AI Demos Research — the structured-intelligence platform. Every verdict on these pages opens to the execution behind it.
Graded 29 October 2026

Does Landing AI keep code blocks intact in Markdown?

Landing AI did not keep the PDF’s code blocks intact in Markdown. With no fenced or other preformatted wrapping, the code rendered as paragraph text.

1 of 1 test case failed

Every test case in this scenario has a result.

Pass rate0%0 of 1 gradable
Coverage1 of 1test case with a result
0PassNo test case passed.
1FailMissed at least one thing it was expected to do; the reason and proof are with each test case below.
0Not gradableEvery result here could be graded.
0UntestedEvery test case in this scenario has a result.
The pass rate is a summary. The evidence is the test case below: what we sent, what we checked, what the tool returned, and the proof.

The test case

Each test case is judged on its own: Pass, Fail, Not gradable, or Untested. The scenario result above is based on this test case.

Code block — fenced, structure intactFailEvidence
What we sent

We submitted the complete three-page operations-note PDF to Landing AI’s document-to-Markdown API in one synchronous request.

InputDocument (PDF)
Input PDF: operations note with Python and shell code blocks on page 2
Open the PDF ↗
What the tool returned

Copied from Output file (Markdown), lines 61–68, 86–88

Markdownthe whole file · cited lines 61–68 and 86–88 marked
Tarnbeck Institute of Hydrology · Data Services
# Reloading a station export after a gauge reset

Operations note DS/OP/07 · Data services · Reviewed 14 January 2026


**Symptoms**
The station's daily total is short and the loader's own log says nothing is wrong. In the log the day
looks like a clean run: the export was fetched, every block was read, and the summary line reports
fewer readings than the station's counter did. The gap is always at the end of the day rather than
the beginning, and it is always a whole number of blocks.


The other sign is a duplicate warning in the block log, with two entries carrying the same sequence
number and different checksums. One warning is normal after a reset. A run of them means the
reset happened part way through the day and everything after it is being rejected block by block.


# Why the loader rejects the second block

Block sequence numbers come from the station and not from the loader. The loader keeps the
highest number it has committed for the day and treats anything at or below it as a repeat, which is
the right thing to do on a poor link: a station that retransmits sends the same block twice, and the
loader must not count it twice.


A reset returns the counter to zero, so the blocks after it carry sequence numbers the loader has
already committed. Nothing about them looks new and the loader discards them for exactly the
reason it should. The reset time is the only thing that distinguishes a continuation from a repeat,
and the loader does not know it unless it is told.


**The fix**
When a gauge is reset in the field the station's counter returns to zero, and the export for that day
contains two blocks carrying the same sequence number. The loader accepts the first and rejects the
second. That is the right behaviour for a duplicated transmission and the wrong behaviour here, and
the day's record ends up short by however much arrived after the reset.


The fix is to requeue the affected blocks with the reset time supplied, so that the loader treats the
second block as a continuation rather than a repeat. The function below is the one to call. It lives in
tarnbeck.stations.maintenance and is safe to run more than once: a block that already carries a
good checksum is skipped.


DS/OP/07 · Data services


Page 1 of 3


<!-- PAGE BREAK -->

Tarnbeck Institute of Hydrology · Data Services


DS/OP/07


from tarnbeck.stations import Export, reload_window
def reload_after_reset(station, reset_at):
   export = Export.open(station)
   window = reload_window(reset_at, margin_minutes=45)
   for block in export.blocks(window):
       if block.checksum_ok():
           continue
       export.requeue(block, reason="gauge reset")
   return export.commit(dry_run=False)


Run it from the loader host and not from a workstation, and take the reset time from the field log
rather than from the station clock, which is the thing that was wrong. The margin defaults to
forty-five minutes either side of the reset; widen it only if the field log is vague about the time,
because a wide window requeues blocks that were never in doubt.


If commit returns a count lower than the number of blocks in the window, stop there and raise it
with data services before running anything again. A short count means blocks were rejected for a
second reason, and requeuing them repeatedly will not find out what it was.


**Running it**
Run it from the loader host, in the maintenance environment, with the station code and the reset
time as they appear in the field log:


$ ssh loader-01.tarnbeck.example
$ tarnbeck-maint reload-after-reset --station VQ1 \
--reset-at '2026-01-29T09:15' --confirm


The command prints the window it will use, the number of blocks in it and the number it intends to
requeue, and then waits. Read the three numbers before answering. If the window is wider than the
field log justifies, stop and narrow it rather than letting it run.


Without --confirm the command does everything except commit, which is the safe way to see
what it would do. The dry run takes a few seconds and there is no reason to skip it.


**Afterwards**
Check the day's totals against the gauge's own counter before closing the ticket. The loader reports
the number of blocks committed and not the number of readings, and a day can come back with
the right number of blocks and the wrong number of readings if a block was truncated in the field.


Note the reset in the station log, with the time taken from the field log and not the time the reload
ran. The next person to look at a gap in the series will look there first, and a reload with nothing
written beside it is indistinguishable from a gap nobody has dealt with.


DS/OP/07 · Data services


Page 2 of 3


<!-- PAGE BREAK -->

Tarnbeck Institute of Hydrology · Data Services


DS/OP/07


**If the count is short**
A short count means blocks were rejected for a second reason and the reload has not found out
what it was. Do not run it again: a second pass requeues the same blocks, reports the same count,
and fills the log with attempts that bury the original failure.


Take the sequence numbers the command reports as rejected and look them up in the block log. A
checksum failure on a block truncated in the field cannot be fixed by reloading and needs the
station visited. A rejection with no reason recorded is a loader fault, and goes to data services with
the run's log attached.


