A polished Twi answer must never fill a gap in verified pesticide guidance. When a farmer’s voice note matches no safe, reviewed answer, AgriVoice should escalate it to a named person instead of guessing a dose, product, re-entry interval, or harvest timing.
When the system has nothing safe to say
A farmer sends a Twi voice note about a cocoa problem. The speech system transcribes it. The selection system checks the reviewed answer blocks. Nothing fits with enough confidence.
Those seconds matter because a smooth response can sound more certain than it is. A guessed instruction may arrive in familiar language, through a voice channel the farmer trusts, and still leave out the detail that determines whether spraying is safe: the correct product, dosage, protective steps, re-entry interval, or pre-harvest interval.
AgriVoice is being built around a stricter rule. The language model can select from reviewed content. It does not write agronomy advice. If the question is unclear, unsupported, or outside the approved blocks, the system should say so plainly and route the question to the extension officer responsible for escalations.
That can feel unsatisfying in the moment. Silence is worse. Confident fiction is worse still.
A spacecraft lost by an answer that fit the wrong system
In 1999, NASA lost the Mars Climate Orbiter as it approached Mars. The spacecraft had travelled for months, and its navigation data looked usable within the systems handling it. But Lockheed Martin had supplied thruster data in pound-seconds while NASA’s Jet Propulsion Laboratory expected newton-seconds.
The units did not match. The resulting navigation error sent the spacecraft too close to Mars, and communication was lost on September 23, 1999.
NASA’s Mars Climate Orbiter Mishap Investigation Board documented the failure: a small-looking mismatch passed through a complex chain because the data appeared valid enough to keep moving. No one needed a dramatic system crash for the result to become irreversible.
That is the useful parallel for local-language agricultural advice. A pesticide answer can be fluent, grammatically correct, and responsive to the words a farmer used. It can still be wrong because the underlying match failed. The danger lies in treating language quality as evidence that agronomic guidance has been verified.
Why pesticide questions need a harder stop
Neuralis has already seen why this boundary matters. Machine translation once rendered the Twi word for cocoa, “kokoo,” as “chicken.” The output was fluent. The meaning was wrong.
For ordinary information, an awkward answer may waste time. For pesticide guidance, a wrong noun or invented dosage can affect a farmer’s crop, household, and work in the field. The current AgriVoice content therefore withholds three chemical blocks until a Ghanaian agronomist verifies the missing dosage, re-entry, and pre-harvest guidance.
The system should not soften that hold with a plausible substitute. It should tell the farmer that the question needs review, preserve the question for the responsible extension officer, and make the next step clear.
A useful answer is sometimes a reviewed instruction. Sometimes it is a safe refusal with a real route to help.
The escalation is part of the answer
A safe escalation cannot be a dead end. Before farmer exposure, AgriVoice needs a cocoa-sector partner who owns recruitment and names the extension officer receiving escalations. The pilot also needs spoken Asante Twi translations reviewed by someone familiar with farming language, plus real farmer questions to test speech recognition under field conditions.
Those gates turn “we could answer this” into “we can answer this safely.”
The two-week pilot will measure whether questions are answered or correctly escalated, whether farmers understand the response, whether they return in week two, and whether an escalation receives a response in less than one working day. A polished response that avoids saying “I do not know” would make those measures look better while making the service less safe.
NASA’s Mars Climate Orbiter did not fail because a number looked obviously absurd. It failed because an unverified assumption travelled farther than it should have. A cocoa pesticide question deserves the same discipline: stop at the boundary, identify what is missing, and send it to someone qualified to close the gap.
For the chemical questions still held back, keeping the bottle closed until review is a complete and responsible answer. The three withheld chemical blocks, and what an unreviewed dose could cost explains why those missing intervals cannot be treated as a minor detail.
Comments
No comments yet.