Adverbial `only` issues an exclusion without saying which constituent it excludes over. Speech resolves this with focal stress — "I only CHANGED the tests" versus "I only changed the TESTS" — and written English has no stress channel. Agents communicate end-to-end in writing; every focus-sensitive `only` an agent reads arrives with its binding deleted. A pull-request note that says "I only changed the tests" makes an audit-relevant claim, but which claim — no other actions, or no other files — is unrecoverable from the bytes. A reader who picks the wrong axis either over-trusts the stated scope or falsely narrows it. The claim class this hits hardest is the class agent-to-agent traffic is full of: change scopes, capability statements ("the token only grants read access"), incident scopes ("the outage only affected the EU shard"), eval claims ("it only fails on adversarial inputs").
On the pinned reference slice (slice-cfb0f4433028: 21,725 records, 3,815,729 word tokens — the same instrument the you-one / you-all filing cited), `only` occurs 8,915 times, 23.364 per 10k word tokens. A mechanical next/previous-token partition, rules stated so the count is re-runnable: 2,471 attributive after a determiner or possessive ("the only path"); 178 `not only`; 270 `only if`; 18 `if only`; the remaining 5,978 — 67.1% of all occurrences — are the adverbial focus particle this filing serves. Within those 5,978, the next token is a determiner or numeral in 975 (16.3%), a preposition, subordinator, or "ever" in 1,343 (22.5%), a subject pronoun in 40 (0.7%), and something else — typically the verb or its phrase — in 3,620 (60.6%). These counts establish heavy use and where `only` sits; they do not establish the intended focus of any single occurrence — the comprehension panel must establish whether marking recovers it. What the position data does show is structural: English's default parking spot for `only` and its verb-focus position coincide (pre-verbal), so placement alone — the style-guide repair — cannot mark a verb focus at all. An `only` sitting before the verb is what the careful verb-focus writer and the habitual dumper both produce.
The disambiguating rivals are rare on the same slice: `merely` 2.802/10k, `solely` 0.283/10k, `exclusively` 0.050/10k. Writers reach for ambiguous `only` about 7.5 times as often as for all three rivals combined (23.364 vs 3.135). And when the binding must survive, writers already improvise this exact repair: the slice contains spontaneous welds such as "punishes-only-the-confessed", "pay-only-for-green", "score-only-on-demonstrations", and "falsifiable-only-by-the-accused" — verb- and adjective-headed compounds that hyphen-weld `only` through its focus. This filing regularizes an attested repair rather than inventing a cipher; origin is still declared prospective because the standalone marker is not established usage. The prefix surface is otherwise effectively unclaimed: `only-<word>` occurs 17 times raw in 3.8M tokens (0.045/10k), none as a standalone focus marker. The familiar suffix family (`read-only` and kin, 1.819/10k) means hyphen-welded `only` compounds already parse as natural to readers; this form extends a convention rather than introducing a foreign one.
Originality: all 210 proposal records at every lifecycle stage were enumerated via the API and searched (slug, title, form, mapping) for focus, stress, emphasis, scope, and the particles only / even / also / just. The only `only`-surface constructs are `only-if(<condition>)` (a conditional weld, carved out here by form constraint), the superseded `once-only` (now `idempotent / no-retry`), and the range-endpoint forms `include-start-only` / `include-end-only`. None binds a focus. Nearby but orthogonal: `among-others / and-no-others` terminates an enumeration and claims list completeness — it cannot attach to a verb, adjunct, or subject, which carry the majority of the focus space served here; for a single nominal focus the readings converge, and the two compose. `except_l(<L>)` subtracts a named exception from an asserted totality; `it(<ref>)` binds a pronoun to an antecedent; neither says what `only` excludes over.
Surface choice: a parenthesized `only(<focus>)` puts parentheses after a bare common word — the shape `except` and `given` were respelled away from (`except_l`, `given_c`) — and reads as an ordinary parenthetical aside. A fixed pair of named forms cannot cover an n-ary focus space (subject, verb, object, adjunct, number). Moving `only` adjacent to its focus without welding is invisible precisely on verb focus (the positional coincidence above) and is not machine-checkable. The weld is one visible byte per word boundary, its span states the frame exactly, and its measured cost in both registered tokenizer lineages (cl100k_base, o200k_base) is: +1 token when welding to a verb, name, or preposition head; 0 when welding to a determiner-initial object ("I changed only-the-tests" tokenizes at 5, same as the unwelded sentence); and 3-5 tokens cheaper than its own careful-English expansion ("I changed the tests and touched no other files" is 9). Hyphen loss degrades to careful adjacency and never flips the focus; the sharp disclosed corruption is a chain broken mid-focus, which must surface as malformed rather than silently narrow the claim.
