Try it / tally sheet

211294C0100

INE 2024 tally sheet, type 2 phone photo Puebla certified by comparison sample: paper rows do not sum

Measured on AMD: Qwen3.8-27B, 779.8 s, $0.000423stored read

Option 03: row 8: the printed label read is not the one expected here: this field goes to a person, with its photo

Gemma's read breaks one of the paper's own checks, so the gate does not accept it.

Straightened by its corner marks; the right-hand column (officials' names and signatures) is blacked out before the page is stored or sent to the model.

Watch it run

Read this sheet again, live

Gemma-4-31B-IT reads the masked page on AMD; the gate checks each field as the answer arrives.

About a minute. One live read runs at a time for the whole site; yours waits in a short queue and shows its place.

How it's checked

Gate: Routed: TOTAL digits 262, words 162, rows sum 262; 5 row labels not where expected

The paper's own checks

✓ rows sum to TOTAL 262 = 262

✓ votaron + representantes = section 5 262 = 262

✓ section 5 = section 7 262 = 262

✓ section 7 = TOTAL 262 = 262

✕ words = digits on every field 22 of 23

Paper inconsistency

TOTAL: on the paper the words say 162 and the digits 262. The official capture applied its words-prevail rule and keyed 162; the rows and section 7, and the certified count (262), support 262. Not a capture error: a person should settle it, which is why the gate routed this form.

Receipts

Every disputed field of the sample audit of 300 sheets gets a Merkle receipt: its record, the SHA-256 of its crop, and a proof path to one published root. This sheet is not in that sample; see a receipt verified on sheet 122655C0100.

Separately: the official capture was wrong here

The official typed record (PREP), not Gemma's read

Ballots taken from the boxes: the official capture keyed 162; the paper says 262.

This compares the official typed record with the paper, not Gemma's read.

Full trace: two real sheets, request to receipt · Source code

All fields (23), and other runs of this sheet

Gemma · measured run Qwen3.8 · re-read after masking Qwen3.8 · second reader

Other runs (4)

Gemma · re-read after masking (attempt 1)Gemma · re-read after masking (attempt 2)Gemma · re-read after masking Gemma · earlier prompt

Qwen3.8-27B, re-read of masked forms with the frozen prompt. sourceresults/verify_spike_qwen.jsonl Read in 779.8 s, $0.000423.

Field	Digits	Words	Official capture	Certified
Unused ballots section 2	170	170	170		✓
Voters on the list section 3	255	255	255		✓
Representatives who voted section 4	7	7	7		✓
Voters + representatives section 5	262	262			✓
Ballots taken from the boxes section 7	262	262	162^†		✓
Option 01	61	61	61	61	✓
Option 02	9	9	9	9	✓
Option 03 row 8: the printed label read is not the one expected here	1	1	1	1	!
Option 04	13	13	13	13	✓
Option 05	6	6	6	6	✓
Option 06 row 11: the printed label read is not the one expected here	29	29	29	29	!
Option 07	113	113	113	113	✓
Option 08 row 13: the printed label read is not the one expected here	4	4	4	4	!
Option 09	2	2	2	2	✓
Option 10 row 15: the printed label read is not the one expected here	0	0	0	0	!
Option 11 row 16: the printed label read is not the one expected here	0	0	0	0	!
Option 12	12	12	12	12	✓
Option 13	0	0	0	0	✓
Option 14	1	1	1	1	✓
Option 15	0	0	0	0	✓
Unregistered candidates	1	1	1	1	✓
Null votes	10	10	10	10	✓
TOTAL section 6	262	162	162^†	262	!

Options 01–15 are the form's vote rows in printed order, shown without party names. Digits and words are the model's two independent readings. Official capture: the double-keyed official capture of this image (PREP). Certified: the certified district count (Cómputos), which has no section 2-5 or 7 values. Amber: the capture differs from the model. † eye-checked: the capture mis-keyed the field. Against the official capture this read differs on 2 of 22 fields; after eye-checking, 0.

Now on 300 sheets: the sample audit