In the previous article, I documented the complete ChatGPT retranslation process for the 103 official course pages of A Tour of Go in Simplified Chinese.
The entire process ultimately resulted in:
8 initial full-translation batches
+
3 revision batches
=
11 ChatGPT batchesBatches 001–008 provided the initial complete coverage of the 103 pages, while Batches 009–011 continued to address terminology and language quality issues discovered later.
The previous article is recorded here:
https://www.shuijingwanwq.com/2026/08/19/26947/
At this point, a very natural thought arises:
Since all 103 pages have now been translated and passed the existing validator, can we directly overwrite the official zh-CN?
I ultimately decided not to do this.
Because this round further confirmed one thing for me:
Correct structure and sufficient language quality are two different issues.
The automatic validator can answer the following questions well:
Is present.Section intact?
Are link targets correct?
Are code and directives correct?
Are protection tokens complete?
Does the restored structure match the source?But it cannot fully answer:
Are there subtle deviations in technical meaning?
Is the Chinese natural enough?
Is the same terminology consistent across different pages?
Is there obvious translationese?
Is it fluent enough to serve as an official Go tutorial for beginners?Therefore, after completing the ChatGPT retranslation of the 103 pages, I did not immediately move to production release.
Instead, I added an independent step in between:
A complete semantic quality review.
1. Passing automatic validation does not mean semantic review is over
The .article for A Tour of Go is not standard Markdown.
A course page might simultaneously contain:
present.Section
links
inline code
preformatted code
.play
.image
emphasis structures
other present directivesIf the translation model incorrectly alters any of these, the final page might fail to parse or render correctly, even if the generated Chinese looks perfectly normal.
Therefore, the automatic validator has always been a prerequisite for an official candidate.
But after completing the ChatGPT retranslation of these 103 pages, I became increasingly clear about one thing:
The validator is a necessary condition, but not proof of language quality.
For example, consider these two types of translations:
Structure is completely correct
Technical meaning is mostly correct
Chinese is understandableAnd:
Structure is completely correct
Technical meaning is accurate
Terminology is consistent
Chinese is natural
Conforms to official tutorial styleBoth might pass the exact same structural validator.
However, if the goal is to officially replace the already live zh-CN, we should clearly strive for the latter state.
Therefore, this time I did not equate:
all validators passeddirectly with:
ready for official release2. Re-grading all 103 pages with a complete semantic quality review
To avoid checking only a few representative pages, I ultimately expanded the review scope to cover all:
103 / 103official course pages.
This round used an A/B/C/D four-tier system to record the current semantic and language quality of each page.
These A/B/C/D grades are not results from the present parser, structural validator, or automated tests; rather, they are an internal quality grading system dedicated to this complete semantic review.
The four tiers roughly indicate:
A
Ready for direct release.
Technical meaning is accurate and faithful to the source, Chinese is natural,
and terminology and pedagogical expression meet official release standards.
B
Already meets release quality; no modifications required.
Only minor, optional opportunities for expression optimization exist.
Keeping the current translation will not block the official release.
C
Revision recommended before release.
There are obvious translationese, terminology issues,
unclear expressions, or other problems where fixing them yields clear quality improvements.
D
Must be fixed.
Contains mistranslations, omissions, technical errors,
or obvious ambiguities that might affect the reader's correct understanding of the source text.From a release perspective, this can be further understood as:
A → Directly meets official release quality
B → Meets release quality; modifications are optional optimizations
C → Revision recommended before release
D → Must block release and be fixedThere is also a very important review principle here:
“It could be phrased differently” is not, by itself, a reason for downgrading.
Natural language can almost always be continually rewritten.
If we constantly downgrade pages to B or C just because the model can suggest an alternative expression, the review will never truly converge.
Therefore, what needs to be judged is:
Whether the current translation has clearly identifiable issues, and whether making changes will genuinely result in clear quality improvements.
The final review results were:

Final statistics:
Formal course pages : 103
A : 103
B : 0
C : 0
D : 0
Missing pages : 0
Duplicate pages : 0
Extra pages : 0In other words, when freezing this round of semantic review:
All 103 pages entered tier A.
However, there is one part of the process worth explaining specifically.
3. B pages could have been left unchanged, but I still moved 2 pages to A
During the semantic review, the results were not immediately:
A=103
B=0
C=0
D=0At that time, there were:
2 pagesrated as B.
According to the grading criteria above, B itself already meets official release quality, so theoretically it could have been left completely unmodified.
This means that even if we had stopped at:
A=101
B=2
C=0
D=0these two pages would not have blocked the official release.
But in the end, I still decided to adjust these two pages.
The reason was not that B was “unqualified.”
On the contrary, it was because:
Only 2 pages remained, and the cost of modification was already very low.
