# Critique Memo: The July 2026 IACC Strategic Plan Working Draft, Measured Against the IACC Record (2007–2023)

**Date:** 2026-09-16
**Prepared by:** Mike
**Sources:** `iacc_linguistic_comparison.ipynb` (every figure below is regenerated by one run of it); outputs in `outputs/iacc_linguistic/`; draft PDF `iac/IACC Strategic Plan Working Draft July 17.pdf`; IACC publications archive `iac/IACC_Publications/` (downloaded from iacc.hhs.gov 2026-09-15)

---

## 0. What this memo is

A consolidated, verified account of how the July 2026 IACC Strategic Plan working draft differs from every document the Interagency Autism Coordinating Committee published between 2007 and 2023 — linguistically, lexically, and structurally — with plain-language explanations of each measurement method, and an interpretation section that separates **measured fact** from **inference**.

**Corpus:** 46 documents in the comparison set — the draft plus 8 earlier strategic plans (2009–2023) and 37 other IACC publications (Summaries of Advances, Reports to Congress, portfolio analyses, letters). 15 inventoried documents were excluded from comparisons as too short or as easy-read/at-a-glance editions (different register, too small for stable statistics).

**Verification status:** every headline number below was cross-checked twice — (a) the report against its own CSV exports, and (b) independently, by re-extracting the draft and the 2023 plan with a different PDF tool (poppler `pdftotext` rather than the notebook's `pymupdf`) and re-counting. Raw and cleaned rates are flagged as such. Raw-text comparisons include reference lists and are used only to check direction and magnitude.

---

## 1. The methods, in plain language

Each measurement is defined here once, then used freely below.

- **Per-1k / per-10k (normalised rate).** A count divided by document length, expressed per 1,000 or 10,000 words. This is the "normalization": it makes an 80,000-word draft comparable to a 10,000-word 2009 plan. "Should: 10.5 per 1k" means the word *should* appears about 10–11 times in every 1,000 words.
- **Flesch Reading Ease (FRE).** A 0–100 readability score from average sentence length and syllables per word. Higher = easier. 60–70 is a news article; 30 is a university textbook; **below 10 is extreme** — dense legal or scientific prose. Formula: `206.835 − 1.015×(words/sentence) − 84.6×(syllables/word)`.
- **Flesch–Kincaid grade (FK).** The estimated US school grade needed to follow the text, from the same two ingredients. 12 = high-school senior; 16 = college graduate; **18+ = postgraduate**.
- **Coleman–Liau, Gunning Fog, SMOG, ARI.** Four other readability formulas using word/sentence length instead of syllables. Used here as corroboration, not headline.
- **MATTR (moving-average type–token ratio).** Lexical diversity: slide a 500-word window through the text, count how many distinct words appear in each window, average. Higher = richer, less repetitive vocabulary. Unlike raw type–token ratio it is not biased by document length.
- **Nominalisation percentage (`nominalisation_pct`).** The share of words ending in -tion/-sion/-ment/-ness/-ity/-ance/-ence/-ism — verbs and adjectives frozen into nouns ("we decide" → "a determination was made"; "analyse" → "analysis of the analysis"). Nominalisation is the classic marker of bureaucratic and scientistic register: it hides actors ("who decides?") inside abstract nouns.
- **Deontic index.** Among all modal verbs, the share expressing obligation: (must + shall + should + "required/need to") ÷ all modals (incl. will/would/may/might/can/could). High = the document *commands*; low = it *describes or predicts*.
- **Hedges / boosters.** Rates of uncertainty markers ("may, suggests, preliminary, unclear") versus certainty markers ("clearly, demonstrates, essential, definitive").
- **Passive constructions per 100 sentences.** A heuristic count of be-verb + participle ("was conducted"). Passive voice hides who acts.
- **Z-score vs earlier plans.** The draft's value minus the earlier plans' mean, divided by their standard deviation. |z| > 2 is the conventional "unusual" line. **Caveat:** there are only 8 earlier plans, and they are stylistically consistent, so their standard deviations are small and z-values are inflated. The *direction* of every z reported here is trustworthy; treat the *magnitude* as "far outside the range," not a precise distance.
- **Keyness (weighted log-odds, Monroe et al. 2008).** For each word: how much more (or less) the draft uses it than the earlier plans, as a log-odds ratio with its standard error, stabilised by a prior so rare words don't dominate. Reported as a z-score: red = distinctively draft, blue = distinctively earlier plans.
- **TF-IDF cosine (content similarity).** Each document becomes a vector of its content words/bigrams weighted by how distinctive they are; cosine (0–1) measures topical overlap. Higher = same subject matter.
- **Burrows' Delta (style distance).** Uses only the 150 most frequent words — mostly function words (the, of, should) — z-scored per document, mean absolute difference. A standard stylometric/authorship measure of *how* a text is written, regardless of topic. Higher = more different style.
- **8-word shingles (text reuse).** Slide an 8-word window through the text; the share of those windows found verbatim in earlier documents measures how much of a text is inherited boilerplate or copied passage.
- **Spearman trend (term_trends).** Rank-correlation between a word's rate and publication year across the 45 pre-draft documents. A word "rises" or "falls" across 2007–2023; the draft is then judged to **continue** or **reverse** each trend relative to the 2018-onwards average.

