Producer vs Consumer Tasks

Aarne Ranta's distinction — producer vs consumer tasks — reframes how
we think about AI applications:

Consumer tasks: the system reads or verifies things produced by others.
Examples: parsers, type checkers, theorem provers, search engines, formal
verifiers.

Requirements:
- Must cover ALL of the input language (no "I don't know" for valid inputs).
- High precision: errors are visible to the human.
- Generic: the consumer often doesn't know the application domain.

Producer tasks: the system creates new artifacts from scratch.
Examples: translators, code generators, formalizers, summarizers, LLMs.

Requirements:
- High precision is harder (no ground truth to compare to).
- Can be domain-specific (a medical translator only needs medical precision).
- The producer chooses what to say.

The asymmetry. Producers face a much harder problem than consumers — they
must generate correct output without any input to verify against. Consumers
just have to validate.

The LLM angle. Large language models are producer-heavy: they generate
text. They are NOT reliable consumers — they hallucinate. A common production
recipe:
1. Use an LLM to produce a draft (producer).
2. Use a symbolic tool (type checker, parser, theorem prover) to verify
(consumer).

The hybrid wins: the LLM gives coverage and fluency, the symbolic tool
gives precision and guarantees.

Where symbolic shines in producer-consumer: in CONSUMER roles — type
checking, parsing, formal verification. These are exactly the tasks where
"close enough" is not acceptable.

Tasks divide by whether the same artefact is consumed by humans and machines (consumer) or produced for both (producer).

Consumer tasks — content is mostly for human reading, with machine help as a bonus:
- Machine translation: humans translate by hand for nuance; machines provide quick first drafts.
- Multilingual documentation: human translators handle idioms and culture; machines translate boilerplate.

Producer tasks — content is generated primarily for machines (and humans inspect):
- Machinery control: PLC programs, robotics — correctness is paramount, generated for execution.
- Program verification: proofs checked by theorem provers; humans write specifications and proofs.
- Medical technology: drug-interaction checkers, dose calculators — automated, certified.

In producer tasks, formal correctness is essential (the artifact runs). In consumer tasks, fluency and style matter more than formal correctness.

Citation: This framing follows Aarne Ranta's work on Grammatical Framework and producer/consumer languages (Ranta 2017, Lecture Notes on Grammar Formalism). The producer/consumer distinction is also implicit in the PSSH (Physical Symbol System Hypothesis) view: symbols manipulated by machines only matter when the result is consumed by another machine process.

Symbolic methods are particularly strong when:
1. The domain has clear rules (mathematics, logic, formal languages).
2. Provenance matters — we need to explain why a conclusion follows (proofs, derivations).
3. Correctness is mandatory (medical, legal, financial, safety-critical code).
4. The artefact is reused across multiple consumers (compilers, proof checkers, knowledge bases).

Symbolic methods are weaker when:
1. The domain is fuzzy (vision, speech, taste).
2. Mass data is available but rules are hard to write (translation between natural languages).
3. Errors are tolerable (a 95% correct chatbot is fine; a 95% correct dosage calculator is not).