Parity

▶️ Try it live in the Playground — write and run .nim / .aowl in your browser, no install.

Every stage of this toolchain is written from scratch and then held to one standard: run the same input through the reference tool and through ours, and the two outputs have to be identical down to the byte. Same parse tree, same typed IR, same generated code.

That bar is unforgiving on purpose. Every naming decision, emission order, whitespace convention and line-info offset has to match, and at the end you are looking at a diff that is either empty or it isn't — there is no partial credit and no room to talk your way past it. Clearing it is what makes a stage an actual drop-in replacement instead of something that looks like one.

We have not cleared it everywhere. This page is the scoreboard, including the parts that aren't green.

Where each stage stands

Stage	ours	vs.	from scratch	byte parity
parse	aowlparser	nifler	✅	byte-exact on the whole compiler tree (184/184); 91–283 files byte-exact on the stdlib, 100% structural everywhere
semcheck	aowlsem	nimsem	✅	corpus 924/941 byte-exact (15 of the 17 misses are oracle-side); own dependency closure 31 of 56 modules byte-exact; std/system semchecks with 89 differing tokens over ~92,600 lines — all Windows, 2026-09-09, aowlsem commit d4955d3e. Linux moddiff baseline: 47 of 55 (2026-08-19). Not comparable across platforms.
lower	aowlhexer	hexer	⏳	runs the reference's 25 passes with two of our own fixes on top (below), so it is near-identical rather than identical by construction; the from-scratch rewrite is next
C codegen	aowlc	lengc	✅	end-to-end correct today (runs, ASan-clean); text byte-parity with lengc is the active push
interpret / VM	the Nimony interpreter in engine (was aowli)	(new)	✅	a tree-walker and a bytecode VM over typed .s.nif; 460 / 460 runnable corpus programs byte-identical to native compile-and-run, and the cross-check corpus at 461 agree / 0 diverge on the shared-GC runtime (2026-10-08)
emit → TS / Py / JS / WASM	aowlts · aowlpy · aowljs · aowlweb	(nimony backends)	✅	idiomatic, readable output; behaviour-verified against native, run-for-run

✅ = written from scratch · ⏳ = still reuses the reference implementation

Parser results

The front of the pipeline is done. aowlparser is diffed against native nifler: structural means the token trees match with line info stripped, byte-exact means the .p.nif files are identical including every offset.

corpus	files	structural	byte-exact
nimony/src (the compiler itself)	184	184	184
nimony/lib (stdlib)	105	105	91
upstream Nim/lib	310	310	283
curated	172	172	156

0 crashes, 0 hangs across all four. The byte gaps that remain are written down one by one rather than left as a mystery — see Parity & gaps.

Semantic checker

aowlsem is a clean-room replacement for nimsem, and it is the stage where parity is currently being ground out construct by construct. The dated gate verdicts live on the aowlsem page; on Windows on 2026-09-09 they read corpus 924/941, moddiff 31 exact of 56, sysdiff 89 differing tokens over all of std/system.

Correction, 2026-09-09. This page said "498/498 modules byte-exact, including all of std/system" from 2026-07-28. That number was an hconv-site census from aowlsem's REQUIREMENTS.md, not a module count, and std/system was not — and is not yet — byte-exact. The aowlsem page has the full note.

What's left is not a long tail of near-misses. Differential runs over the whole nimcache oracle set point at specific missing features — anonymous sum-type construction and of-pattern matching are the two biggest — rather than at drift inside code that already works. That's the good kind of remaining work: you can name it, and finishing one construct moves a countable number of modules. Details on the aowlsem page.

Lowering

aowlhexer still runs the reference compiler's 25 lowering passes, so its .c.nif matches by construction almost everywhere. There are two deliberate exceptions, both in places the reference marks its own code as unfinished:

• Captured var / out parameters are now rejected. A closure that captures a var T parameter holds a pointer into the caller's frame, so an environment that outlives the call is left pointing at dead stack. The reference leaves this as a TODO in lambdalifting; we make it a lowering-time error.
• Move analysis reports the right token. In mover, the "other usage" that blocks a sink was recorded after the cursor had already advanced past it, so the diagnostic pointed at a closing paren. It now points at the use.

Both make aowlhexer stricter or more accurate than the original rather than looser, so either one can reject a program the reference accepts. That is the trade we want. The passes get rewritten onto an aowl-owned core next, and then this row stops saying ⏳.

How we measure

Every claim on this page comes out of a differential harness: run the same input through the reference tool and ours, normalize only what genuinely doesn't matter, diff the bytes. A green diff is the only thing that counts as done for a construct — everything else is a punch-list item, not a rounding error. That is how the parser got where it is, and it is how each remaining stage closes.

If a number here looks wrong to you, we would rather hear it than not. The harnesses are in the repos, and the argument is welcome on Discord.