AgriVoice is forbidden to finish any answer when a farmer’s question only partly matches its reviewed guidance. It must stop at the last verified detail and send the missing part to a named extension officer, because a plausible completion could become unsafe agronomy advice.
In April 1970, the Apollo 13 crew faced rising carbon dioxide while sheltering in the lunar module. The command module carried square lithium hydroxide canisters, but the lunar module used round openings. The parts did not fit, and the crew’s survival was still uncertain.
NASA engineer Ed Smylie and a team in Houston worked with materials already available aboard the spacecraft. They developed an adapter, tested the procedure on the ground, and then relayed the instructions to the astronauts. NASA’s historical account of Apollo 13 documents the improvised carbon dioxide removal system and the mission’s safe return.
The important detail is the test. Mission Control did not fill a gap in the instructions with something that merely sounded workable. Engineers established what the crew had, assembled a specific procedure, and checked it before asking three astronauts to depend on it.
AgriVoice follows the same safety principle on a smaller but still consequential scale: a missing detail must remain missing until someone qualified verifies it.
A question that matches only halfway
Consider a composite example based on the kinds of questions the AgriVoice pilot is being built to handle.
A cocoa farmer records a question in Asante Twi. He identifies a problem on his farm, names a chemical product, and asks how to use it. Part of the question matches a reviewed content block. The library may contain approved guidance about recognising the problem, using protective clothing, keeping other people away during spraying, or following the product label.
Then comes the unfinished part: how much chemical should he add, when may workers safely return, or how long should pass before harvest?
Those details cannot be inferred from the surrounding text. Product concentration can matter. The crop situation can matter. The approved label can matter. A sentence that begins with verified safety guidance can still become dangerous in its final clause.
AgriVoice therefore selects only reviewed content blocks. The language model chooses among those blocks; it does not write agronomy advice. Three chemical blocks remain deliberately held because dosage, re-entry intervals, and pre-harvest intervals have not yet been cleared by a Ghanaian agronomist.
If the farmer’s question reaches beyond the reviewed material, the response stops and the case moves to a person. That refusal is the system working as designed.
Fluency can hide the missing evidence
A generated answer can sound complete even when its evidence is incomplete. That is especially risky in voice.
The farmer does not see uncertainty markers buried in a paragraph. He hears a confident sentence in his language. Natural pronunciation and a familiar conversational rhythm can make an unsupported detail feel as reliable as the reviewed guidance before it.
Neuralis has already seen how language systems can alter important domain terms. In one translation test, the Twi word for cocoa, “kokoo,” returned as “chicken.” The output was fluent. It was also wrong.
That experience shaped AgriVoice’s boundary. Reviewed English content must receive spoken Asante Twi translation from a farming-aware translator. Chemical guidance stays unavailable until an agronomist verifies the missing safety details. Questions outside the approved library escalate instead of receiving freshly composed instructions.
This is also why a refusal should explain what happens next. “I cannot verify that dosage” gives the farmer a boundary. “I am sending this question to the extension officer” gives the boundary a useful destination. The pilot requires a named officer who owns those escalations and responds within an agreed operating process.
For a fuller view of that handoff, see why human escalation makes AgriVoice safer. The same risk appears in the half-filled spray tank AgriVoice must refuse to complete.
Build the stopping point before the answer
Safety reviews often focus on what a system should say. AgriVoice also defines where it must stop.
Before farmers enter the two-week pilot, the approved blocks need spoken Asante Twi reviewed by a farming-aware translator. A Ghanaian agronomist must either clear or continue withholding the chemical blocks, including dosage, re-entry, and pre-harvest guidance. Staff must rehearse the full path and confirm that no unsafe answer reaches a farmer.
The pilot scorecard makes the standard measurable. One unsafe pesticide answer is a stop or redesign signal. Successful handling includes correct escalation, so the system receives credit for recognising its limit instead of guessing past it.
Apollo 13’s improvised adapter worked because the missing connection became an engineering problem to test on Earth before instructions reached the spacecraft. AgriVoice treats an incomplete match the same way. The unresolved detail goes to a qualified reviewer, and the sentence remains unfinished until evidence supplies the ending.
Comments
No comments yet.