---

## 2. Measured findings

### 2.1 Size and shape

| Measure | Draft | Earlier plans (mean ± sd) | z |
|---|---|---|---|
| Words (cleaned prose) | 81,788 | 32,326 (max 60,494, SP-2023) | — |
| Flesch Reading Ease | **0.57** | 27.9 ± 8.2 | −3.3 |
| Flesch–Kincaid grade | **17.9** | 14.6 | elevated (within ±2 range) |
| Coleman–Liau | 21.6 | 15.9 ± 1.1 | +5.0 |
| Syllables per word | 2.21 | 1.86 ± 0.05 | +6.8 |
| Letters per word | 6.62 | 5.64 ± 0.17 | +5.9 |
| Words of 7+ letters | 49.4% | 37.9% ± 2.2 | +5.4 |
| Polysyllables (3+ syll.) | 35.4% | 24.0% ± 1.5 | +7.7 |
| Nominalisation % | 11.77 | 7.55 ± 0.77 | +5.5 |
| MATTR-500 | 0.551 | 0.490 ± 0.01 | +5.8 |

The draft is the most lexically diverse, longest-worded and most nominalised document in the record. On readability the honest statement is narrower than "least readable": the draft is the most extreme document in the corpus on Flesch Reading Ease (0.57) and Coleman–Liau, the two formulas that weight word length most heavily, but it ranks **second or third behind SP-2023** on Flesch–Kincaid (17.9 vs 18.5), Gunning Fog (21.8 vs 22.0), SMOG (18.0 vs 18.8) and ARI (19.4 vs 20.6), because its sentences are shorter (19.2 words vs 28.4). What is unambiguous is word length: 49.4% of its words run 7+ letters against 40.6% in SP-2023, and 35.4% are polysyllabic against 24.0% across earlier plans — both the highest in the record. FRE 0.57 is also an extrapolation below the calibrated floor of that scale and should be quoted as "at the bottom of the scale", not as a reading level. Independently confirmed by raw re-extraction: words of 12+ letters make up **7.9%** of the draft vs **5.0%** of SP-2023; of 10+ letter words, **38%** are nominalisations vs **31%**.

**Appendix C is the register at its limit, and it is a formatting failure rather than a prose one.**
Appendix C (pp. 270–271) is the draft's highest-grade section at **FK 30.6** against the document's
17.9, and the highest of all 37 sections; 7 sections sit above FK 20 and the median is 18.0. The
mechanism is specific: it is a bulleted list with the bullets removed. The "Communication-Adapted
Clinical Protocol" block runs **373 words in 5 sentences — 58 semicolon-delimited items, a longest
sentence of 121 words, and 1.1 finite verbs per 100 words** against 3.6 across the draft. Semicolons
run at 15.3 per 100 words in that block, against 0.9 across the draft and 0.1 in SP-2023. The draft
uses 346 bullet glyphs elsewhere, so this is a typesetting choice, not an artifact of PDF extraction.
Across the whole draft there are **26 such blocks, 1,790 words (2.1% of the text), 207 enumerated
items**; the list is in `list_as_prose_blocks.csv`.

Two consequences, pointing opposite ways. For the committee: this is the most actionable comment in
the package, because the fix is free — set the protocol contents, the escalation criteria, the minimum
data set (22 fields inside one 83-word sentence) and the technology criteria as numbered lists. As
typeset, a clinician cannot check whether an item is present and an implementer cannot enumerate the
data fields, which is what an appendix of this kind exists to permit. For the analysis: this register
is *not* what drives the readability result. Removing every one of the 26 blocks moves the draft's
grade level by 0.1, so the finding rests on the main prose, and a critic cannot dismiss it as an
appendix artifact.

