The Three Chemical Blocks AgriVoice Withheld, and What Unsafe Advice Could Cost

A chemical question should receive reviewed guidance only when the system identifies the request clearly, selects an approved content block, preserves every required safety instruction, and has enough confidence to proceed. If any check fails, AgriVoice must stop the automated Twi recording and send the question to the named extension officer.

In April 1970, the Apollo 13 crew faced rising carbon dioxide levels after an oxygen-tank failure forced them into the lunar module. The available lithium hydroxide canisters did not fit the equipment that now had to keep the crew alive.

At Mission Control in Houston, flight director Gene Kranz and the ground team could not send the astronauts a plausible improvisation and hope for the best. Engineers had to devise a workaround using materials available aboard the spacecraft, test it on the ground, and give the crew a procedure they could follow. NASA’s Apollo 13 history documents both the danger and the improvised carbon dioxide removal system that helped bring James Lovell, Fred Haise, and Jack Swigert home.

A cocoa farmer’s chemical question carries smaller stakes, but the decision has the same shape. A fluent answer is insufficient. The content, conditions, and instructions must fit before anyone acts.

The first check identifies what the farmer actually asked

The clock begins when a farmer sends a Twi voice note through WhatsApp. Speech recognition turns the recording into text, but recognition alone cannot establish meaning with enough certainty for chemical guidance.

Farm speech may include background noise, English product names, local pronunciations, or several questions in one recording. A transcription can look convincing while changing the important noun. Neuralis has already seen machine translation turn “kokoo,” meaning cocoa, into “chicken.” That error is a warning against treating fluent language as verified understanding.

AgriVoice therefore checks whether the transcript points clearly to a reviewed question category. Is the farmer asking how to identify a disease, when to apply a treatment, how much product to mix, when someone may re-enter a sprayed field, or when cocoa may be harvested after spraying?

Ambiguity changes the route. The system should request clarification or escalate instead of choosing the nearest-sounding answer.

Approved content must match the whole question

When the request is clear, the reasoning layer selects reviewed content block IDs. It does not compose new agronomy advice.

That boundary matters most with pesticides. A block that names a chemical but lacks a verified dosage, re-entry interval, or pre-harvest interval cannot become safe through confident delivery. AgriVoice currently withholds three chemical blocks for exactly this reason. A Ghanaian agronomist must resolve those gaps before the blocks can be released.

The selection also has to match every part of the request. If a farmer asks about dosage and when it is safe to return to the field, a block covering only dosage is incomplete. If the transcript contains two possible products, the system cannot silently choose one.

This is where the Apollo 13 analogy earns its place. Mission Control tested a specific procedure against the equipment and materials actually aboard the spacecraft. AgriVoice must test a selected answer against the farmer’s actual question and the approval status of every required instruction.

For a closer look at why incomplete pesticide content stays withheld, read The Three Chemical Guidance Blocks AgriVoice Withheld, and What They Could Prevent.

Safety content cannot disappear in translation or speech

A selected block still has to survive the trip into spoken Asante Twi.

The Twi text must have been reviewed by a farming-aware translator. Product names, quantities, warnings, re-entry guidance, and pre-harvest guidance must retain their meaning. AgriVoice cannot use an unchecked machine translation for advice that a farmer may treat as authoritative.

The system then checks whether the answer contains the mandatory safety content associated with that block. It also needs to avoid adding an unrelated block that introduces unsafe instructions. The final recording must be understandable through the existing voice, including code-switched terms farmers commonly use in English.

These checks explain why speed is measured alongside safety, rather than above it. The pilot’s performance target is a median response below eight seconds for automated answers, excluding human escalation. A fast unsafe answer fails. A slower escalation can be the correct result.

The safe exit is a person, not a guess

If the question remains ambiguous, the selected content is incomplete, a chemical block is still held, or required safety language cannot be preserved, automation ends. The farmer receives a clear indication that the question needs human review, and the case enters the extension officer’s queue.

That path must have an owner. Before the pilot begins, a cocoa-sector partner must name the extension officer who receives escalations. The pilot scorecard then measures whether those cases receive a response within one working day at the median.

The two-week pilot with 20 to 50 farmers will test the full decision under real conditions. Neuralis will measure successful answers and correct escalations, comprehension, repeat use, response time, and whether any unsafe pesticide instruction reaches a farmer. The safety target is zero.

Apollo 13’s ground team did not remove uncertainty by speaking with greater confidence. They moved from a dangerous mismatch to a tested procedure, then communicated it. AgriVoice follows the same discipline: use reviewed guidance when every check passes; otherwise, hand the question to the person accountable for the answer.