TikChat passive Twitch browser observation, November 15–16, 2026 Unit: author-plus-plain-text signature groups, NOT unique messages. Observation duration: 106 minutes 21.621 seconds. Sending and independent backend receipt: unmeasured. Observation started during the broadcast and ended at the owner request while the native page still indicated LIVE. METHOD ### Establish the native chat and production session Before formal collection, we saved the protocol, DOM reader, counter and prepared controlled-message plan with their SHA-256 hashes. Calibration was excluded from the reported sample, and fresh starting buffers were read immediately before the formal window. The operator's source scope was checked. Twitch showed Connected, search was empty and the Hide A–D replies filter was off. On this third-party channel we observed natural chat only: no synthetic messages were sent and no provider disconnects were forced. ### Exclude the initial reference buffer We excluded **130 groups already present in the starting native buffer**. Those groups remained excluded if the same author-and-text combination appeared again later. The initial TikChat buffer could establish prior presence for a group first observed later in the native reference. This prevents the starting backlog from being counted as new reference material. It also means the result is not a count of every message posted during the window. ### Read both browser views repeatedly The target was approximately one paired read every five seconds. The completed record contains **968 native Twitch reads** and **978 TikChat reads** after the baseline; ten TikChat reads fell within the final comparison window. Each read inspected up to 500 recent native rows or 800 recent matching Twitch rows in the operator. The native buffer actually contained 145–150 rows per read. Actual intervals, row exclusions, bounded reads and buffer-overlap warnings were recorded rather than assuming uninterrupted sampling. ### Match author-and-text groups and compare content Displayed authors were normalized to Unicode NFC, whitespace was collapsed and letter case was lowered. Plain text used NFC and collapsed whitespace with trimmed edges; its letter case and punctuation remained significant. Recognized CTCP ACTION wrapping was removed for presentation comparison. The grouping text excluded inline images. Their alt labels were included in a separate complete-content comparison. Emote-only rows with no usable plain text, and rows without a usable author or body, were excluded from the group denominator. Their repeated DOM observations are recorded separately. An ephemeral salt and temporary digests supported comparison without writing comments or viewer names to the evidence. A group was marked ambiguous when the same signature occurred multiple times in one buffer or its complete content representation changed. Ambiguous matched groups were excluded from unambiguous content-equality claims. ### Save aggregates and complete the final comparison Progress snapshots were saved approximately every five minutes. An unmatched group was initially pending. After 60 seconds without a match it became unresolved, while remaining eligible for a later match. At the owner's stop request, we closed the native-reference window and applied the planned 90-second final comparison window. Only operator reads inside its cutoff were counted. Working comparison maps and the salt were disposed after aggregate export. The observer made no deployments, service restarts, production fault injections, additional TikChat collectors or additional private-API polling loops. Normal application collection continued. Server resource use was not measured. LIMITATIONS ### Group identity is weaker than message identity Repeated author-and-text combinations collapse, and a later repetition can establish a group match without proving the earlier occurrence was received. Baseline groups remain excluded on repetition, and emote-only messages are outside the denominator. Twitch does provide message IDs in its IRC protocol; our limitation was the browser DOM available to the reader. See [Twitch's IRC documentation](https://dev.twitch.tv/docs/chat/irc/). Our [99-minute YouTube chat study](/blog/large-youtube-live-chat-test-99-minutes/) used message IDs, so its percentage and this group ratio must not be ranked as equivalent capture-rate measurements. ### The native reference was sampled | Read interval | Native Twitch | TikChat operator | |---|---:|---:| | Median | 5.002 s | 5.002 s | | 95th percentile | 13.090 s | 13.164 s | | Longest | 118.726 s | 118.729 s | The first formal native read started 16.141 seconds after the window began; the last ended 24.452 seconds before it closed. One consecutive nonempty native-buffer pair had no shared group keys. Overlap elsewhere does not prove full coverage, and the number of missed messages is unknown. The final 90-second comparison window contained ten operator reads. The first began 27.609 seconds after reference closure and the last ended after 73.311 seconds. The final 16.689 seconds were not sampled. No post-cutoff reads were included. This was not continuous recording. ### Both browser views had bounded buffers The operator reader deliberately used at most 800 recent matching rows; 966 reads were marked as bounded. That flag is not an application-loss count. Excluded or repeated rows in the log are repeated DOM observations, not distinct failed comments or confirmed duplicate-message defects. ### Observation timing was not transport latency Seven groups were first matched more than 60 seconds after their initial native observation. Sampling pauses and repeated content can contribute to that interval. It is not a measurement of provider delay or a transport-latency percentile. ### This was one first-party production session There was no independent reviewer, random sample of channels or separate backend-receipt measurement. The version checks were point checks, not continuous runtime attestation. CPU, RAM, I/O and network use were not measured, so avoiding deployments and restarts is not proof of zero resource impact. During initial calibration, Twitch displayed a video-playback-unavailable notice; the native chat subsequently loaded and was sampled. Video and audio were not the reference. An optional-variable cleanup warning after final export was corrected without affecting counts; it is disclosed in the error record. The planned controlled series of 120 synthetic messages was **prepared but not sent**. Late source attachment and provider reconnect were not tested. Joining our observation midway through a broadcast is different from measuring the application's source attachment, and refreshing an operator tab would not test the server's provider connection. LEGACY FIELD DEFINITIONS sampled: new distinct normalized author-plus-plain-text groups observed after baseline exclusion; NOT unique comments matched: reference groups for which at least one same author/plain-text signature occurred in the observed TikChat DOM pending: no match yet and first observation no more than 60 seconds old unresolved: no established match after the observation threshold/final cutoff; not proven loss matchedPercent: matched / sampled groups, not all-message capture rate baselineExcluded: 130 native groups already visible at baseline; excluded even on later repetition exactRows: legacy counter field: distinct signature groups within that read; not provider IDs duplicateDomRows: within-read occurrences beyond the first of the same signature; may be real repeated comments, not duplicate DOM bugs; observations repeat across reads unkeyedRowObservations: rows without usable author/body, summed across repeated reads unknownRowObservations: unsupported rows or empty plain text, including emote-only rows, summed across repeated reads truncatedReads: bounded extraction of last 800 operator rows when more were present; not app reception failure possibleBufferGap: successive nonempty buffers had zero shared keys; only a warning, not a count of lost messages; nonzero overlap cannot prove completeness content.normalizedEqual: complete rendered representation agreed in a matched group not marked ambiguous; pairing itself fixes plain text and author content.ambiguousMatchedGroups: matched groups with same-key co-occurrence or content revisions; excluded from an unambiguous content-equality claim referenceRevisions/operatorRevisions: digest change events for a known group, not necessarily different messages; includes baseline byKind.text: legacy reader class for accepted native text-bearing rows, not independently verified Twitch PRIVMSG categorization backendReceiptCheck: legacy fields in original export; reads=0 means backend was unmeasured, not that messages were not received cohorts: groups first seen in five-minute intervals, evaluated using final matches checkpoints: cumulative state at each actual snapshot time; pending/unresolved counts can change later samplingLog: relative read times, bounded row counts and coverage warnings; no comments or fingerprints The article presentation was prepared after the observation. The frozen prospective protocol and original evidence are preserved in research-package.zip. This text does not replace that protocol.