The most-used 12+ letter words, side by side:

- **Draft 2026:** implementation (348), communication (303), developmental, gastrointestinal, participation, translational, accountability (116), mitochondrial, surveillance (90), coordination, consolidated, stabilization, standardized, longitudinal, inflammatory.
- **SP-2023:** interventions (314), disabilities, communication, developmental, co-occurring, environmental, understanding (92), opportunities, participants, evidence-based, perspectives, underrepresented, autism-related.

### 2.2 Voice: from describing to directing

| Measure | Draft | Earlier plans | z |
|---|---|---|---|
| "should" per 1k | **10.53** | 1.07 ± 0.30 | **+31.1** |
| Deontic index | **0.66** | 0.21 ± 0.05 | +8.9 |
| "will" per 1k | 0.23 | 2.76 ± 0.84 | −3.0 |
| Hedges per 1k | 4.8 | 7.7 | (notable) |
| Boosters per 1k | 2.6 | 2.2 | (flat) |
| First-person plural per 1k | 0.09 | 1.19 | (notable) |
| Passives per 100 sentences | 18.4 | 25.1 | (lower in draft) |

Earlier plans *survey and predict* ("will," hedged findings, a collective "we"). The draft *instructs*: "should" is ~10× denser than the earlier-plan mean — the single most extreme measurement in the study. The draft is grammatically a directive to federal agencies, not a committee's survey of a field.

### 2.3 Naming autism (per 10k words; draft vs earlier-plan mean)

| Expression | Draft | Earlier plans |
|---|---|---|
| ASD (case-sensitive) | **1.2** | 156.9 |
| Person-first ("people with autism") | **2.6** | 54.8 |
| Identity-first ("autistic people") | 22.0 | 9.2 |
| "on the (autism) spectrum" | **0.0** | 9.9 |
| "autism spectrum disorder(s)" | 0.1 | 5.2 |
| "disorder(s)" | 4.8 | 21.3 |
| "condition(s)" | **26.9** | 15.6 |
| **"profound autism"** | **5.5** | **0.0** |
| nonspeaking / minimally verbal | 3.5 | 2.2 |
| neurodivers* | 0.0 | 0.2 |
| vaccin* | **0.0** | 4.5 |

**The naming debate is the wrong frame, and the eight-plan mean is the wrong baseline.** Person-first usage has been falling across autism writing for a decade by deliberate community preference, so its collapse here is partly a field-wide trend the draft inherits rather than a choice it makes; and identity-first usage in the draft is *above* the eight-plan mean but **2.7× below SP-2023** (22.0 vs 58.7 per 10k). The reported "90% identity-first share" is therefore a share of a shrinking denominator, not evidence of an identity-first turn.

Counting every naming convention together removes the problem. Measured as a single non-overlapping pass over cleaned text — person-first phrases, identity-first phrases, "on the spectrum", self-advocate/autistic-community, then bare *autism* and *ASD* — the total rate of explicit reference to autism or autistic people has been remarkably stable for seventeen years, and the draft is far outside it:

| Plan | 2009 | 2010 | 2011 | 2012 | 2013 | 2017 | 2019 | 2023 | **2026 draft** |
|---|---|---|---|---|---|---|---|---|---|
| all autism references /10k | 314 | 286 | 261 | 256 | 199 | 264 | 285 | 278 | **108** |
| of those, references to autistic *people* /10k | 79 | 86 | 72 | 56 | 39 | 78 | 99 | 113 | **28** |
| federal agency acronyms /10k | 9 | 15 | 32 | 22 | 20 | 8 | 30 | 12 | **167** |
| agency names per autism reference | 0.03 | 0.05 | 0.12 | 0.09 | 0.10 | 0.03 | 0.11 | 0.04 | **1.54** |

Eight plans spanning 2009–2023 sit in a band of 199–314 references per 10,000 words (mean 268, sd 33). The draft sits at **108 — 4.8 standard deviations below that mean, 2.6× below SP-2023, and 1.8× below even the lowest earlier plan (SP-2013, 199 per 10k).** Both components fall: person-referring mentions from 113 to 28 per 10k, and autism-as-topic mentions from 164 to 81. The finding does not depend on which label is in fashion, on the naming debate, or on the choice of baseline — it holds against SP-2023, against the plan mean, and against every individual plan.

