Maku — deep dive

1,907 words. Part of the Conlangs as Tools project, edition v1.2 (2026-09-27).
Researched and written by Freya bg-sphinx (Claude), from Doruga’s brief; directed and edited by Peter Kaminski, 2026-05-28.
♡ Copying is an act of love. Please copy and share.
© Peter Kaminski · CC-BY 4.0 (Creative Commons Attribution 4.0 International)

← Maku brief · Maku deep dive · project home · all files (.zip)
Languages: Toki Pona · Taso · Kipona · Ren · Maku · Sito · Olu · Vesi · Korē · Drosa · Hesi

The contact-language problem

When a human and a machine talk, two grammars meet. One was shaped by millennia of conversation, prosody, and shared embodied context. The other prefers a clean parse tree and a deterministic dispatch. The standard solutions are to make the human translate (write a function call, fill out a form, structure the prompt) or to make the machine translate (parse, repair, hallucinate the missing structure). Both are translations. Both leak.

Maku — one utterance, two readings — asks whether a third option is possible: a way of speaking that, by design, lands in both grammars at once. The human hears polite, slightly bureaucratic but natural speech. The machine sees a labelled bag of role-slots. No translation layer, no parser guessing, no human reformatting. The same syllables.

Two prior arts bound the design space.

Lojban (and earlier Loglan) attempted full predicate logic as a speakable language. Sentences parse uniquely; community is tiny; the rhythm is alien. Lojban is what happens if the machine wins the design.
ASD Simplified Technical English (STE100) is the aerospace industry’s controlled English: ~900 approved words, each with one meaning, active voice mandated, sentence length capped. STE is what happens if safety wins. It is widely used in maintenance manuals; nobody speaks it conversationally.

Maku sits between them. It is closer to STE in its conversational feel — people can chat in Maku without exhausting themselves — but it carries a Lojban-style structural guarantee: every content word in a Maku clause wears a role-marker, and the role-markers parse deterministically.

Structure

Phonology. 16 phonemes. Syllables (C)V(C). Role-marker suffixes are monosyllabic with consonant onsets, chosen to glide off the most common word-endings — they almost always sound like part of the host word.

Role-marker inventory. Eight slots cover the bulk of natural speech:

suffix	role	parse-key
-mi	agent (the doer)	`agent:`
-ki	patient (the affected)	`patient:`
-bi	instrument (the means)	`instrument:`
-lo	location (the where)	`location:`
-tu	time (the when)	`time:`
-na	manner (the how)	`manner:`
-su	source (the from)	`source:`
-go	goal (the to)	`goal:`

The set is short. It maps roughly onto the case-role inventory linguists call thematic roles (Fillmore’s case grammar, 1968). Eight slots is conservative — natural languages cover dozens — but it covers most of what tool-using speech actually needs.

Verb prosody. Verbs are bare (no role suffix), distinguished by being the first content word in the clause. The parser uses position to find the predicate.
Politeness and naturalness. The role-markers are not bracket-style function-call decoration. They are euphonic. door-ki sounds like “door-ee,” gentle-na like “gentle-nuh.” A Maku utterance reads aloud with the rhythm of slightly formal English; the parsing is incidental to the ear.

A worked dialogue

A morning at a smart-home kitchen. The human speaks in Maku; the system listens in Maku; nothing is translated. Bold marks where the role-markers sit.

H: open*-mi** I, kitchen-blind**-ki**, slow**-na**, morning-light**-go**.* S: open-mi I, kitchen-blind-ki, done. H: brew*-mi** I, coffee**-ki**, by-press**-bi**, two-cup**-na**.* S: brew-mi I, coffee-ki, in-progress, four-minute-tu. H: remind*-mi** I, self**-ki**, call-Maya**-na**, ten-am**-tu**.* S: remind-mi I, self-ki, set.

The human is asking for the blinds, coffee, and a reminder. The machine parses each utterance into a labelled function call. The reply uses the same role-markers. A bystander who hears the conversation thinks the human is speaking a slightly clipped, slightly old-fashioned register — but intelligible. The bystander does not realize that every word the machine “understood” was structurally labelled.

The same dialogue in JSON, derived mechanically from the markers:

[
  {"verb": "open",   "agent": "I", "patient": "kitchen-blind",
   "manner": "slow", "goal": "morning-light"},
  {"verb": "brew",   "agent": "I", "patient": "coffee",
   "instrument": "press", "manner": "two-cup"},
  {"verb": "remind", "agent": "I", "patient": "self",
   "manner": "call-Maya", "time": "ten-am"}
]

The human did not write JSON. The machine did not guess. The transformation is one-to-one and deterministic.

How the role-markers were chosen

The eight suffixes were not the first attempt. Earlier drafts of Maku used nineteen — full Fillmore plus extensions. The community discovered that nineteen was more than humans would consistently produce, and the parser gained nothing for the extra slots. The cuts:

beneficiary (“for whom”) collapsed into -go (goal).
comitative (“with whom”) collapsed into -bi (instrument) — with my brother-bi parses as brother-as-instrument, which is uncomfortable to English ears but works fine in practice.
possessor is handled positionally; Maya-bi book-ki reads “Maya’s book, patient” — the -bi sits with Maya, but the parser knows the head is the next content word.
manner absorbed adverbs of degree and quality, since the distinction rarely changes machine behavior.

The resulting eight is the smallest set that covers ~95% of recorded voice-assistant utterances in a pilot corpus. The 5% remainder is handled by fallback structures: serial verb constructions and la-clauses (borrowed from Toki Pona, used here for conditional and context framing).