As the entire project neared completion, what needed to be handled was no longer dozens of pages, but just two very clearly scoped minor revisions.
At this point, the marginal cost of continuing to revise was low, but it allowed the final state to converge from:
101 A's
+
2 B'sto:
103 A'sSo, I ultimately completed the revision of these two pages.
This decision also reflects my purpose in using the A/B/C/D grading this time:
The grades are primarily used to determine whether a release should be blocked, rather than forcing all pages to endlessly pursue the exact same phrasing.
If there had still been dozens of B pages at that time, I might not have rewritten all of them just to turn the number into A=103.
But with only 2 pages left and a very low modification cost, it was reasonable to go ahead and complete the final language optimization.
Therefore, the final result:
A=103
B/C/D=0was not because:
B must also be fixed.
but because:
B was already ready for release; it was just that only 2 pages remained, the cost of continued optimization was low, and completing the final convergence was a convenient step.
4. Why I raised my overall evaluation from about 94 to about 98
In the previous round of official zh-CN quality evaluation, my internal engineering judgment was roughly:
94 / 100More precisely, this can be understood as:
93~95It was already considered a relatively high-quality official translation.
The main issue was not a large number of obvious errors, but rather that it was still possible to feel:
A slight translationese
Some expressions not being natural enough
Individual terminology still having room for further unification
Some pages could be closer to the style of an official Chinese technical tutorialThis was also the reason I continued researching minimal-protect, Static Context, and the Translation Engine.
The entire experiment ultimately proved that the real key to significantly improving language quality was not further reducing structural protection, but:
Changing the Translation Engine.
After ChatGPT completed the full-page retranslation of all 103 pages, followed by subsequent revisions and a complete semantic review, I am now more willing to place the overall quality evaluation of this version of zh-CN at:
About 98 / 100It is important to emphasize here:
The score of 98 is not an automated score calculated by a program, nor is it an official Go evaluation.
It remains an internal comprehensive judgment reached after combining the entire round of translation, review, and engineering validation.
The main evidence supporting this judgment includes:
103/103 completely retranslated
Unified semantic review for all pages
Final A=103
B/C/D=0
Further unified terminology
Known bad terminology scanned and cleared
Cleanup of normal English natural language omissions
Multiple rounds of revision
Same structural validator passed againCompared to the original score of about 94, the difference is no longer just a subjective feeling on a few representative pages; it ultimately covers all 103 official course pages.
5. Why A=103 is still not 100
Two concepts need to be distinguished here:
Tier A
≠
100-point perfect translationA means:
The current page has reached the official release quality set for this round, and there are no clear semantic, terminology, or expression issues worth continuing to block the release.
A score of 100, however, implies a significantly higher standard:
Almost no reasonable room for editorial improvementAlthough this round completed a page-level semantic review of 103 pages, it did not undergo a more costly process:
Having independent Go technical experts and professional Chinese technical editors perform publication-level final proofreading sentence-by-sentence across all 103 pages.
Moreover, natural language does not have a single correct translation.
A single English sentence could completely have two Chinese expressions that are both accurate and natural.
One might be more concise, while the other might better fit a certain editorial style.
Therefore, the remaining roughly 2 points are more accurately described as:
Uncertainty reserved for future editorial-level tweaks and extremely low-probability omissions.
It does not mean we already know that “2% of the sentences still contain errors.”
Thus, I believe:
About 98is more appropriate than:
100as an engineering quality evaluation for the current version of zh-CN.
6. A=103 was not achieved in the first review
Seeing the final result:
A=103
B/C/D=0makes it easy to think that the 103 pages were already in this state after the first check.
In reality, they were not.
After Batches 001–008 completed the first full coverage, subsequent reviews still found areas worth adjusting.
This led to the creation of:
Batch 009
Batch 010
Batch 011These batches were no longer for filling in missing pages.
They were specifically tasked with:
Terminology unification
Clear language quality revisions
Final optimization of a small number of B-tier pages
Final expression revisions for individual pagesTherefore, once historical batches were formed, I did not go back and directly modify the original evidence.
Instead, I followed the principle:
Old batch
→ remains unchanged
Quality issues discovered
→ create a new revision batchThis allows the final page to not only know:
What the current translation in use is.
but also to continue tracing:
Why it changed from an earlier version to the current version.
7. The final canonical does not all come from the first full batch
This can be seen directly from the provenance of the final status.tsv.

welcome/1 still comes from Batch 001, while concurrency/1, methods/17, and methods/19 adopt revisions from subsequent Batch 009 and Batch 011 respectively; all ultimately reach ready status and pass the canonical validator.For example:
welcome/1
→ chatgpt-zh-CN-001indicates that the early version of this page already met the requirements.
