼
results/production_scored_*.json ← the 3-axis gate
The transposition domain is set in spec/brief_<lang>.json — a domain line plus the
scene-world elements the generator must deploy (in its own words) while holding the form.
It is language-specific: build_specs.py exits with an error if a brief's lang field does
not match --lang. To transpose into a different world, edit the brief and rebuild. See
docs/Target_Domain_Brief.md for what goes into it.
pip install -r requirements.txt # Stanza, anthropic, etc.
python -c "import stanza; stanza.download('it')"
cp .env.example .env # then add your API keys (see below)
python scripts/build_rubric.py # validate/emit the rubric
python scripts/build_specs.py --lang it --degree D2 # → spec/specs_it_D2.json
python scripts/run_homeo.py --backend anthropic --model claude-opus-4-8 \
--lang it --degree D2 --n 6 --tag opus48
python scripts/score_production.py # → results/production_scored_*.jsonrun_homeo.py accepts --degree S|D0|D1|D2|D3 and three backends (anthropic,
openrouter, vllm-http); the same rubric drives every degree. The Opus/Italian
gradient over all five rungs is in results/ — the two tight rungs (D0, S) drive
coherence to zero, which only recovers from D1 down.
The cloud backends read two keys from a .env file in the repo root (git-ignored —
never commit it). Copy the template and fill in only what you need:
cp .env.example .env
# edit .env:
# ANTHROPIC_API_KEY=sk-ant-... # Opus generation + the frame/coherence LLM-judges
# OPENROUTER_API_KEY=sk-or-... # open-weight models (Qwen, Kimi, …) via OpenRouter--backend anthropicneedsANTHROPIC_API_KEY.--backend openrouterneedsOPENROUTER_API_KEY(provider/quant pinned — seeMANIFEST.md).score_production.pyalways needsANTHROPIC_API_KEY: the frame-divergence and coherence judges call Claude no matter which backend produced the candidates. (SetSEM_JUDGE_MODELto override the default judgeclaude-sonnet-4-6.)
Local / self-hosted models (--backend vllm-http) need no key — point --endpoint
at any OpenAI-compatible /v1/chat/completions server:
# stand up your own (example): vLLM serving an instruct model
python -m vllm.entrypoints.openai.api_server --model <hf-model> --port 8000
# then generate against it
python scripts/run_homeo.py --backend vllm-http --endpoint http://127.0.0.1:8000 \
--model <hf-model> --lang it --degree D2 --n 6 --tag localThe ACES SLURM scripts in scripts/ (qwen_aces.sbatch, maverick_aces.sbatch) are
worked FP8-MoE examples of standing up such a server on H100s.
| dir | contents |
|---|---|
scripts/ |
the pipeline (build_rubric, build_specs, run_homeo, score_production, the metric libraries, ACES SLURM scripts) |
spec/ |
the rubric (patterns/constraints) + the transposition brief (brief_{it,en}.json) + the built Italian form-targets specs_it_{S,D0,D1,D2,D3}.json |
texts/ |
the public-domain Italian source (Manzoni 1840) |
results/ |
the generations + scored outputs + metric-validation data |
docs/ |
the rubric workbook + Pipeline_Scripts.md, Data_Files.md, Tools_and_Stack.md |
queneau/ |
the Queneau ladder of the CHR 2027 paper: scripts, specs, every generation and score (see its README) |
Runs span Opus 4.8 (Anthropic), Qwen3-235B (OpenRouter and an ACES H100 self-host, fp8),
and Kimi-K2 1T (OpenRouter, fp8). OpenRouter routing is pinned per provider with
quantizations:["fp8"] and allow_fallbacks:false, and the resolved provider is recorded
in every result file. See MANIFEST.md for slugs, parameters, and environment, and
docs/Tools_and_Stack.md for the full stack.
The Italian source is public-domain (Manzoni, 1840) and is included. The English source used
in the paper (the revised Colquhoun translation, ed. Reynolds) is in copyright: only the
incipit we analyse (texts/incipit_en.txt, about 680 words) is included, as quotation for
scholarly analysis, with the specs and results derived from it; the complete translation is
not. The same holds for the Queneau passage and its two translations in queneau/.