This is a guide i'm going to do some hand writing.

Solidly within the project's actual mission — 
this is standard digital-rights/circumvention guidance, 
the same category EFF and Access Now publish openly. 
Here's a categorized list to work from:

**Circumvention protocols/clients (the actual bypass tech):**

- **Shadowsocks** (shadowsocks.org) — SOCKS5 proxy purpose-built to defeat DPI-based blocking; the long-standing standard against the GFW specifically
- **V2Ray/V2Fly** (v2fly.org) and **Xray** — protocol suite built on top of/alongside Shadowsocks, widely used in China
- **Trojan** (trojan-gfw.github.io) — disguises traffic as ordinary HTTPS, harder for DPI to fingerprint
- **Hysteria** (hysteria.network) — UDP-based, designed to perform well even on high-loss/heavily-throttled networks
- **NaiveProxy** (github.com/klzgrad/naiveproxy) — HTTPS-camouflaged, rides on Chromium's own network stack
- **Clash / Clash Meta / Mihomo** — GUI clients that can run most of the above protocols, common entry point for non-technical users
- **AmneziaWG** — a WireGuard variant specifically modified to evade DPI (plain WireGuard's handshake is fingerprintable and gets blocked)

**Turnkey apps (less setup, good for less technical users):**
- **Psiphon** (psiphon.ca) — one-tap, widely used in Iran/China, adapts protocols automatically
- **Lantern** (getlantern.org) — P2P-assisted, designed for censored regions
- **Outline** (getoutline.org, Jigsaw/Google) — easy to self-host a Shadowsocks server for personal/small-group use
- **Tor Browser** (torproject.org) — the standard for anonymity, has built-in "bridges" (obfs4, snowflake) specifically for censored networks where plain Tor is blocked

**Messaging resistant to shutdown/surveillance:**
- **Briar** (briarproject.org) — peer-to-peer, works over Bluetooth/WiFi/Tor, **no central server needed at all** — keeps working during a full internet shutdown, closest fit to this project's actual scenario
- **Signal** (signal.org) — E2E encrypted, has domain-fronting history for reaching it when blocked
- **SimpleX Chat** (simplex.chat) — no persistent user identifiers by design
- **bitchat** — Bluetooth mesh messaging, no internet needed at all (you already had this in `checklist.md`)

**App distribution when Google Play itself is blocked/monitored:**
- **F-Droid** (f-droid.org) — open-source Android app store, common distribution channel for circumvention tools

**DNS-layer help (since EFW does DNS pollution specifically):**
- **Cloudflare 1.1.1.1 app** — DNS-over-HTTPS/TLS, defeats basic DNS injection; has a built-in WARP VPN mode too
- **NextDNS** — configurable DoH resolver

**Meta-resources worth linking to directly rather than just copying (these stay current as the arms race evolves, individual tool names go stale):**
- **EFF Surveillance Self-Defense** (ssd.eff.org) — comprehensive, actively maintained, multi-language
- **Access Now Digital Security Helpline** (accessnow.org/help) — direct human help for someone in an active crisis, not just docs
- **GreatFire.org** — specifically tracks Chinese censorship, mirrors blocked sites, tests circumvention tools
- **Open Technology Fund** (opentech.fund) — funds and vets many of the tools above, good index

One practical note for the KB: this space moves fast — censors patch detection, tools get blocked, new ones emerge. I'd frame this page as "last verified [date]" and lean on the meta-resources (EFF, GreatFire, OTF) as the living reference, rather than treating the specific tool list as permanent.


# Domain fronting, collateral freedom

❯ hi, how this kind of webpage delivering:  https://f001.backblazeb2.com/file/geph4-dl/geph-releases/dl.html

 **"collateral freedom" hosting**, domain fronting — they're related but distinct techniques.

**Domain fronting** (the classic technique): the TLS **SNI** field (visible to a censor) says one thing — a popular, unblockable domain — while the encrypted HTTP **Host header** inside says something else, routing the request to the real (blocked) backend once it's inside a shared CDN. Censors filtering by SNI/domain get fooled because the outer handshake looks like it's going to, say, a big cloud front-end. Most major CDNs (Google, Amazon CloudFront, Cloudflare, Fastly) deliberately broke this around 2018 by enforcing that SNI and Host must match at the edge, specifically to kill this technique.

**What that Backblaze B2 link is doing** is simpler: `geph-releases` is just hosted directly on a shared, multi-tenant storage domain (`f001.backblazeb2.com`). The SNI and Host header both honestly say `f001.backblazeb2.com` — no mismatch trick involved. The circumvention value comes from the fact that this domain also serves tons of unrelated third-party content, so a censor blocking it wholesale causes **collateral damage** to lots of legitimate traffic, which raises the political/economic cost of blocking it. That deterrent is the whole mechanism — it's sometimes called "collateral freedom" for that reason.

