---
title: "Why raddr"
output: rmarkdown::html_vignette
vignette: >
%\VignetteIndexEntry{Why raddr}
%\VignetteEngine{knitr::rmarkdown}
%\VignetteEncoding{UTF-8}
---
```{r, include = FALSE}
knitr::opts_chunk$set(
collapse = TRUE,
comment = "#>"
)
```
This article is the case for the package: the disagreement it reports, where
that disagreement was measured, and what raddr deliberately leaves to other
tools. For a hands-on tour, start with `vignette("introduction", package =
"raddr")`.
## The problem, in one string
One literal, one machine, three different hosts:
```
"0177.0.0.1"
strict (RFC dotted-quad; Python ipaddress, Go, Rust) -> reject
whatwg (= what browsers do) -> 127.0.0.1
pton (POSIX inet_pton, Apple libc) -> 177.0.0.1
aton (BSD inet_aton) -> 127.0.0.1
```
Every existing library picks one of those readings and discards the rest. raddr
does not pick. `addr_parse()` takes no mode argument — it returns every reading,
always, with the reason codes that explain each one.
### Dialects sit on two axes
Not one list, but what a *standard* requires ("on paper") versus what an
*implementation* actually does ("in reality"):
| Axis | Dialect | Models | Stability |
|---|---|---|---|
| paper | `strict` | RFC dotted-quad grammar; Python `ipaddress`, Go, Rust | fixed |
| paper | `whatwg` | WHATWG URL host parser; what browsers do | fixed, versioned spec |
| reality | `pton` | POSIX `inet_pton` | **platform-varying** |
| reality | `aton` | BSD `inet_aton` | stable in practice |
Two more dialects are precedence orderings over those reality primitives, not
separate parsers:
```
addr_getaddrinfo = pton, falling back to aton
addr_curl = aton, falling back to addr_getaddrinfo
```
curl falls back to the resolver *entry point* rather than to the bare parser
under it, because that is what real curl reaches — measured against curl 8.20.0,
after the derived version of this line turned out to be wrong for IPv6.
### Measured divergence
Measured on macOS Darwin 25.4.0 arm64, Apple libc, 2026-07-26; the `curl`
column against real curl 8.20.0, 2026-07-28:
| input | `strict` | `whatwg` | `pton` | `aton` | `getaddrinfo` | `curl` |
|---|---|---|---|---|---|---|
| `127.0.0.1` | 127.0.0.1 | 127.0.0.1 | 127.0.0.1 | 127.0.0.1 | 127.0.0.1 | 127.0.0.1 |
| `0177.0.0.1` | reject | 127.0.0.1 | 177.0.0.1 | 127.0.0.1 | 177.0.0.1 | 127.0.0.1 |
| `192.0.010.1` | reject | 192.0.8.1 | 192.0.10.1 | 192.0.8.1 | 192.0.10.1 | 192.0.8.1 |
| `192.0.048.1` | reject | reject | 192.0.48.1 | reject | 192.0.48.1 | 192.0.48.1 |
| `4294967296` | reject | reject | reject | 0.0.0.0 | 0.0.0.0 | 0.0.0.0 |
| `1.2.3.` | reject | 1.2.0.3 | reject | reject | reject | reject |
| `2130706433` | reject | 127.0.0.1 | reject | 127.0.0.1 | 127.0.0.1 | 127.0.0.1 |
| `10.048.1.1` | reject | reject | 10.48.1.1 | reject | 10.48.1.1 | 10.48.1.1 |
Two rows carry most of the package's value:
- **`192.0.048.1`** — curl reaches a host a browser refuses to dial.
- **`4294967296`** — `aton` wraps modulo 2^32 to `0.0.0.0`; the standards reject.
### For IPv6 the disagreement inverts
The two paper dialects disagree about IPv4 on six of the eight rows above. About
IPv6 they agree on every input measured, and all of the divergence moves to the
reality side:
| input | `strict` | `whatwg` | `pton` | `aton` | `getaddrinfo` | `curl` |
|---|---|---|---|---|---|---|
| `::1` | ::1 | ::1 | ::1 | reject | ::1 | ::1 |
| `00001::` | reject | reject | 1:: | reject | 1:: | 1:: |
| `::1.2.3.04` | reject | reject | ::102:304 | reject | ::102:304 | ::102:304 |
| `fe80:abcd::1` | fe80:abcd::1 | fe80:abcd::1 | fe80:abcd::1 | reject | **fe80::1 %43981** | **fe80::1 %43981** |
| `fe80::1%lo0` | reject | reject | ::1 %lo0 | reject | ::1 %lo0 | ::1 %lo0 |
- **`fe80:abcd::1`** — two libc entry points on one machine return different
bits for one string. Apple's `getaddrinfo` reads the second hextet of a
link-local address as a scope ID and clears it; `inet_pton` does not. This is
the IPv6 counterpart of `0177.0.0.1`. curl reaches the entry point that lifts,
so `pton` is the odd one out on this row.
- **The `curl` column is a copy of the `getaddrinfo` one here**, because `aton`
rejects every IPv6 literal and the composition is nothing but its fallback.
For IPv4 the two differ on `0177.0.0.1` and `192.0.010.1` above.
- **`00001::`** — `inet_pton` counts only the four *significant* hex digits and
lets the leading zeros run as wide as they like. The standards cap the digits
outright.
- **The zone ID** lives in its own field, never in the address bits, and does
not participate in equality: `fe80::1%lo0 == fe80::1%en0` is `TRUE`, and
`addr_zone()` is how you tell them apart.
## raddr is a shower, not a protector
It reports facts. It never returns a verdict, a risk score, or an allow/deny
decision. Security policy is a separate concern and belongs to a separate
package.
### What raddr will never do
| Excluded | Owner |
|---|---|
| DNS resolution, hostname lookup | `ssrfr` / `curl` |
| HTTP, redirects, connection pinning | `ssrfr` |
| Allow/deny policy, risk scores, verdicts | `ssrfr` |
| Cloud-metadata endpoint tables | `ssrfr` |
| "Most restrictive reading wins" convenience | nobody: not needed, `ssrfr` fixes one reading (INV-1) |
| Geolocation, ASN, country data | nowhere |
| IDNA, punycode | `punycoder` |
| Public-suffix logic | `pslr` |
| URL parsing, scheme/port policy, reg-name-vs-IP host form | `rurl` |
| General CIDR set algebra (collapse, exclude, subnets) | `ipaddress` |
| `X-Forwarded-For` extraction (HTTP header parsing, not address parsing) | the application (no stack package) |
| Visualization | `ggip` |
### Why not just use `ipaddress`?
`ipaddress` is a good package and raddr does not replace it — the table above
assigns CIDR set algebra to it permanently. But it answers a different question.
It gives you *one* reading of a literal and is silent about the rest: it
transforms `0177.0.0.1` into `177.0.0.1` without a word, which is exactly the
transformation that changes which host you reach. raddr's premise is that
`0177.0.0.1` has several defensible answers and that you should see all of them.
raddr is pure R, performs no network access, and depends on `vctrs` and `rlang`.