An unmatched reviewed block should trigger a visible escalation for the partner institution, not a generated pesticide answer. For AgriVoice, uncertainty is a safety signal that needs an accountable person and a clear response path.
A warning light during Apollo 13
In 1970, Apollo 13 lost much of its normal capability after an explosion in its service module. Jim Lovell, Fred Haise and Jack Swigert still had a problem inside the spacecraft: carbon dioxide was rising, and the lunar module had a limited supply of lithium hydroxide canisters.
The available square canisters did not fit the lunar module’s round openings. NASA engineers on the ground had to work from materials already aboard the spacecraft and send the crew a procedure they could follow. The solution became a famous example of controlled improvisation under pressure, documented in NASA’s Apollo 13 history and in Lost Moon by Jim Lovell and Jeffrey Kluger.
The important part is the sequence. A dangerous condition became visible. People with responsibility understood what it meant. They worked from known materials and constraints. Then they gave a bounded response.
A pesticide question that does not match a reviewed AgriVoice block deserves the same discipline. The system has detected a limit in what it can safely say. That limit should travel to the cocoa-sector partner as a clear red flag.
A Twi question with no safe answer
Imagine a farmer sending a voice note in Asante Twi through WhatsApp. The farmer describes a cocoa problem, names a chemical product, and asks how much to apply before returning to the farm.
AgriVoice first has to understand the speech well enough to identify the question. It then tries to select from reviewed advice blocks. Those blocks are the product’s safety boundary: the language model selects an approved block ID, rather than writing agronomy advice itself.
If the question fits a reviewed block, the farmer can receive spoken Twi advice that has been translated and reviewed for farming use. If it does not fit, the system should say so plainly and route the question to the named extension officer.
That outcome may feel less satisfying than an immediate answer. It is still useful. The partner institution learns what farmers are asking that the approved content does not cover. The farmer avoids receiving a confident instruction with an unverified dosage, re-entry interval, or pre-harvest interval.
AgriVoice currently withholds three chemical blocks because those details have not been verified by a Ghanaian agronomist. That is a deliberate safety decision. A system that fills those gaps with fluent language creates a risk that sounds helpful right up until it causes harm.
Make the escalation visible to the institution
A safe escalation cannot disappear into an inbox with no owner. Before the farmer pilot begins, the cocoa-sector partner must name the extension officer who receives these questions and commits to handling them.
The routing record should preserve the practical facts the institution needs: the reviewed blocks considered, the reason no safe match was found, whether the question concerns a pesticide, and whether speech recognition or code-switching may have made the request unclear. The officer can then respond, request clarification, or identify a missing content area for expert review.
This creates a useful distinction between two kinds of failure. A farmer may ask something outside the content pack, which calls for human follow-up. Or the system may struggle to understand the voice message, which calls for better field speech evaluation. Treating both cases as a reason for the model to guess hides the information the pilot is supposed to produce.
The two-week pilot has a clear safety measure: zero unsafe pesticide answers. It also measures successful answers as questions answered or correctly escalated. A correctly escalated question counts because the product has preserved the farmer’s path to help while keeping unreviewed advice out of the conversation.
The need for a named owner is explored further in The Named Extension Officer AgriVoice Needs, and What an Unanswered Question Could Cost.
Turn red flags into reviewed coverage
Each unmatched question should be reviewed after the fact. Some may reveal a missing block that an agronomist can safely approve. Others may show that farmers use a different phrase for a disease, product, or farm practice than the content team expected. Some may expose an ASR error that changes the meaning of a Twi message.
That review loop matters because machine translation has already produced a serious domain error in this work: `kokoo`, meaning cocoa, came back as “chicken.” A fluent sentence can still carry the wrong noun. The case where cocoa became chicken is a reminder to test the whole route, not only the final voice.
Apollo 13’s carbon dioxide problem did not become safer because someone ignored the warning light. It became manageable because the warning reached people who could act within known limits. AgriVoice should give partner institutions the same kind of signal: here is the farmer’s question, here is why approved advice did not cover it, and here is the person who must decide what happens next.
Comments
No comments yet.