What fills the space is federal machinery. Every plan from 2009 to 2023 named autism between 8× and 36× more often than it named a federal agency; **the draft names agencies 1.5× more often than it refers to autism at all**, and collective referents (population, cohort, subgroup, subtype) outnumber references to autistic people by 49.5 to 27.1 per 10k — an inversion of SP-2023, where person references outnumbered collective ones roughly 4:1. This is the strongest single result in the study and the one least exposed to methodological objection.

Independent raw cross-check against SP-2023 (uncleaned text; reference lists included — direction/magnitude only):

| Expression | SP-2023 /10k | Draft /10k |
|---|---|---|
| autistic individual(s) | 14.7 | **8.3** |
| autistic adults/children | 11.4 | **8.0** |
| autistic people | 4.1 | 4.3 |
| spectrum (any use) | 69.9 | **26.9** |
| on the spectrum | 17.4 | **0.9** |
| profound autism | 0.0 | **4.6** (50 occurrences) |
| nonspeaking / non-speaking | 0.5 | **2.8** |
| minimally verbal | 0.5 | 1.4 |

**The label the community prefers was adopted; the individuated person was written out.** Mentions of "autistic individuals/adults/children" *fell* versus SP-2023; "autistic people" held flat; the spectrum frame — the accepted diversity framing — was effectively deleted; and "condition" replaced "disorder" as the default noun.

### 2.4 What is to be done: treatments up, interventions down, medications up 10×

Per-10k, cleaned (draft vs earlier-plan mean): implementation 39.0 vs 4.1; service(s) 28.2 vs 62.1; support(s) 61.7 vs 33.2; intervention(s) 20.9 vs 46.9; biomarker(s) 24.8 vs 5.6.

Raw cross-check vs SP-2023 (direction/magnitude only):

| Term | SP-2023 /10k | Draft /10k |
|---|---|---|
| treatment(s) | 6.3 | **19.9** |
| intervention(s) | 45.4 | **19.6** |
| medication(s) | 2.0 | **20.0** |
| prescrib* | 0.5 | 3.2 |
| dose / dosage | 0.5 | 1.4 |
| placebo | 0.7 | 0.7 (flat) |
| behavioral | 7.6 | 9.0 |

The swap is specific: **medication vocabulary rose tenfold while interventions halved and placebo stayed flat** — this is pharmacology talk, not generic clinical-trial talk. Keyness bigrams corroborate: "medication effects," "immune metabolic," "trials design," "clinically meaningful," "biomarker-defined." **Leucovorin — 24 mentions — has never appeared in any earlier IACC publication.**

### 2.5 Causes and biology (per 10k; draft vs earlier-plan mean)

immun* 27.8 vs 5.9; regression 21.6 vs 1.4; safety/wandering/elopement 29.8 vs 7.6. New therapeutic-domain architecture: nine "Priority Therapeutic Domains" covering neurotransmission, immune/autoimmune, folate metabolism (leucovorin), GI/microbiome, neurodevelopmental regression, mitochondrial/redox, autonomic dysfunction, sleep/circadian, epilepsy — plus motor planning/praxis and adaptive trial design.

**"Profound autism" as policy instrument.** The draft proposes "adoption of a **standardized functional designation of profound autism for research and policy purposes**," and couples the category to prevalence: "the share of children with profound autism — minimal or no functional speech and who require continuous care — **has risen alongside the broader trend**." The term appears ~50 times; it never appeared in the entire 2007–2023 record. The rhetorical effect is to convert rising diagnosis counts into rising *severity* claims.

### 2.6 Governance vocabulary (per 10k; draft vs earlier-plan mean)

Agency acronyms (HHS/NIH/CDC/FDA/CMS…) 166.9 vs 18.6 (9×); federal 71.2 vs 10.0; **IACC itself 4.5 vs 38.0** — but that figure needs its SP-2023 comparator: the eight-plan mean is carried by SP-2019 (70.5) and the 2009–2010 plans (~52), while **SP-2023 was already down to 15.7**, so the draft's additional decline is 3.5×, not 8.4×, continuing a fall that began before it (an independent re-count on raw text with running headers stripped symmetrically gives 3.9 vs 10.5 per 10k). New coinages never before seen in the record (176 words, 5+ uses): dashboard (42), workplans (42), compact (39), comparative-effectiveness (38), NAPTI (38), biomarker-defined (37), implementation-ready (34), praxis (32), tier (31), prescribing (21), leucovorin (18), deterioration (18), civil-rights (17), communication-access (17), decompensation (17). The draft proposes two new federal instruments: **NAPTI** (National Autism Precision Therapeutics Initiative) for translation, and **Autism.gov** (37 mentions) for navigation and "public accountability."

