The morning began at 95,669 pass, 11 fail, 0 skip on a hosted run whose notes say, in so many words, "This is not a 100% conformance result." An Annex B fix (9f9e3d56) found a real hidden defect behind an obsolete expectation: a block function was not visible before its textual declaration. Then 1539eb4b completed the pinned Intl.DateTimeFormat shard with generated CLDR 48 data and reported the corrected core profile separately from the unmodified one.

The nine remaining unmodified failures are contradictions inside the pinned corpus, corrected by five patches with recorded hashes; Node 24.19 fails four of the same executions. v0.0.18 (fc474d15, the repository's only pull-request merge) shipped with the first WebAssembly archive that includes Python. The landing page was updated to show both numbers.

Until today zipp.org was a single React page whose text lived inside a JavaScript bundle. It now has static HTML pages for the JavaScript engine, Python, WebAssembly, GPU compute, Test262, benchmarks, architecture, embedding, sandboxing, comparisons, releases, the story and three long-form articles, each with its own title, description, canonical URL, breadcrumbs and structured data.

Every word of those pages is in one file, landing/site/content.json, and a small renderer turns it into HTML at build time, so the text is in the served document before any script runs. The same build writes sitemap.xml, robots.txt, a real 404.html, the journal's RSS feed and a search index. Unknown URLs now return a genuine 404 instead of a copy of the home page.

378b18de adds the Python frontend, the javascript and all WebAssembly variants and the playground. 5fd22bcb is titled "Python: Zipp's own full-language implementation": the RustPython parser for the AST, Zipp's own symbol table and emitter to register bytecode, and a JavaScript-shaped runtime for the object model, validated byte for byte against CPython on a committed corpus.

8bb14918 integrates the GPU Lab so Python graphs can execute on WebGPU, WebGL2, compiled WASM kernels or JavaScript. 9651dd40 makes Python projects folders, adds the Torch subset and puts the lab in the playground. 955914f3 adds asynchronous dense GPU training through torch.compile(step, training=True) and pins down the virtual-filesystem change protocol. The docs are careful: a Python and Torch subset, not CPython or PyTorch.

The correctness audit starting from ee107d4b (v0.0.16) found host values and exceptions lost across a callable Proxy's apply getter, a structured argument returned as Opaque and a string as an unrelated Date, TypedArray species and byte-copy semantics, and six reproduced defect regressions. The follow-up covered ArrayBuffer, SharedArrayBuffer and DataView re-entrancy, a CAS retry in SharedArrayBuffer.prototype.grow, strict ToNumber versus BigInt across a dozen algorithms, Promise combinators and iterator closing.

v0.0.17 (127477bd) is "audited engine hardening". Both audit reports are in docs/audits/, and the scorer fix that came with them exposed 42 false passes in the older corpus (95,897 of 95,942 before the fixes, 95,939 after).

014c1b70 and cfb83cf3 fix the 11 September audit's embedding-API and host-boundary findings. b92a268c adds browser Worker smoke tests and a reference host adapter tested in real Chromium, Firefox and WebKit Workers. ab56a840 and 14770703 rework CI: exact-revision release gating, an accountable quarantine, scheduled security checks, and a minimum gate on push with the heavy suites as local commands and a weekly security run.

The canonical capture at 14770703 (zipp 0.0.15) is publishable and worse: headline ten rows at 1.003× Node against 0.878× on 2 September, with bytecode-vm at 3.59×. A paired A/B against the audited baseline puts the baseline within 1% of the new build on every row, so the regression opened earlier, and the suspect is v0.0.15's switch to specification-order method calls by default. The ledger records the trade instead of hiding the capture.

The first run on the newly pinned Test262 corpus also landed today: 95,665 of 95,680, then 95,669 after two JIT-only defects were fixed the same day.

v0.0.15 (e6e0f65d) bundles tickets B280 through B288: spec-order method calls by default, an accel spec grammar, transactional host.call, zippProfile() so a host can tell which build it loaded, and a CI gate. B280 is the one with a price: ZIPP_RELAXED_CALL_ORDER=1 still exists as a diagnostic and is documented as observably wrong for a getter or proxy trap, and it will never ship. The performance cost showed up in the 11 September capture.

