A self-check before you write the report. Not one row here says "high accuracy": the rubric measures whether what you got is stated honestly.

The rubric

#	what the report must say	good	weak
1	What you measure — how many images, how many classes, how many in data/ and test/	"400 new images, 100 each from four classes"	"accuracy 94 %" with no image count
2	Which number you publish	the new-image one, with ± and the image count	the one that looks best, or the validation one
3	What you decided on	validation only; it is shown that test was not used for the decision	"I chose the one that gave the best result" (which?)
4	How often test/ was looked at	a concrete number, equal to the number of runs	not mentioned
5	What you tried — every run, including those that did not help	a table of all of them; the run that did not help is mentioned	only the winner shown
6	Where it goes wrong — a confusion matrix and at least three mistakes described	matrix + 3 files with an explanation	only the average accuracy
7	What it matters for — which mistake costs you more (precision or recall, for which class)	one concrete sentence	"both matter"
8	What is unresolved — what the model cannot do, what is unknown	a concrete limit ("no 'other' class", "20 photos per class")	"I will keep improving it"

Five common distortions

None of them is deception; each is an ordinary accident that "improves" the number:

1. Publishing the validation number instead of the new-image one. In lesson 11: 95.6 % against 93.5 %.
2. The best epoch, picked on the same set you publish. On validation it will always look good.
3. test/ used for choosing. Even "I just looked once to be sure" — if you changed anything afterwards, it is no longer a test.
4. The best of many runs, without saying there were many. In lesson 7 six runs with different seeds gave 87.2 % validation for the best, while their mean was 86.2 %.
5. A single number without an error. 160 validation images give ±3.2 points; 400 new ones, ±2.4.

How to read your own results

• If validation and test differ by less than the error — you have established no difference; say: "the difference fits inside the error".
• If augmentation (or anything else) made a difference smaller than the error — do not rank. Choose by validation, but do not claim it is significant.
• If a class has only 20 test/ images, then when its recall is about 90 % the error is ±13 points (the drills of lesson 11).
• If the model does everything badly — that is also a report. Only with a matrix that shows where.

Ask yourself

If somebody who has not seen my data read only my report, could they repeat what I did, and would they know what to believe?