Abhishek S.
Shipping in public. Listening in private.

Abhishek

I lead women’s Indo-Western & Premium at Max Fashion. I also wrote the AI that runs the buying floor.

Rare profile. Category operator who ships production code.

Senior Buying Leader · Max Fashion Women’s Indo-Western & Premium · 530+ India stores NIFT ’12 · On the women's wear floor

abhishek@bengaluru ~ %
>role: senior buying lead
>dept: women’s indo-western + premium
>floor: 530+ stores india

Tuyuca Evidentiality

To say that a man played football in Tuyuca, the verb may have to disclose whether I saw him, heard the game, found his footprints, received a report, or merely judged it likely. Janet Barnes documented these five evidential categories in 1984. English lets a claim arrive without provenance; Tuyuca often makes that omission ungrammatical.

Five endings, five routes to a claim

Tuyuca is an Eastern Tukanoan language spoken in the Vaupés region of Colombia and Brazil. In many independent clauses about past or present events, the verb carries an evidential suffix. That suffix can also encode tense and agreement, so the source label is part of the sentence’s machinery rather than a detachable phrase such as “apparently.”

Barnes used variants of one football claim to expose the system:

Form Category Basis of the claim
díiga apé-wi Visual I saw him play
díiga apé-ti Nonvisual I heard the game
díiga apé-yi Apparent I saw physical traces
díiga apé-yigi Secondhand Someone told me
díiga apé-hĩyi Assumed Prior knowledge supports it

The split between nonvisual and secondhand is a revealing cut. Hearing players on a field is direct sensory access; hearing a person recount the match is transmission through another mind.

What the suffix does not prove

An evidential is not a truth certificate. A speaker can misuse a visual form just as an English speaker can falsely say “I saw it.” Grammar demands a category of access, not honesty.

The boundary between information source and confidence is also contested. A secondhand report can be trusted completely, while a visual impression can be doubtful. Alexandra Aikhenvald’s 2004 account treats evidentiality and epistemic modality as distinct categories, but actual usage lets source, certainty, person, and context bleed into one another. Even -wi is not confined to detached visual observation: Barnes recorded it with first-person actions that a speaker performs rather than watches.

That caution matters beside concept wug test. Children can learn a compulsory distinction and apply it productively without that distinction fixing how they reason about every case.

Provenance before probability

Tuyuca grammar labels the path into a claim. concept bayes theorem asks how strongly that path should alter belief:

P(H | E) ∝ P(E | H) × P(H)

The evidential identifies the kind of E; Bayesian inference weighs it against H. concept information theory measures how much uncertainty a signal removes, while evidentiality preserves a clue about where the signal came from.

The same missing field haunts concept hallucination. A machine-generated sentence may come from a retrieved document, a remembered pattern, or an inference assembled during generation. Tuyuca treats the difference as grammar. Many software systems treat it as optional metadata.

An open question

If every generated sentence had to be marked observed, retrieved, reported, inferred, or assumed, would unsupported claims become easier to catch, or would machines merely learn five new ways to sound certain?

Key Sources

Further Reading

See Also

Abhishek's take

The feature I want to borrow from Tuyuca is compulsory provenance, not compulsory certainty. “Inferred from traces” tells me more than a confidence score stripped of origin.

Tags: #evidentiality #tuyuca #amazonian-languages #grammar #epistemology #data-provenance