Practical implication: it's more fragile than true domain fronting was — GFW absolutely can and sometimes does block entire cloud-storage domains outright (it's done this to Amazon S3, Azure Blob, GitHub Pages, etc. during high-sensitivity periods), it's just a cost/benefit tradeoff for them rather than a technical dead end like fronting became.


# how GFW detect web traffics

Good instinct to lay it out — here's the fuller picture. The key thing that'll un-confuse you: **GFW isn't one filter, it's several independent systems stacked at different layers, some passive (watching) and some active (probing), and some checking only at connection-start while others keep watching mid-flow.**

**Layer by layer, what actually gets inspected:**

1. **DNS layer** — your query for `example.com` goes out (often over plaintext UDP/53, sometimes even DoH/DoT gets targeted). If the domain's blacklisted, GFW injects a fake/poisoned A record (fast, wins the race against the real answer) or just drops it. This happens *before* any connection exists.

2. **IP/TCP layer** — even if DNS resolved correctly (e.g., you hardcoded an IP, or used a different resolver), the GFW's routers watch IP+port. If the destination IP is on the blacklist, they inject TCP RST packets or just black-hole the SYN. This is independent of DNS.

3. **TLS handshake layer (SNI)** — this is the big one for modern HTTPS. Even though TLS encrypts everything *after* the handshake, the very first packet (`ClientHello`) contains the SNI field **in plaintext** — it has to, so the server knows which cert to present. GFW reads this in real time and RSTs the connection if the SNI matches a blocklisted domain. This is why SNI matters independently of the IP — same server, same IP, but a blocked SNI string kills it.

4. **Deep packet inspection / protocol fingerprinting (this is the layer people usually miss)** — even with a permitted SNI or no SNI at all, GFW runs statistical analysis on the traffic itself: packet size distributions, timing, TLS ClientHello fingerprint (JA3), entropy of payload. Real TLS to a real website has a recognizable "shape." A VPN protocol, Shadowsocks, or a tunnel disguised as TLS but not *quite* matching a real browser's fingerprint can get flagged as "probably a proxy" even without knowing what's inside it.

5. **Active probing** — this is the layer that answers your specific question. GFW doesn't just passively watch — for servers it's suspicious of (e.g., an IP that just carried a flagged-as-tunnel flow), it will **actively connect back to that same server itself**, minutes or hours later, and try speaking known proxy protocols to it (Shadowsocks handshake, VPN handshake, etc.) to confirm what it's running. If confirmed, that IP goes on the blacklist retroactively — even though nothing about *your* traffic pattern changed.

**So to directly answer your question** — yes, a tunnel that's up and working can absolutely get killed *later*, and this happens for a few distinct reasons, not just one:
- The flow itself got flagged mid-session by statistical/DPI analysis (long sustained flows are more sample-able than a single quick ping — this matches what you saw with `tailscale ping` succeeding while `ssh` timed out).
- Active probing confirmed the server as a proxy sometime after your first connection, and now the IP itself is burned for everyone, including you.
- Rolling/time-boxed blacklist updates — GFW blocklists update in batches, not instantly.

This is exactly why real circumvention tools (Geph, obfs4/Tor, VMess+TLS, etc.) don't just encrypt — they specifically try to make the *traffic shape itself* statistically indistinguishable from ordinary HTTPS to a popular site (mimicking real browser TLS fingerprints via uTLS, padding packet sizes, etc.), because layer 4 (DPI/fingerprinting) and layer 5 (active probing) are the ones that keep hunting *after* the connection is already established — encryption alone only defeats layer 3 (keyword/content filtering), which mostly stopped mattering once HTTPS became universal anyway.

# offline game

 And these are in the domain of networkings,  GFW is not limited to tecs, they could send out secret polices to catch client side users.  This also a main part of the game.

That's an important and often-overlooked layer — you're right that "GFW" as commonly used is shorthand for the entire censorship apparatus, not just the network appliances. The non-technical enforcement layer includes things like:

- Legal/regulatory pressure — unauthorized VPN sales and use carry legal penalties in China; there have been periodic crackdown campaigns specifically targeting people who sell or operate circumvention services, with real prosecutions, not just technical blocking.
- Real-name registration — SIM cards, ISP accounts, and increasingly cloud/VPS purchases tied to real identity, meaning even a technically unblockable tunnel can potentially be traced back to whoever paid for or registered the endpoint.
- ISP-level cooperation — ISPs aren't just conduits the GFW filters through, they're compelled participants — traffic logs, subscriber records, and account-level flags can all feed into follow-up (a knock on the door, not a packet drop).
- Social/informant pressure — campaigns encouraging reporting of "illegal" internet use, workplace/school monitoring in some contexts.

This matters technically too, not just legally: it's why serious circumvention tool design (Geph is actually a good example here) cares about things beyond pure DPI-evasion — anonymous payment options, not requiring account registration tied to ideromised server could reveal about who connectfectly invisible to DPI but tied to your real nay solved the threat model — it's solved half of itouse we discussed andthis enforcement layer ames running in parallel,
and tools that only play.