Chromium 156 reported Windows after a DevTools override, while its captured TCP SYN kept the same option layout as the Linux baseline. Heretic's TCP parser classified both captures as Linux. Its stack assessor reported stack.os-contradiction only when supplied with the Windows claim.
We tested two fresh headless Chromium processes on one Linux kernel, with a local HTTP server and packet capture. This answers a specific question: did the tested browser identity override also change the connection fields used by Heretic's TCP OS matcher?
Change the claims, then check the request
The browser was Chromium 156.0.8078.79, controlled with Playwright Core 1.56.1. Both processes ran in the same container image. Runtime networking was disabled except for loopback, and neither process used an existing user profile.
The baseline used the browser's defaults. Before navigating the second browser, we called Network.setUserAgentOverride with a Windows user-agent string and these platform fields:
{
platform: 'Win32',
userAgentMetadata: {
brands: [{ brand: 'Chromium', version: '156' }],
fullVersion: '156.0.8078.79',
platform: 'Windows',
platformVersion: '10.0.0',
architecture: 'x86',
model: '',
mobile: false
}
}The DevTools protocol definition specifies separate overrides for navigator.platform and the metadata returned by navigator.userAgentData and sent in Client Hint headers. We checked the executed result in the page and at the HTTP server.
| Observed value | Default browser | Overridden browser |
|---|---|---|
navigator.platform |
Linux x86_64 |
Win32 |
navigator.userAgentData.platform |
Linux |
Windows |
Received Sec-CH-UA-Platform |
"Linux" |
"Windows" |
| OS text in received User-Agent | X11; Linux x86_64 |
Windows NT 10.0; Win64; x64 |
The received User-Agent matched the page's value in each case. The default string included HeadlessChrome/156.0.0.0; the supplied string included Chrome/156.0.8078.79. We changed these OS-identity fields together, so this experiment does not attribute effects to one override field in isolation.
Match each request to its SYN
A browser can open more than one connection. Selecting the first packet seen on a listening port would leave an attribution gap.
The HTTP server recorded the peer source port for the document request. The capture selected an incoming SYN with that source port and the server's destination port, with SYN set and ACK clear. Exactly one captured SYN matched each request. The capture skipped the outgoing duplicate produced by Linux loopback and rejected truncated reads or reported packet drops. Both capture processes reported zero kernel drops. The saved records retain the document request's source port. The capture checked the listener port at runtime; the extracted reproduction now retains that port too.
The saved frames passed through Heretic's actual Ethernet, IPv4 and TCP parsers. Both produced:
| Captured field | Default browser | Overridden browser |
|---|---|---|
| Observed TTL | 64 | 64 |
| Advertised MSS | 65495 | 65495 |
| TCP window | 65495 | 65495 |
| Window scale | 10 | 10 |
| Option bytes, length | 20 | 20 |
| Option order | MSS, SACK permitted, timestamp, NOP, window scale | MSS, SACK permitted, timestamp, NOP, window scale |
Here is the option-byte structure from both SYNs, with the changing timestamp value shown as TSval:
02 04 ff d7 | 04 02 | 08 0a TSval 00 00 00 00 | 01 | 03 03 0a
MSS 65495 SACK timestamp, echo zero NOP scale 10Sequence numbers, IP IDs and timestamp values differed. The result concerns the parsed fields and layout, not identical packets. The large MSS belongs to this loopback fixture; it is not an Ethernet-path measurement.
Execute the stack comparison
We inserted each captured SYN and its observed browser OS claims into a synthetic admitted session, then called Heretic's actual stack assessor. The rest of the admission and verdict pipeline did not run.
The current matcher associates this option layout and window scale with its Linux catalogue entry. The stack assessor reads admitted UAData platform first, then navigator platform, then the effective User-Agent. It compares that claimed family with the matched family.
| Browser claim | TCP catalogue match | stack.os-contradiction |
|---|---|---|
| Linux | Linux | Absent |
| Windows | Linux | Present, Composite |
Both rows also produced the same weak path-MTU evidence from the loopback MSS. Absence of the OS mismatch in the first row is not a clean bill for a headless browser. This test executed only the stack comparison.