Dropped words (used in 3+ earlier plans, absent from the draft): sharing, enhance, **scientists**, **evidence-based**, repository, mutations, infants, prevented, collaborative, **advocates**, explore, **vaccine**, gender, international, ABA, survey, males.

### 2.7 Similarity, style, and institutional memory

- **Content:** nearest earlier document is SP-2023 at cosine 0.29 — versus 0.41 average similarity between consecutive earlier plans. The draft is further from its own series than any plan was from its predecessor.
- **Style:** nearest by Burrows' Delta is SP-2011 (Delta 1.13); the draft clusters away from every recent plan.
- **Text reuse:** **0.03%** of the draft's 8-word shingles occur anywhere in the earlier record (68 shared shingles total; longest shared passage: 16 words), and **0.02%** come from SP-2023. The 24% average for earlier plans should not be used as the comparator: it is carried by the annual-update era, when each plan was a revision of the last (SP-2010 inherited 52.8%, SP-2013 40.1%, SP-2011 31.1%). The modern plans were already near-clean-sheet documents — SP-2017 inherited 4.4% from the record and 1.1% from its predecessor, SP-2019 19.4% and 5.9%, **SP-2023 14.7% and 0.94%**. Against that recent norm the draft is roughly 50× lower, not the ~800× the historical average implies. The rupture is real but its size has been overstated.
- The draft's own preface states why, in part: *"Our drafting process did not have the benefit of a complete institutional record or formal handoff from the prior planning cycle."* But earlier plans were built largely from the same *published* record and still inherited a quarter of their text — the near-zero reuse indicates a deliberate clean room, not merely a missing archive.

### 2.8 Trends reversed

Of 80 terms with a strong monotonic trend (|ρ| ≥ 0.5) across 2007–2023, the draft **reverses 55 — 69% of its trend set**. **This number needs a control, and the control has to be time-symmetric.** Scoring an earlier document by refitting the trend on the whole corpus and comparing it to the 2018-onwards average is not the rule applied to the draft: a document published before that window sits on the reversing side of any rising trend by construction, which is why that version returns 75–79 of 80 for the 2009–2013 plans and makes the draft look ordinary. Applied properly — each document judged against **its own past**, trends refitted on the documents preceding it and compared with its own trailing eight-year window — the 28 earlier documents reverse **25–76%** of their own trend sets (median **48%**). A reversal share near half is the corpus norm, and the measure still has no magnitude floor: 61 of the 80 terms are reversed by 8 or more documents in the retrospective diagnostic and none by the draft alone. But the draft is at the top of the distribution, not in the middle: **rank 2 of 29 documents, and the highest of any strategic plan** (SP-2013 25%, SP-2019 42%, SP-2023 48%, SP-2017 61%, draft 69%). Full working in `trend_reversal_walkforward.csv` and `reversal_control_by_term.csv`.

**The pairwise check, which needs no baseline at all.** Comparing each plan only with the plan before it, on one fixed vocabulary of 386 terms: the draft changes **62.2%** of it by a factor of two or more — the largest step in the series, against 50.5% for 2019-vs-2017 and 13–35% across the annual-update era — with **45.1% of terms at least halved** and the median term at **0.60×** its SP-2023 rate. That is the strongest form of the rupture finding, because it involves no trend fit, no window and no baseline choice. But on the strict pairwise *reversal* rule (plan-series trend, 2× magnitude floor) the draft reverses 32.4% against **SP-2023's 42.8%**: SP-2023 was an expansionary rewrite whose increases reversed falling trends, while the draft mostly continues declines downward. The draft is the most divergent plan-to-plan step in the record, not the most trend-reversing one — full table in `consecutive_plan_change.csv`.

