What the sources actually say
The official English documentation explains the threat model first. In the simplest case you are dealing with passive DPI: equipment that can read passing traffic and inject its own packets, but cannot drop your packets. If a request is prohibited, passive DPI injects its own RST packet toward the client and optionally an HTTP redirect.
If the fake packets from DPI are only sent to the client, carefully written firewall rules can drop them, but that requires manual, per-ISP traffic analysis. More importantly, the documentation states that if the DPI also sends an RST to the server, there is nothing you can do after the fact. That is why the project is aimed at preventing the ban trigger from firing, not at eliminating its consequences.
Prevention works by sending what the DPI does not expect: traffic shaped so the DPI fails to recognize a prohibited request, while the destination server still reconstructs it correctly. This is the core idea the documentation calls DPI desynchronization.
How it works, step by step
- Your application sends a normal request, for example an HTTPS handshake to a blocked site.
- Zapret intercepts the outgoing traffic on your own machine or router. On Linux this is done with an NFQUEUE handler (nfqws); a stream-level variant (tpws) works through a transparent proxy.
- Instead of forwarding the request as-is, Zapret modifies it: it can split the request into TCP segments, inject fake packets the DPI will see but the server will reject or never receive, or apply small header-level changes.
- The documentation gives concrete stream-level examples: some DPIs do not recognize an HTTP request split into two TCP segments (first "GET", then the rest); others stumble when the Host header is written in a different case, when an extra space is added after the method, or when a dot is appended to the hostname.
- The server reassembles the original, valid request and responds. The DPI, having seen something it cannot match against a prohibited signature, does not fire the block trigger.
Why prevention instead of cure
The documentation is explicit about the asymmetry: once a passive DPI has injected an RST toward the server, the connection is dead and no client-side tool can undo it. Everything Zapret does is therefore designed to keep the DPI from ever classifying the request as prohibited. This also explains why effectiveness is ISP-specific: different DPI vendors and configurations fail on different tricks, so a strategy that works on one network may do nothing on another.
What is source-backed vs community-reported
- Source-backed (official documentation): the passive DPI model, RST and HTTP redirect injection, the statement that server-side RST cannot be undone, and the prevention-first design goal.
- Source-backed (official documentation): the segmentation, Host-header case, extra-space, and trailing-dot examples of requests that evade some DPIs.
- Community-reported: which specific strategy works on which ISP is tracked in Flowseal discussions and issues, and changes over time. Those reports are evidence about particular networks, not general laws.
When the reported workaround may not apply
Per the official documentation, the whole approach fails when DNS answers are falsified and cannot be protected (for example via DNS-over-HTTPS or DNSCrypt), when blocking is done by IP address, or when traffic is forced through a full TCP-terminating proxy. In those cases no packet-level desynchronization helps, because the problem is not DPI classification at all.