There was no need to rewrite an already good translation just to make all pages come from the latest batch.
Whereas:
concurrency/1
→ chatgpt-zh-CN-009indicates that it ultimately adopted a subsequent revision.
Another example:
methods/17
methods/19
→ chatgpt-zh-CN-011These were not frozen as the final version until the very last revision batch.
So the principle of canonical promotion is not:
Find the first 103 pages with complete coverage
↓
Overwrite all official translationsbut rather:
Check each page
↓
Determine the final valid revision for that page
↓
Verify complete provenance
↓
Then promote to official canonical8. Historical evidence cannot be modified just to make official files “look clean”
When actually preparing for canonical promotion, I encountered an issue that seemed minor but was actually very important.
The candidates saved in historical batches did not have completely uniform file endings.
After taking statistics, I found:
23 candidates
→ one newline at the end
80 candidates
→ two or more newlines at the endIf these historical candidates were copied byte-for-byte directly into the official canonical:
git diff --checkwould produce warnings related to trailing whitespace.
The simplest way to handle this seemed to be:
Directly delete the extra newlines from the historical candidates.
But I did not do this in the end.
Because these historical files already serve the role of evidence.
If I modified the historical candidates just to make the final submission cleaner, it would break the previously established:
raw response
↓
restore
↓
historical candidateexact correspondence.
The final frozen rule became:
Historical evidence
raw
↓
restore
↓
batch candidate
Keep historical bytes unchangedThen, at the official promotion boundary:
Promotion boundary
batch candidate
↓
deterministic EOF canonicalization
↓
validator
↓
canonical candidateOnly when crossing into the official canonical is deterministic EOF normalization executed.
The rule is very simple:
Only delete extra \n at the end of the file
Ensure exactly one \n at the end
Body text and other bytes remain unchangedThis protects the historical evidence while ensuring that the candidate files in the official repository remain standardized.
9. Canonical promotion must also prove the translation actually comes from the corresponding ChatGPT evidence
When actually preparing for promotion, I specifically strengthened the evidence chain.
Because just knowing:
a candidate exists in a certain batchwas not enough.
Theoretically, if someone manually modified a candidate later, and promotion only checked the final structure, it would still be possible to officially promote a file with unprovable origins.
Therefore, the final promotion was not just “find the file and copy it.”
It also required re-proving the entire chain of origin.
This roughly included:
input recorded in the manifest
↓
input SHA at save time
↓
current Default protector regenerates input
↓
byte-level comparison
↓
ChatGPT raw response
↓
restore
↓
historical candidate
↓
exact byte comparison
↓
validator
↓
promotionThis way, the provenance of the final official candidate was no longer just:
This file is located in the ChatGPT Batch directory.
but rather:
It can be proven that it indeed comes from the saved ChatGPT raw response, and has passed through deterministic restore and the same validator.
10. When retry history exists, the final valid attempt must also be correctly identified
The previous article documented two official retry pages:
moretypes/1
concurrency/1Upon reaching the promotion stage, another issue needed to be resolved:
If a page has a retry history, can promotion correctly identify and select the final valid attempt?
For example:
attempt-001
→ validation failed
attempt-002
→ validation succeededPromotion then needed to prove:
That the valid result corresponding to attempt-002 was the one ultimately promoted.
Just because attempt-001 was still retained in the historical directory, it could not erroneously fall back to the earlier failed result during subsequent processing.
Therefore, retry provenance rules were further strengthened.
The final requirements were similar to:
Retry history exists
↓
Must identify the final valid attempt
Attempt numbers must be consecutive
Prior validation history must exist completely
attempt-999 cannot appear out of nowhere
Historical attempts cannot be skipped
Provenance cannot regress to an already failed old attemptThis part seems very strict, but the purpose is simple:
Official promotion must not only prove “this file works,” but also prove “it indeed comes from the final confirmed valid attempt in this retry history.”
11. Official apply: 102 out of 103 pages changed
Only after all semantic reviews and promotion mechanisms were frozen was the official apply actually executed.
The results were as follows:

First official apply:
pages : 103
changed : 102
unchanged : 1
EOF normalized : 80The:
unchanged : 1here does not mean that one page did not use the new ChatGPT translation.
It means that the current canonical candidate for this page was already exactly identical to the final target, so there was no need to produce a new file diff.
What is truly worth noting is:
changed : 102102 out of the 103 official candidates actually changed.
At the same time:
EOF normalized : 80exactly corresponds to the 80 historical candidate files found earlier that needed EOF standardization.
12. Why another dry-run is needed after apply is complete
After the official apply was completed, I did not commit immediately.
Instead, I re-ran a promotion dry-run.
The result:
changed : 0
unchanged : 103This step verifies:
Whether promotion has reached a deterministic, stable final state.
