AMD Developer Hackathon ACT III · Intelligent Industry

SAYAGAIN

Nobody on the channel noticed. The radio
log did.

Checks each read-back of a radioed safety order against the plant registry; wrong ones, and any it cannot check, reach the supervisor.

MISMATCH

H-01 · CH-01 · English

tag: XV-203 ≠ XV-208 XV-208 is on line A (registry)

OPERATORH-01-0

Close valve two-oh-three, line Bravo.

FIELD TECHH-01-1

Copy, closing two-oh-eight, line Bravo.

speaker text·scripted
extraction Gemma-4-31B-IT · AMD-hosted API
referee local · 1.07 ms

zsh · build/

jq -r '…' bench/results/text/llm-judge-r1-reaudit.json

llm_judge false-VERIFIED 2 false alarms 1/48

jq -r '…' bench/results/text/ladder-r1-reaudit.json

Gemma-4-31B-IT false-VERIFIED 0 false alarms 2/48

regex false-VERIFIED 0 false alarms 0/48

# 3 more rungs, false-VERIFIED 0 on each

CH-01 · how the channel hears it today

Two lines. Nobody flags them.

OPERATORClose valve two-oh-three, line Bravo.

FIELD TECHCopy, closing two-oh-eight, line Bravo.

no flag · no card · the crew walks to XV-208

In the registry the demo runs on, XV-203 is on line B at the press station; XV-208 is on line A, in clarification.

Three-way communication

Read-back is the check. Nothing checks it.

Wrong read-backs caught, of every 100 · aviation, Helmke 2021

tower63

approach50

filled: caught by the controller · empty: went through

No rate

published for plant radio

The supervisor queue · stored shift replayed

The supervisor gets the one wrong word.

SayAgain supervisor queue: the list of exchanges, and card H-01 with two-oh-eight blocked red under the order two-oh-three, buttons Confirm error and Dismiss

1the diff, checked against the registry
2the read-back word that differs
3a human confirms or dismisses

One model job · everything else is code

The model points. Code decides.

1 · the exchange the model read

order

Close valve two-oh-three, line Bravo.

read-back

Copy, closing two-oh-eight, line Bravo.

H-01 · written words, text·scripted

2 · what Gemma-4-31B-IT returned

{
  "commands": [
    {
      "action_span": "closing",
      "line_span": "line Bravo",
      "state_span": null,
      "tag_span": "two-oh-eight"
    }
  ],
  "kind": "readback",
  "retracted_spans": []
}

AMD Token Factory
developer.amd.com.cn
median latency 4.304 s, first pass
cost USD 4.8937e-06

3 · what code decided

instruction: grounded + covered (1 command(s))
read-back: grounded + covered (1 command(s))
audio: no gate (no token probabilities, no unprompted pass)
instruction: XV-203 line B (registry ok)
compare: tag: XV-203 ≠ XV-208

MISMATCH

tag: XV-203 ≠ XV-208

speech-to-text: not run yet · span extraction: AMD Token Factory · referee: local, plain Python · queue: local

receipt, as stored · request_id c9647a04cfcc4eeaadbfa965a4d74c6d · model_served self-dploy/Gemma-4-31B-IT (sic, as served) · host developer.amd.com.cn · latency_s 35.4972 (a slow call; the first pass’s median was 4.304 s) · cost_usd 4.8937e-06

XV-205 · a valve the plant registry does not have

Judging is not the model’s job.

LLM-as-judge
DeepSeek-V4.1-Flash · no registry

VERIFIED

Order and read-back match; the valve does not exist.

It also passed 2 wrong read-backs.

ABSTAIN

R-01 · CH-01 · English

ABSTAIN · unknown tag cannot clear this exchange: a human must check it

OPERATORR-01-0no recording

XV-205 not in registry

Close valve two-oh-five, line Charlie.

FIELD TECHR-01-1no recording

Copy, closing two-oh-five, line Charlie.

speaker text·scripted
radio none (no audio)
extraction Gemma-4-31B-IT · AMD-hosted API
referee local · 0.768 ms

Model ladder · AMD Token Factory · 89 held-out · text·scripted

2 versus 0 wrong read-backs passed.

The LLM-as-judge baseline passed 2. Each of the five rungs passed 0.

rungfalse-VERIFIEDfalse alarms on 48 correct read-backs · dashed: 5 % budgetrule

regexrung 0 · no model00/48meets

MiniCPM5-2Bsmallest on the ladder048/48all 48 to a human

Gemma-4-31B-ITthinking off · re-audit02/48meets

DeepSeek-V4-Flashthinking off · re-audit04/48over budget

Qwen3.8-27Bthinking on · re-audit04/48over budget

LLM-as-judgebaseline · DeepSeek-V4.1-Flash · no registry21/48passed XV-205

On written words, regex meets the rule too; audio decides.

Who pays

Priced per channel. Bought by HSE.

1,318mills, Indonesia 2022 (Trase)

× 30 %addressable*

× 6channels per mill*

× USD 2,400per channel-year*

≈ USD 5.7 Ma year, mills only (estimate)

* assumption

Wedge

Free 30-day listen-only baseline

Who else listens to plant radio

Loggers transcribe. None we checked compares.

	transcribes radio	compares a read-back	plant tag registry
NICE Inform AI	yes	—	—
Eventide NexLog + Speech Factor AI	yes	—	—
Motorola Assist	yes	—	—
IntelliPERMIT	—	radioed codes, not speech	yes
HAAWAII (air-traffic research)	yes	yes, aviation	no
SayAgain	planned	yes	yes

— not in the public documentation we read, 2026-10-09 · sources: evidence/sources.json

README

What we abstain on.

ABSTAIN

No recorded speech has been measured.

The plant, its 120 tags and every exchange are fictional…

…what a language model adds is unproven until the audio rows exist.

No audio evaluation in this submission

Audio rowthe rule on recorded speech
Failure case−6 dB look-alike tags
Kill testEnglish, Bahasa or auto
Blind auditorby ear vs the queue

The ask

Put it on one mill’s radio.

One mill. Thirty days. Listen only.

Q1 2027paid pilot, one palm-oil mill
Q3 2027DCS check: did it close?
Q4 2027Malaysian palm belt

github.com/edycutjong/sayagain

Releases

Edy Cu · Jakarta