Applying a magnitude floor — used at least 5 per 10k in the 2018+ documents and at least halved by the draft — leaves **13 substantial reversals**: health (111.1 → 47.0), disparities (19.8 → 1.3), living (14.5 → 5.4), adulthood (11.9 → 2.2), experience (11.5 → 4.0), receive (11.4 → 4.0), physical (10.8 → 2.7), racial (9.5 → 0.7), improving (8.2 → 2.4), meet (7.9 → 2.7), suicide (6.6 → 2.2), sex (5.4 → 1.6), justice (5.1 → 0.0). Report these rather than the fifty-five.

Sex and gender language deserves its own line, because the deletion is entirely the draft's and not a continuation of any trend: *women/woman/girls/female(s)* falls from 152 uses in SP-2023 (15.0 per 10k) to **2 uses** in the draft (0.19 per 10k); *gender\** from 103 uses to **1**; *men/boys/male(s)* from 54 uses to **0**.

### 2.9 Inside the draft

37 outline sections of 300+ words; every section is ~100% novel text (no inherited boilerplate anywhere). Most prescriptive sections (should per 1k): Autonomic Dysfunction (18.5), Epilepsy (17.9), Neurotransmission (17.2), Sleep (16.3), Caregiver Stabilization (16.2). Densest prose (FK grade): Appendix C (30.6), Contributing Factors and Causality (22.2), Mitochondrial/Metabolic (20.6), Implications for Federal Action (20.6), Communication Access for Nonspeaking People (20.3). The section written *about* nonspeaking people requires postgraduate reading level. No easy-read or plain-language companion exists for the draft (the 2023 plan had one).

---

## 3. Interpretation

### 3.1 What the data shows (measured)

Four mutually reinforcing shifts, each independently measurable:

1. **Genre shift: survey → directive.** The epistemic machinery of the old plans (will/may/suggests; hedges; "we") was replaced by obligation grammar ("should" ×10; deontic index 0.66 vs 0.21) and management vocabulary (dashboards, workplans, implementation, accountability). The document directs agencies; it does not deliberate.
2. **Subject shift: services → mechanism.** Services and interventions fall by half; treatments triple; medications rise tenfold; biomarkers, subgroups, and therapeutic domains organise the document. The unit of analysis is no longer a person receiving support but a biological subtype receiving a compound.
3. **Naming shift: person-reference thinning.** The community-approved *label* (identity-first) was adopted — superficially a concession to self-advocates — while individuated person references fell, the spectrum frame was deleted, and autism was recast as a "condition." Simultaneously, a severity frame was built: "profound autism" (0 → ~50), nonspeaking foregrounded, tied to an explicit claim that the severe *share* is rising.
4. **Register shift: the jargon wall.** The least readable document in the record (FRE 0.57; FK 17.9; 49% long words; nominalisation z = +5.5), with no plain-language companion, including in its sections about people who cannot read it.

Plus one meta-finding: **institutional rupture.** 0.03% inherited text against a 24% series norm; 55 of 80 long-run vocabulary trends reversed; the committee's own name nearly absent from its own plan (IACC 4.5 vs 38.0 per 10k). Seventeen years of negotiated, incremental, publicly-built language — discarded in one draft.

### 3.2 The interpretation we favour (inference, argued)

**Jargon here is not a byproduct; it is audience selection.** A document at FK 18 does three things at once:

- **Self-advocates cannot complain precisely.** You cannot object to a sentence you cannot parse. Objection costs are raised above the capacity of the constituencies most affected — including the nonspeaking people the document discusses in its densest prose.
- **Scientists cannot rebut cheaply.** Pseudo-precise framing ("biomarker-defined confirmation," "standardized functional designation," "immune-metabolic subgroups") sounds falsifiable but is asserted, not derived; a rigorous rebuttal must unpack each compound claim. Asymmetry of effort: a page to assert, a monograph to refute.
- **Laypeople hear credentials.** Long words, acronyms, and therapeutic-domain architecture read as expertise to journalists, legislators, and parents. The register manufactures authority while withholding the accessibility that would let readers check it.

This is a documented rhetorical pattern in fields whose claims weaken under scrutiny: insular jargon, scientific phrasing, and complexity signalling substitute for evidence that survives inspection. The draft's register is consistent with that pattern — and inconsistent with the record of a committee that previously published easy-read editions so affected people could answer back.