If the second dry-run still showed:
changed > 0it would mean that promotion itself might still have issues like:
Non-deterministic processing
State not fully persisted to disk
Unstable file generation resultsUltimately:
0 changed
103 unchangedproved that recalculating with the same inputs and the same historical evidence would not continue to modify the canonical.
In other words, it possessed the crucial property of:
Idempotency.
13. The final official commit contains only 103 data files
After promotion was complete, I performed one final check on the Git change set.

bfa9f92 for the official ChatGPT canonical promotion. This commit contains only 102 changed official candidates and 1 status.tsv, totaling 103 files, with no unexpected paths; the working tree remains clean after the commit.Final commit:
bfa9f92
data: promote ChatGPT retranslation candidates to official zh-CN candidatesStatistics:
103 files changed
628 insertions(+)
533 deletions(-)Classified by path:
canonical candidates : 102
status.tsv : 1
unexpected paths : 0That is:
102 official candidates
+
1 status file
=
103 filesThis commit did not casually mix in:
promotion implementation code
test code
historical batches
raw responses
other project filesThis is also a boundary I want to maintain long-term:
Official data promotion and code changes implementing the promotion should not be mixed in the same Git commit.
14. After promotion was complete, another “test failure” occurred
After the official candidates were replaced, the full test suite did not pass entirely on the first run.
Several tests failed at that time.
After further investigation, it turned out that the new ChatGPT translation had not broken the program.
The real reason was:
The old tests treated the previous version’s official Chinese sentences themselves as a fixed test baseline.
For example, a test had directly depended on a specific old translation.
When ChatGPT changed it to a new expression with the same technical meaning but more natural Chinese:
Program was correct
Structure was correct
Page was correctthe test failed because:
the Chinese sentence was no longer identical word-for-wordThis actually exposed another issue:
What exactly should tests protect?
If the goal of the test is to ensure that a page still maintains correct publication semantics, then it is more appropriate to check:
Stable semantic anchors
Page structure
Publication order
Status fields
Provenancerather than permanently freezing entire Chinese sentences.
Therefore, I only updated the test baselines related to the new canonical.
Ultimately:
go test -mod=readonly -count=1 ./...passed completely.
At this point, the new ChatGPT canonical truly completed repository-level official validation.
15. Moving from about 94 to about 98 adds more than just model capability
Looking only at the surface, this round of changes can be simply summarized as:
GLM-5.2
↓
ChatGPTThe change in Translation Engine was certainly the key reason for the significant improvement in language quality.
But what really pushed the original score of about 94 to what I am willing to evaluate as about 98 was not just the model itself.
It also included the subsequent:
103 pages full coverage
↓
Complete semantic review
↓
A/B/C/D grading
↓
Terminology unification
↓
Revision batches
↓
Final A=103
↓
Provenance verification
↓
Retry provenance
↓
Canonical promotion
↓
EOF canonicalization
↓
Re-run dry-run
↓
Full testsIn other words:
A stronger Translation Engine provides better raw translations, while a complete engineering process is responsible for proving that these translations are ultimately worthy of entering the official project.
Without the former, language quality is hard to improve significantly.
Without the latter, even if the model translates well, it is hard to confidently replace 103 official pages all at once.
16. Translating 103 pages and 103 pages being officially releasable are two different milestones
In retrospect, one of the most valuable lessons to retain from this experience is not to conflate:
translation completedwith:
release readyas the same state.
By the end of the previous article, we had achieved:
103/103 ChatGPT retranslation completedBut only by the end of this article did we complete:
103/103 semantic review completed
↓
2 pages were B, but already met release quality
↓
Continued optimization due to low modification cost
↓
Final A=103, B/C/D=0
↓
Final revision confirmed
↓
Canonical promotion
↓
102 candidate changes
↓
103/103 promotion converged
↓
Full tests passedAt this point, I am finally willing to say:
The new ChatGPT zh-CN is no longer just an experimental candidate, but the official canonical.
Combining the old version and the complete review results of these 103 pages, my internal engineering evaluation of the overall quality of the current A Tour of Go Simplified Chinese has also increased from the previous approximate:
94 / 100to:
About 98 / 100But even after reaching this point, there is one last question left unanswered:
Does replacing the official canonical in the repository with the ChatGPT translation mean that users will definitely see this new translation when they open the website?
The answer is still no.
Official release, production deployment, public web pages, CDN caching, and real Run / Format all belong to the next layer.
Therefore, the third article will no longer discuss the Translation Engine, nor will it repeat the deployment script design process written about previously.
It will only answer the final question:
How to prove that the 103 official course pages in the production environment have truly all been updated to this round’s new ChatGPT translation, which underwent complete semantic review and canonical promotion.
