Study #5 on the modelun harness. The consensus census measures which answer a model picks when many are valid ("Name a tree.") in plain chat. This study repeats the census with one change: the reply format. "Name a tree. Reply with JSON only, in the form {"word": ""}." No schema enforcement, no constrained decoding, no token masking — just the request. The question is whether the serialization register shifts which answer gets chosen.

1. The format tax is progressive. Field mean surprisal drops 1.80 → 1.58 bits under JSON, and the drop concentrates on the divergent tail: deepseek-v3.2 −1.31 (the census's lone frontier-lab explorer loses half its personality), hermes −0.91, gpt-4o-mini −0.92, gpt-4-turbo −0.87, gpt-5.6-sol −0.65 — while the conformist floor barely moves (sonnet-5 −0.05, grok-4.5 −0.07). The scorecard's spread compresses from ~2.2 to ~1.5 bits. On "Pick a word.", serendipity goes from 40% of the field to 63%, distinct answers 53 → 28.

2. Mostly a sharpener; re-indexes only where the plain mode is weak. The JSON field mode matches the plain mode in 28/31 categories, <4% of mass lands on answers the plain census never saw, and the tail thins. The three flips are all weak-mode categories — insect: ant 43% → butterfly 60%; board_game: chess → monopoly; dance: salsa → tango — and the new modes smell like the example-payload canon of documentation corpora. β-story reading: the register amplifies its own conditional prior, which only wins where the chat prototype was weak.

3. Defaults are register-indexed — the register erases about half of them. Across the panel, models hold 144 stable off-modal defaults (same non-modal answer four-of-four in chat); only 47% survive into JSON. Retention is a per-model trait: gpt-5.6-terra loses its signature defaults (mustard→ketchup, mango→apple, four-of-four flips) keeping 1/4; claude-fable-5 keeps 5/7 (mustard, mango survive; gouda→cheddar does not) — and acquires five new JSON-only stable defaults (cerulean for color, carpenter for occupation, 4-of-4 in the register, never in chat). llama-4-maverick holds the panel's largest defaults set (10) and keeps 6, which is what its positive Δ is made of. So "conviction vs costume" is a continuum, and the true object is a per-register defaults profile — a model's personality is indexed to the channel, the same way the census found answers indexed to wording. (All read from deterministic 4-of-4 answer behavior; the surprisal deltas for individual models are n.s. — see finding 5.)

4. The register gradient (full-battery, compliance-conditioned, permutation-tested). JSON and XML compress the field and the effect is highly significant on both exchangeable units (field entropy over 31 categories: −0.20 bits, p=.0006; per-model Δ-surprisal over 44 models: −0.22, p=.0002; XML similar). YAML and CSV show no significant net effect (p=.75/.35 by model) — YAML's dramatic any-word concentration (65% serendipity) is real but category-specific, washing out battery-wide. Brackets significantly loosens the field (+0.12 bits entropy, p=.014; +0.13 per model, p=.009): an arbitrary non-data wrapper makes the field more varied. The json-vs-yaml gradient is itself significant (paired by category, p<1e-4). Compression is thus specific to the answer-delivery formats models are trained to speak — which favors the tool-use-post-training story over pure corpus register, though the any-word result keeps both alive. Tests: probe_significance.py (sign-flip permutation, 20k draws). Two companion phenomena: format incompetence (granite/mythomax emit CSV wrappers ~1% of the time; their unwrapped replies read as spurious divergence — always condition on compliance) and a genuine register- diverger (llama-4-maverick: 100% compliant in json/csv and still +0.4 to +0.7 in every format). Parse survival ≥99% per column after junk-guarding, plus an echo guard that drops the category noun / template placeholder (a literal [city], answer, word) as a non-answer on every column alike — it trims the raw brackets loosening from +0.16 to +0.13.