**The severity pivot deserves scrutiny, not reflexive acceptance.** "Profound autism" enters federal planning here for the first time, ~50 occurrences, explicitly coupled to a rising-share claim. That framing serves whoever benefits from a public image of autism as predominantly severe, mute, and biologically treatable: it re-anchors prevalence debates, justifies pharmacological investment, and marginalises the self-advocate framing (spectrum, neurodiversity — both deleted from the lexicon). Whether this reflects the sincerely held priorities of a new committee or a coordinated agenda cannot be decided by linguistics alone. What linguistics *can* say: the pivot was executed silently, without acknowledging the 17-year vocabulary it replaced, and in prose calibrated to be unanswerable by the people it describes.

**Deletion by silence is the document's most characteristic move.** "Vaccine" — 4.5/10k across the earlier record, mostly in debunking contexts — occurs zero times. "Neurodiversity": zero. Suicide, equity, justice, accessibility, veterans: all rising trends, all reversed. Nothing was argued against; the words simply no longer exist. A reader of the draft alone cannot detect this. Only measurement against the baseline can.

### 3.3 Honest counterweights (what the data cannot prove)

- **Intent is not measurable.** Genre change, jargon, and novelty are also consistent with an inexperienced committee, consultant-drafted text, or haste under statutory deadline. The preface's "no handoff" admission is a genuine mitigating fact for the rupture finding, and the title page marks the document "Working Draft — Subject to Errors and Revision", released for public review and discussion on 31 July 2026. It is a draft for comment being measured against finished publications.
- **Two facts cut directly against the "jargon as audience selection" reading, and should be stated before a critic states them.** First, the draft requires plain-language and accessibility standards for the products it proposes, in ten separate passages — "federal plain-language, accessibility, multilingual, and AAC-accessible standards" for autism.gov, family-facing tools, benefits guidance, and screening materials. A document choosing register to exclude self-advocates would not mandate AAC-accessible design for its own outputs. Second, the draft carries 65 pages of references organised by section with 55+ DOIs, which is not what "asserted, not derived" predicts; testing that charge requires auditing what those citations actually support, which this analysis has not done. The absence of an easy-read companion is also a draft-versus-publication comparison: the 2023 easy-read edition accompanied a final plan.
- **Some shifts are defensible on merits.** Identity-first labelling, attention to nonspeaking people, caregiver stabilisation, and life-course coverage (aging, transportation, housing sections are all present) answer real community demands. The critique is not that these appear; it is what was *removed* to make room, and the register in which it was done.
- **The Autism CARES Act of 2024** broadened the plan's remit toward services and policy, which legitimately pushes toward implementation vocabulary. It does not, however, explain the collapse of "services," "interventions," "community," or "advocates."
- **The z-scores** rest on 8 prior plans; "FRE 0.57" rests on heuristic sentence-splitting. Directions are solid; exact magnitudes are approximate.
- **Raw cross-check rates** in §2.3–2.4 include reference lists; use the notebook's cleaned rates for citation.

### 3.4 Bottom line

Within one drafting cycle, the IACC's strategic plan changed its grammar (describe → command), its subject (persons receiving services → subtypes receiving treatments), its naming (spectrum/disorder → condition/profound), its audience (families and community → agency auditors), and its memory (24% inherited → 0.03%). Any one shift could be innocent. Their convergence in a single document — in the least readable, most jargon-dense text the committee has ever produced — is the signature of a worldview being installed, not a plan being updated.

---

## Appendix: key files

- `iacc_linguistic_comparison.ipynb` — full analysis, methods, and smoke tests
- `outputs/iacc_linguistic/report.md` — auto-generated headline numbers
- `outputs/iacc_linguistic/draft_zscores.csv` — all z-scores
- `outputs/iacc_linguistic/lexicon_per10k.csv` — naming/lexicon rates, all documents
- `outputs/iacc_linguistic/keyness_words_vs_prior_plans.csv`, `keyness_bigrams_vs_prior_plans.csv` — distinctive vocabulary
- `outputs/iacc_linguistic/new_vocabulary.csv`, `dropped_vocabulary.csv` — coined and deleted words
- `outputs/iacc_linguistic/reuse_carryover_by_plan.csv` — text inheritance by plan
- `outputs/iacc_linguistic/term_trends.csv` — rising/falling terms and the draft's verdicts
- `outputs/iacc_linguistic/draft_sections.csv` — per-section style, novelty, nearest earlier document