**The block header, for reference**
Each block in an export begins with a fixed header, which is what the loader reads to decide
whether it has seen the block before:


SEQ=00184 STATION=VQ1 START=2026-01-29T08:45Z 
COUNT=0180 INTERVAL=15s CRC=8f2a41d6 

SEQ is the station's counter and is the field a reset returns to zero. START is the station clock, which
is not to be trusted after a reset - which is why the reset time comes from the field log and not from
the export.


CRC covers the readings and not the header, so a block with a good checksum and a wrong
sequence number is possible and is exactly what a reset produces. Nothing in the header says a
reset happened.


**Rolling back**
A committed reload can be undone within the same day by reverting the station's day to the
snapshot the loader takes before it commits. The snapshot is kept for seven days and is named for
the run.


After seven days the only route back is a fresh fetch of the export from the station, which is possible
while the station's own buffer still holds the day - about three weeks - and impossible afterwards. A
reload that has committed and cannot be rolled back is an incident and is raised as one, whatever
the size of the gap.


DS/OP/07 · Data services


Page 3 of 3

<!-- doc_id=parse-01m22n7bvjsdy3vaqw0b9293e9 -->

Returned Markdown: unfenced code with changed indentation

Open the Markdown file ↗
Expected vs. Found
✗Expected: Emit the code as preformatted or fenced blocks, preserving its line breaks and indentation.Found: Python indentation became 3/7/11 spaces rather than 4/8/12; the shell continuation became flush left. Each code line kept its own line, with relative Python nesting preserved.
Supporting proof
Proof 1Screenshot
Side-by-side comparison: source code block and returned Markdown (the image renders the output flush left; the Markdown file shown above keeps 3/7/11-space indents).
Open original ↗
Proof 2Screenshot
Run summary card: successful conversion of the complete PDF using the default pipeline
Open original ↗
Why this result

The full PDF converted successfully with the default pipeline, yet the required code-block formatting was absent. The API documentation offers no code-rendering option that could change this outcome.

Also observed in this test case

Code characters unchanged. The returned Python snippet and shell command contained no altered or invented code characters.

Blank lines removed. Blank lines separating the Python import, function definition and loop in the source were absent from the returned Markdown.

Tested by Chandresh Bisht · evidence dated 9 September 2026

Configuration and setup

How this tool was set up for the run and what the test needed in place. Each row is a fact from the run's records; a fact the records do not hold is left out, not guessed.

Software that produced the output
Landing AI
Build or version
ADE Parse API v2; model_version dpt-3-pro-20260710
Plan or tier
Explore (free tier)
Surface
API
Set up before the run
The complete gauge-reset operations note, containing Python and shell code blocks, had to be converted to Markdown in one pass using the default pipeline without product-specific tuning.
Tested
9 September 2026 · Chandresh Bisht

How this scenario is graded

How we decide Pass, Fail and Not gradable. The same rules apply to every tool tested on this scenario.

How results are decided

Each test case gets one result per tool: Pass, Fail or Not gradable. A test case we haven't run yet shows Untested. There are no partial results.

The rules
  • Pass: the tool did everything the test expected, and nothing it said contradicts the correct answer.
  • Fail: at least one expected behaviour clearly didn't happen; the test case's result says which one and quotes the tool.
  • Not gradable: our evidence couldn't settle the outcome (for example a record we needed is missing). It is never counted as a fail, and the test case's result says what's missing.

Where this sits in the benchmark

This page is one cell of a larger study: one tool, one scenario. Only this benchmark's frame appears here.

History of this result

What has happened to this result since it was first published. Runs and grades are never overwritten: a retest or a re-grade publishes a new result and keeps the earlier one readable.

from the publication record
21 November 2026First publishedConverting a complex PDF into clean Markdown with a hosted API v1

Act on this result

Nothing filed here edits the run or the grade. A challenge opens a review, and a review can produce a new run or a re-grade — which becomes the current result and leaves this one in the history.

This matches what I see

You run the same kind of test against your own setup and get the same behaviour.

Agree →
This does not match

Yours behaves differently. Tell us what you got, with a screenshot if you have one.

Disagree →
Point out an issue

Something here is wrong — a reference value, a transcription, a grade.

Report an issue →
Request a retest

On a newer build, a larger dataset, or your own setup.

Request a retest →
We have fixed this

Tell us what changed and we schedule a rerun of the failing test case. The old result stays as history.

Vendor notice →
Filed against this evidenceNothing yet. Challenges, counter-evidence and fix notices appear here with their outcome, and stay on the page after they are resolved.
Cite this result
aidemos.com/benchmarks/pdf-to-markdown-apis/results/landing-ai/a-document-containing-a-code-block · 0 pass · 1 fail · coverage 1/1 · graded 2026-10-29

The same record is available as structured data through the AI Demos MCP server, with the counts, the coverage and every per-test-case reason carried as fields.

Verify the proof files

These files support this result. Open a file to inspect the original evidence.

Proof 1 · Screenshot ↗image/png · 204 KB
Input · Document (PDF) ↗application/pdf · 78 KB
Output file (Markdown) ↗text/markdown · 7 KB
Proof 2 · Screenshot ↗image/png · 96 KB
File fingerprints (SHA-256)

A fingerprint identifies the exact file used for this result.

Proof 1 · Screenshot7eb015e3373ab9c2372c64588261e8922868702c1debd0241079519705753f0f
Input · Document (PDF)cd27d1e77ec9818ff769bde1b49d95e711373888507af4b007054c27b7a52db8
Output file (Markdown)9d20c510df491f4201aeac1cc2aafa15e9997b2636379c08720dd87b860dc91d
Proof 2 · Screenshotc30a46e240a552ab820a7799ee7ed3793bf39f47347365c91b641e18d12d9417