5. The Δ ordering has exactly three tiers of meaning — read it that way. (a) Six models have individually significant Δs (BH-FDR q=.10), all compressions: deepseek-v3.2 −1.31, gpt-4o-mini −0.92, gpt-4-turbo −0.87, gpt-5.6-sol −0.65, qwen3 −0.45, gemini-3.1-pro −0.31 (hermes −0.91 now just misses the threshold). (b) One family-level effect: the Llama-derived group (maverick, llama-3.3, mythomax [a community Llama-2 merge]) diverges under JSON — no member significant alone, but pooled Δ +0.44 (p=.009) and +0.70 vs the rest of the field (two-sample permutation, p<.0001, surviving a ×19 family-selection penalty). Interpretation is open: hermes-4 (also Llama-based, heavily Nous-retuned) compresses hardest of anyone, so base weights alone don't explain it — but with community derivatives splitting both ways (mythomax up, hermes down), a clean "follows the tuner" story doesn't hold either. The mechanism is unresolved. (c) The remaining ~38 models are a noise plateau (Δ range −0.67…+0.56, none distinguishable from zero individually): their ordering means nothing, exactly the census's tiers-not-ranks rule applied to deltas. The conviction/costume contrast (fable vs terra) rests on discrete 4-of-4 answer behavior, not on Δ magnitudes, and is unaffected.

6. The register moves the mode, not the temperature. Field-mean self-distinctness is nearly flat across columns (plain 0.42 → json 0.39) while surprisal drops 0.22 bits — the collapse is positional, not a sampling change (and so not a serving-layer artifact). The within-model narrowing that does exist is progressive like everything else: hermes 0.71 → 0.52, deepseek-v3.2 0.56 → 0.42, while the conformist floor and fable (0.32, conviction ≠ noise) don't move.

Contrast with the census's "unusual fruit" probe: a semantic reframe re-indexes (new column, new mode — durian), while a channel reframe sharpens (same column, higher peak). The question picks the distribution; the register sets its peakedness — mostly by moving mass onto the mode, not by cooling the sampler.

The structured-output literature measures accuracy under format requests (Let Me Speak Freely, EMNLP 2024 — contested by the dottxt rebuttal), validity-vs-correctness under decoder-level enforcement (the 2026 "constraint tax" line), and engineering fixes for diversity within grammars (automata steering). Nobody measures which answer gets chosen under a mere format request, across a cross-vendor panel. Prior-art map: section 9 of the modelun bibliography. Note both obvious names are taken — "constraint tax" (arXiv:2605.26128, 2606.25605) and "The Format Tax" (arXiv:2604.03616) — and both refer to enforcement/accuracy effects, which this study deliberately is not.

• Baseline: the consensus study's plain-chat transcripts (same 31 questions, same 44-model roster, same conditions: no system prompt, temp 1.0, 4 runs).
• JSON: full 31-category battery, all 44 models × 4 runs.
• XML, YAML, CSV, brackets: each runs the full 31-category battery too (all 44 models × 4 runs), to separate serialization from mere structure across the whole panel.
• Metric: the census's answer-choice surprisal, computed per format column; the per-model Δ-surprisal (JSON − plain) is the headline "how much personality does the register erase" number. Junk guard + parse: format wrapper stripped mechanically (regex per format), then the census normalization applies, plus an echo guard that drops a reply echoing the category noun or the template placeholder ([city], answer, word).

Wrapper-hit rate per model per format falls out of every run for free, and it is kept deliberately: (1) with one-word content, a missing wrapper is pure format incapacity — unlike IFEval/FOFO-style evals, where format and task difficulty entangle; (2) it is the mandatory control for any distributional claim (non-compliance masquerades as diversity: granite's 1% CSV compliance read as +1.49 bits of "personality" unconditioned); (3) it is a generational capability track — gpt-3.5-turbo speaks YAML at 19% where modern models are ~100%, so format registers have acquisition dates the re-run battery will record. Analogous to self_distinct doubling as the effective-temperature proxy in the census: wrapped/n doubles as the format-capability scorecard.

../../.venv/bin/python run_formats.py     # -> probes/format_register.json (raw replies)
../../.venv/bin/python analyze.py         # comparison vs ../consensus transcripts

Status: findings frozen (2026-07-11, after the finding-5 correction); committed. Data collected against the 44-model roster (2026-07-10). Paper draft in paper/main.tex (abstract + full body prose); blog draft in convovo-site. Depends on ../consensus/ for the stimulus, roster, baseline transcripts, and probe_lib.py.