Read phases without counting TLS twice
The waterfall positions each request relative to the first request in the capture. Its segments show available blocked, DNS, connection, send, wait and receive values. A value of minus one means that the browser did not provide an applicable measurement; it is not a negative delay to subtract from the timeline.
TLS negotiation is represented by the SSL timing field and is already included in connect. Adding it again inflates the total. The workbench reports a discrepancy when recorded duration differs materially from the available phase sum. It preserves the capture evidence rather than rewriting browser values to make the chart look consistent.
Distinguish overlapping time and different byte counts
Several requests can run at once. Summing their durations therefore does not give page-load duration. The request span measures the distance from the first recorded start to the latest recorded finish, while browser page timing fields describe separate milestones. A partial capture can miss important activity before or after this span.
Response body size describes the recorded transferred body. Decoded content size can be larger after decompression. Header size is another measurement, and some browsers report it as unknown. Known totals exclude minus-one values and display the unknown count. Body text removal changes the exported file size without changing what the original network measurement claimed.
Treat cache observations as limited evidence
A 304 status indicates revalidation, while an explicit browser cache flag can identify a cache observation recorded by that exporter. A Cache-Control header describes caching policy and does not prove that a particular request was served from cache. A zero body size also has more than one possible explanation.
Use these clues to decide which requests need investigation. A slow wait phase is not proof of slow application code: proxies, service workers, network behavior and server processing may contribute. Compare representative requests, check the same response and cache conditions, and repeat the capture when making a performance decision. Keep the original evidence private while sharing only the reviewed subset.