When AgriVoice cannot match a pesticide question to reviewed guidance, it stops the voice workflow and sends the question to a named extension officer. That pause protects the farmer from a fluent answer that may contain the wrong dosage, re-entry period, or pre-harvest interval.
A cocoa farmer has one thumb on the sprayer cap. The product is beside him. Water is ready. He asks in Asante Twi how much chemical to mix and when workers may enter the field again.
AgriVoice converts the speech into text and searches its approved cocoa guidance. It finds related material, but no reviewed block that safely answers the whole question. The system does not fill the gaps. It stops.
That silence may feel like failure in the moment. It is one of the most important things the product can do.
The day two filters would not fit
In April 1970, the crew of Apollo 13 faced rising carbon dioxide inside the spacecraft after an oxygen tank explosion forced them into the lunar module. The command module carried square lithium hydroxide canisters. The lunar module used round ones. The crew had filters, but the available parts did not fit the system keeping them alive.
NASA engineers in Houston had to solve the mismatch using materials available aboard the spacecraft. Flight controller Ed Smylie and his team developed a makeshift adapter from items including a plastic bag, cardboard, a hose, and duct tape. Astronaut Jack Swigert assembled the device in space using instructions relayed from Mission Control.
NASA documents the episode in its Apollo 13 mission history, including the improvised adapter that allowed the command module canisters to work in the lunar module system. The critical discipline came before the successful fix: engineers worked with known materials and tested instructions. They did not tell the crew to improvise freely around an uncertain life-support problem.
A pesticide question has lower stakes than a damaged spacecraft, but the mechanism is the same. When the available answer does not fit the actual situation, confident improvisation can turn a gap into harm.
A partial match can still be unsafe
AgriVoice is designed around reviewed content blocks. The language model selects those blocks; it does not write agronomy advice from scratch.
That boundary matters most when a question includes chemicals. A reviewed block might identify a cocoa disease or name an appropriate treatment category while still lacking the dosage on the farmer’s product label. It might say nothing about protective equipment, how long people should remain outside the sprayed area, or how long the farmer must wait before harvest.
Returning the partial block as a complete answer would hide those omissions behind a natural-sounding Twi voice.
The safer response is explicit: the system cannot provide verified instructions for this question, so a person must review it. AgriVoice then routes the case to the extension officer named by the pilot partner. The farmer should hear that the question has been escalated, rather than receiving a guess or unexplained silence.
This is also why the current pilot cannot begin merely because the code path works. Three chemical guidance blocks remain held until a Ghanaian agronomist verifies dosage, re-entry, and pre-harvest details. The refusal is the correct behavior while those facts are missing.
The same principle applies when a question only resembles reviewed guidance. A half-match should trigger caution, not completion.
Fluency can conceal the dangerous part
Voice interfaces create a special trust problem. A response delivered smoothly in a familiar language can sound settled even when its underlying evidence is incomplete.
Neuralis has already seen how language systems can change meaning without announcing the change. In one translation test, the Twi word `kokoo`, meaning cocoa, returned as “chicken.” The sentence could remain grammatical while the subject changed completely.
Pesticide advice leaves even less room for that kind of error. A plausible dosage is not a verified dosage. A general safety warning does not supply a missing re-entry interval. A model’s confidence does not turn absent agronomic review into evidence.
The pilot scorecard therefore treats any unreviewed or unsafe pesticide instruction reaching a farmer as a stop or redesign signal. Success means zero unsafe pesticide answers. Questions may be answered from approved content or correctly escalated, and the escalation must have an accountable owner.
For a closer look at why reviewed Twi alone does not settle every spraying question, see what happens when safe spraying details are missing.
Build the stop before the answer
A useful safety workflow needs more than a refusal message. It needs approved content with clear limits, tests containing ambiguous and ASR-corrupted questions, and a named person who receives the cases the system cannot resolve.
It also needs operational evidence. During the planned two-week pilot with 20 to 50 farmers, Neuralis will measure whether questions are answered or correctly escalated, whether the response is understood, and whether the extension officer resolves escalations within the agreed working window. Repeated confident errors end the experiment.
Apollo 13’s improvised adapter worked because Mission Control respected the mismatch, checked the available materials, and passed a bounded procedure to the crew. AgriVoice follows that lesson at the sprayer cap: recognize when the pieces do not fit, keep unsupported instructions out of the farmer’s ear, and put the unresolved question in qualified human hands.
Comments
No comments yet.