TikChat YouTube browser sample — November 14, 2026 FIRST-PARTY METHOD One real production TikChat session and one public YouTube broadcast. Commissioned by Matt Olech, owner of TikChat; browser observation and comparison executed by an AI assistant. The native reference was sampled, not an independently complete record. The observation window was 5940.299 seconds; final reception checks lasted another 97.248 seconds. ### Establish the native chat and production session The owner started the YouTube source in TikChat. The observer reused that operator tab and opened the native YouTube chat separately. The native view was set to Live chat, and the browser structure and ID transformation were checked before formal collection. A calibration found 205 shared message IDs. Of those, 203 also had matching normalized content, while two involved differences with emotes. These calibration entries were not reported as additional study results. ### Exclude the initial reference buffer A newly opened chat can contain messages posted before observation begins. We recorded the starting buffers and excluded **212 initial native YouTube IDs** from the new reference count. The initial TikChat buffer could still confirm an entry that became visible in the native reference later. ### Read both browser views repeatedly The observer aimed to read both chats about every five seconds. The completed record contains **932 native reference reads** and **945 TikChat reads** after the baseline. The latter includes 13 reads during the final comparison period. Each reference read inspected up to 500 recent native chat rows. Each TikChat read inspected up to 800 recent YouTube message rows. Reading the same ID again did not increase the count of distinct reference messages. Actual read intervals, buffer overlap and exclusions were logged. ### Match IDs and compare message content Matching used provider message IDs transformed into the form represented by TikChat. It did not rely on similar wording, usernames or AI judgments about meaning. For content, the checker read the message text and the alt labels of inline emote images. It normalized Unicode to NFC, collapsed consecutive whitespace and trimmed whitespace at the edges. Letter case and punctuation remained significant. Temporary hashes allowed comparison without retaining the comment text in the evidence files. The checker distinguished matching normalized content, differences involving emote representation with matching ordinary text, and other unresolved content differences. It also checked for differences limited to emoji variation selectors; none were classified in that category. ### Save aggregates and complete the final comparison We saved progress aggregates approximately every five minutes. An unmatched reference entry was initially pending; after 60 seconds without a match, it was classified as unresolved. It could still be matched later. After the owner confirmed the natural end, we stopped adding reference entries and continued reading TikChat for 97.248 seconds. The final ID totals did not change during that period. Working message pseudonyms were then cleared, leaving the aggregate evidence and sampling record. The measurement used local browser reads. The observer did not start another TikChat collector, add a separate TikChat API polling loop, connect to the VPS, deploy code, restart services or change application settings. Server resource use was not measured. LIMITATIONS ### The native reference was sampled The median interval between native reads was **5.086 seconds**, the 95th percentile was **8.103 seconds**, and the longest interval was **75.053 seconds**. Eighteen successive nonempty native buffers had no shared IDs, signalling possible observation gaps. The number of messages missed by the observer is unknown. The first formal reference read began 51.610 seconds after the observation start. The last ended 48.967 seconds before the reference window closed. The 99-minute window therefore must not be described as continuous recording of every YouTube message. ### Both browser views had bounded buffers The TikChat reader deliberately used only the latest 800 matching rows. The log marks that bound as truncation; it is not a count of application failures. Unsupported native DOM rows were excluded from the message denominator. Their repeated appearances in the log are not distinct comments or proven lost messages. ### Some content representations changed The observer recorded **259 changes of the reference content digest for already-known IDs**, and zero such changes in the operator view. This counts changes, not necessarily 259 different messages. The checker kept the latest observed content representation. The causes of those changes were not established. The final content result is consequently a comparison of observed browser representations under the stated rules. It is not a manual visual inspection of every emote or a finding that all 132 other differences were defects in TikChat. ### Observation timing was not transport latency Fifty-two IDs were first found in TikChat more than 60 seconds after their first native observation. Read intervals and pauses by the observer contribute to that difference. It cannot be used as a count of messages proven delayed by the transport, or as a precise latency benchmark. ### This was one first-party production session The owner commissioned the test, and an AI assistant performed the observation. There was no independent reviewer, random sample of broadcasts or control group. The audience reading was a single pre-test context point, not a peak measurement or an audience time series. The evidence is from the browser views. We did not independently audit the production database, record a production release fingerprint or measure server CPU, RAM, I/O or network use. It is therefore not a server load benchmark. The detailed messages and working pseudonyms are not retained for later message-level re-auditing. AUDIENCE CONTEXT Approximately 85000 concurrent viewers in one pre-test browser reading. This supplementary context was recorded in the research conversation, not as a continuous audience dataset. See audience-context.json. No peak, average or server capacity was measured. REPRODUCIBILITY The research-package.zip archive preserves the aggregate record, complete read log, definitions, comparison source and synthetic tests. Its verification script checks aggregate arithmetic and integrity. Message-level matching cannot be replayed after disposal of working data. The publication study-record.json keeps the numerical study results unchanged and adds a supplementary audience note; the original snapshot remains in the ZIP.