The catalogue's internal eligibility weight is uncalibrated. It does not turn these two observations into a detection probability. The captured Linux pair verifies this particular match and mismatch; it cannot establish that no other TCP implementation emits the same layout.
Change a socket setting instead
A separate four-connection control used Python sockets on the same Linux kernel. Its application OS labels were synthetic. Setting IP_TTL to 128 changed captured TTL from 64 to 128. Setting TCP_MAXSEG to 1200 changed the advertised MSS from 65495 to 1200, with the window changing from 65495 to 64800.
All four socket controls retained the same option order and window scale 10. Heretic's parser still selected Linux. Windows labels triggered the stack mismatch at both TTL values; Linux labels did not, including the reduced-MSS case.
These controls rule out treating every changed TCP signature as a changed OS. They also show why the browser result must stay tied to the intervention: the DevTools override left these TCP fields unchanged, while socket settings changed selected fields without changing the catalogue match.
Update, 10 November: when kernel settings remove the match
An additional experiment on 8 November kept Chromium 156.0.8078.79 and the container image fixed, then varied two Linux settings. Each of four isolated containers ran a fresh native-Linux browser and a fresh Windows-claim browser. This adds eight captures to the original pair.
Linux documents tcp_timestamps=0 as disabling timestamps and tcp_sack=0 as disabling selective acknowledgments. We set these at container creation and read them back inside each container before starting Chromium. The kernel documentation defines those controls; the captured packets establish their option order in this experiment.
| Timestamps | SACK | Observed option order | TCP catalogue result | Windows-claim OS mismatch |
|---|---|---|---|---|
| On | On | MSS, SACK permitted, timestamp, NOP, window scale | Linux | Composite |
| Off | On | MSS, NOP, NOP, SACK permitted, NOP, window scale | Linux | Composite |
| On | Off | MSS, NOP, NOP, timestamp, NOP, window scale | Linux | Composite |
| Off | Off | MSS, NOP, window scale | Unknown, uncatalogued layout | Absent |
All four native-Linux controls also had no OS mismatch. Each selected SYN matched the document request's source and listener ports; all eight capture processes reported zero drops. The independent decoder agreed with Heretic's recorded parser output. As in the original experiment, the stack comparison used captured packets and actual browser claims in synthetic admitted sessions.
With both options disabled, the captured options were:
02 04 ff d7 | 01 | 03 03 0a
MSS 65495 NOP scale 10The Windows claim still reached the page and HTTP server. The matcher returned layout-not-catalogued, so the stack assessor emitted no OS contradiction. Disabling either option alone exercised recognised alternative Linux layouts; disabling both removed the catalogue match.
This is an observed limit of this finding. The browser's declared OS remained altered, but the available TCP layout did not supply the match required by this comparison. We did not run the complete assessment or test a different operating system emitting the same layout. These results neither certify the other catalogue entries as unique nor show that a modified browser passes Heretic.
A terminating proxy introduces another endpoint. Our earlier forwarding-versus-termination experiment shows how to identify that boundary. This browser pair contains no proxy, and supplies no comparison across kernels or named anti-detect products.
For a comparable investigation, retain the actual browser values, the request's connection identifier and the selected SYN. Report the packet changes separately from the catalogue match and named finding. That record lets a reader distinguish a failed override, a changed connection, and an OS mismatch produced by the tested rule.
The expanded reproduction package contains the capture programs, saved frames, independent decoder and recorded Heretic parser and stack-assessor outputs. Run python3 option-controls/check-matrix.py for the eight additional captures. Run python3 check.py to verify the saved packet fields and browser/header agreement without Docker. The decoder does not implement Heretic's OS catalogue. Fresh capture instructions require Linux Docker with raw-socket capability; image rebuilds may install a newer Chromium from Debian. The recorded browser version is 156.0.8078.79.