Tickets B274 to B278, measured as diagnostic A/Bs on one machine: a 64K non-ASCII charCodeAt loop 4,462 → 3.2 ms; a tokenizer 9,828 → 8.2 ms; a slice loop 449 → 4.3 ms; join with 300K retained strings 8,423 → 138 ms; a new Function property loop 20.9 → 14.8 ms; a named-property array read 13.3 → 9.6 ms. None of these are headline numbers; all of them were cliffs a real embedder could fall off.

v0.0.11 (9409b8d5): register classes keep tokenizer loops on the INT tier, taking parse-large-js from 1.24× to 0.88× Node, and a forced-JIT segfault on a user-overridden Math.random is fixed. v0.0.12 (13aa7e63): inline stores into pinned young dense arrays and x | 0 fused into the wrapping add. b36702a4 publishes the 8229b3fc capture: headline ten 0.878×, all 13 normal rows 0.614×, hostile 17 0.829×, all 30 equal-row 0.728× Node, 0.594× Bun, 0.460× Deno, with 15 counterbalanced repetitions and 10,000 bootstrap samples.

Nine rows still lose to Node, led by reactish-reconcile at 1.578×, allocation-survival at 1.559× and warm-router at 1.520×. The artifacts print FASTER_THAN_NODE_ON_EVERY_ROW=0.

The tag messages read like a bug tracker for a product that had just met its users. v0.0.7 (d69c7b7b) raised the hardened ArrayBuffer ceiling from 1 MB to 32 MiB: "an 8 MB Game Boy cartridge and a single 8.5 s frame of 48 kHz audio were both impossible." v0.0.8 (32680f06) added renewInstructionBudget, because a 50-million-instruction lifetime budget "was a fuse on every long-running embedder." v0.0.9 (d0c783b1) schedules collections on bytes, not only allocation count: "An emulator encoding a 23 KB framebuffer per frame died ninety seconds into a game." v0.0.10 (ddbf7cd9) sized the instance ceiling at 1 GB and the heap budget at 512 MB.

The day after the first release, the first embedders' workloads exposed the cost of crossing into the VM. v0.0.2 (a70cef47) stopped the host boundary from walking the whole heap on every re-entry and raised parser sandbox limits that had rejected real code. v0.0.3 (d71168a9) removed another O(heap) walk, took JSON.stringify ×4000 from 919 ms to 8.5 ms, shrank the bundle from 6,018,862 to 5,668,533 raw bytes, dropped wasm-opt because -Oz measured smaller raw but larger on the wire and slower, and pre-warmed the playground Worker after finding that a 68 ms delay was V8 compiling, not the engine. v0.0.4 (769c68ce) added getGlobalsFingerprint, a content digest so a mutation through a reference cannot be missed, cutting per-tick host sync from 9.6 ms to 2.1 ms.

96dac4e1 is tagged v0.0.1. be2f5ed4 (B249) records the retained-ten suite at 0.918× Node and hostile at 0.866×, with bytecode-vm from 3.68× to 0.93× and warm-router from 2.18× to 1.62×. 8ac5cc73 computes module [[CycleRoot]] at depth-first-search time and restores three Test262 rows, back to 95,939 of 95,942. ce488bbe archives the campaign ledger (B001 to B252, mostly negative results, with a header warning that its earliest sections were written at 3.31× and have since been refuted), and ac183cff redesigns the README around measurements.

The capture on 29 August came in at 2.157× Node on the headline ten, from 0.965× the day before. The cause was 85fca98a, which de-fused the CallMethod lowering; e2250fca separately traced a 24-row Test262 drift to the same change. 271647ad reversed it (0.977× headline, 0.665× all 13). The same call-order question came back in September as B280, this time resolved in favour of correctness and with the cost measured.