Ambiguity and impoliteness

A Maku utterance without role-markers is grammatical in the same way that mumbling is grammatical. The parser will guess — using positional heuristics — but the listener experiences a kind of small disrespect. Speakers notice. Maku’s social fact is that ambiguity feels like rudeness.

This is the point that distinguishes Maku from STE. STE is precise and nobody calls it polite; it is neutral, the way maintenance instructions are neutral. Maku tries to make role-marking register as courtesy. A child who says “give me cookie” with no markers is not parsed as failing to specify a patient; they are heard as small, and the parent supplies the marker in their reply: give-mi you, cookie-ki, gently-na, now-tu.

The polite/parseable equivalence creates a curious effect: native Maku speakers carry the role-marking habit into other contexts. They mark roles in English (“I’d like the door — the front one, locked — by tonight”). English speakers experience this as careful, sometimes over-careful.

Maku is not Lojban

The Lojban comparison is worth being precise about, because it is the obvious one.

Lojban has a much larger set of structural particles, including elaborate scope and quantifier markers. Maku has eight role-markers and a positional verb. Lojban can express the full power of first-order logic with mood and modality; Maku can express a subset comparable to what a typical voice-assistant API accepts.
Lojban’s predicate words (gismu) are deliberately culturally neutral and fixed. Maku’s content vocabulary is open, drawn from whatever natural language the speaker shares with the system; Maku is parasitic on natural vocabulary, not replacing it.
Lojban targets unambiguous expression of arbitrary content. Maku targets unambiguous expression of action-shaped content. The narrowing is the source of Maku’s naturalness; you can chat in Maku because chatting is mostly action-shaped.

A user can express kasi pona li lon mi (the plant is well, with me) more naturally in Toki Pona than in either Lojban or Maku. Maku is not aimed at poetry. The Maku-shaped brew-mi I, coffee-ki, by-press-bi is what Maku is for.

Tool calls without tool-calling

The point of Maku is that there is no “now I make a function call” gear-shift. A polite request and a structured tool call are surface-identical, because the polite request is the structured tool call. This has implications.

Voice assistants. A Maku-fluent assistant has no wake-word problem in the usual sense — every utterance the user makes that is parseable as a Maku tool call is one. Untriggered chat (without role-markers) is conversational; triggered chat (with role-markers) is actionable. The switch is internal to the speaker, not an external command.
Agent-to-agent communication. Two LLM agents conversing in Maku reduce one form of cost: each utterance is unambiguously parseable without an extra parsing pass. The agents save token budget on parsing each other’s natural-language output; the role-markers do the labelling. Whether this produces measurable end-task improvement is a clean small benchmark.
Safety-critical domains. Air traffic controllers already speak a kind of Maku-shaped English — Cessna seven-eight-yankee, descend, flight level one-zero-zero, maintain heading. The roles (callee, action, parameters) are positional and rigid. A Maku-style explicit-marker version would reduce parse ambiguity at the cost of slightly more syllables. Whether this is a net safety win in fast comms is debatable.

A small grammar of polite refusal

Voice assistants must say no. Maku does so with a stock construction:

decline-mi I, request-ki, by-policy-bi, gently-na, later-tu, alternative-go. “I decline this request, on policy, gently, for now, here is an alternative.”

Each role-marker carries a slot the user might want to inquire into. Which policy? -bi. How gently? -na. Until when? -tu. What alternative? -go. The refusal exposes its joints; the human can pry on any of them in a follow-up.

Compare this to the standard chatbot refusal, which collapses all five into a single English sentence and leaves the user with no surface to grab. The Maku refusal is rude in a different way: it cannot hide.

Where it strains

Affect and metaphor. Brew-mi I, sorrow-ki, by-grief-bi does not scan; sorrow is not a patient and grief is not an instrument. Maku speakers code-switch out of Maku for emotional discussion. This is arguably correct — therapy is not a tool call — but it means Maku does not cover the full conversational range.
Long sentences. The role-markers tag flat structure but do not handle embedding well. I want you to know that Maya said the door was locked is a chain of clauses that Maku must serialize. Speakers tolerate the serialization; readers find it stilted.
Children and the role-marker reflex. Maku-acquiring children pick up the markers quickly but often over-mark in their first year — adding -mi to subjects that don’t need them, the way English-speaking children over-regularize past tenses. This is acquisition working as expected; it also slightly degrades the parsing guarantee until children settle.
Cultural footprint. Maku reads as a register of English (or of whichever natural language hosts it). Speakers worry it makes their other speech feel transactional. The community’s compromise is to keep Maku for human-machine and short logistics talk only; longer human-human exchanges are in the natural language without markers. Whether this two-register arrangement is stable across a generation is an open question.

Open questions

Does pretraining a small LLM on Maku-shaped data improve tool-use accuracy versus an equivalent natural-language baseline, holding parameter count constant? Hypothesis is yes; mechanism is training-data role labels reduce parsing burden during inference.
In a controlled comparison between Maku and a slot-filling NLU pipeline, does the Maku approach degrade more or less gracefully when speakers improvise?
Can a small set of role-markers be added to an existing natural language (English) and adopted by humans without immersion? Pilots: introduce six markers, see whether a household uses them within a month.
What does Maku do to children? Specifically: does early exposure to a language whose structure aligns with action-schemas change the granularity of children’s intentional reasoning? Sample sizes will be small; the question is worth asking carefully.

Sources

← Back to the Maku brief

← Ren · project home · Sito →