The hostile corpus pairs six baselines with six stressors (closures, megamorphic shapes, type churn, throw-catch, surviving allocation, long-lived async) and adds five applications including a React-style reconciler, a warm router, a bytecode VM, a hot module graph and the unmodified nanoid package. Its first captures on 26 August were 1.111× and 1.029×; af7289f1 records 0.961×, "UNDER PARITY (first in series)", and the next two days reach 0.937× and 0.920×.

e89612a6 (24 August) hardens the sandboxes and adds the guarded ARM64 baseline JIT: bounded call-free integer functions and numeric loops with exact-ip fallback, no helpers, no OSR yet, and a bespoke W^X path on Windows ARM64. 4195e063 continues the hardening for untrusted code: compile-time forbid(unsafe_code) under safe-sandbox, a compile error if safe-sandbox and jit are combined, and separate workspaces so feature unification cannot reintroduce a JIT.

0f1a4c71 (23 August) is "wave 19 -- the three rows nobody had decomposed were 60% of the gap", at 1.0997×. c6ffa200, the next day, is "bench: cross Node parity at 0.9695x". Fifty-six days earlier the same suite had been at 4.20×.

The engine had already learned this lesson once in June, when an adversarial audit of the JIT found six soundness bugs, and again on 13 June when "the INT tier was silently returning wrong answers on ordinary loops -- a dense-array tag check scratched a live Bool home register". ee14c71d makes the check systematic: a fuzzer that runs generated programs in every tier and diffs the results, which found five more classes. A later commit found that "the operand table was blind to 185 of 221 opcodes -- 23 wrong-answer shapes, one root cause".

cdf064c4: "the wave-13 capture on the PGO build -- 1.2511x -> 1.1962x [1.192, 1.201], the first capture under 1.20x; typedarray-math reaches node parity". The double tier's unboxed register homes are what carried it.

The log had been quiet since 4 August. 74d03d23 is "the nursery goes default-on": minor collections trace young reachability plus remembered old-to-young edges; promotion is a bookkeeping change on the same heap index because the collector never moves objects. The wave-9 capture moved from 1.2832× to 1.2117×. ZIPP_NO_NURSERY=1 keeps the majors-only mode for conformance runs and ablations.

NURSERY_DESIGN.md was written when GC looked dominant; the ledger later corrected that ("GC is 2-12%, not dominant — B81 corrected" and "it is the COLLECTOR, not the allocator — pool built, measured zero, reverted"). Stage 1 of the nursery was refuted; stage 3 (82e454fa, 72cb640f) landed with a prover that verifies remembered-set invariants, ZIPP_NURSERY_VERIFY=1. The design document now carries a header calling itself historical. 0d124bf1, "feat: add Zipp landing page", is the first version of this site.

5cafcb10. Every capture since has been a PGO build with the profile's hash, the recipe, the compiler, the linker and the source snapshot all recorded in the artifact; the working rule is "PGO retrain before any headline capture" and "never promote a dirty, filtered, non-PGO, or incomplete-engine artifact".

f1b38086: "bench: the first capture this repo can attribute to a commit". The sampling profiler the roadmap had wanted since its third ticket landed the day before. The ledger's measurement-hygiene section records the failures that motivated this: a profiler "lying about a quarter of json-large", a harness that "could name a commit it had never measured", and a leftover ZIPP_GC_STRESS=1 process from a killed session that held a core for hours and was misdiagnosed as an engine livelock.

94e7e0f5 moves the README to the new home. d5d0ae7b (29 July) refreshed the corpus and implemented Iterator.prototype.join, 36 executions that V8 24.12 does not pass either: "the suite is the oracle, not node". 799ead6d recorded the last full-suite Intl402 figure on record, 6,502 of 6,714.

a3d616d3 reaches 99.97% with 30 failures and claims Annex B handling "more conformant than Node", a claim later retracted. d0854919: "test262: 30 -> 5, and none of the 5 is a live engine defect". Of the five, module-level using declarations that were never disposed accounted for six executions fixed the same day, and the mutually exclusive Annex B pair ("95,942/95,942 does not exist for any engine — V8 fails the same one") became the standing caveat.

Four commits in one day: b550a4ca adds Intl.DurationFormat and the IANA database generated from pinned upstream data; 04edd833 adds non-ISO Temporal calendars (Intl402 63.6% → 85.3%); 1ae58724 reaches 99.88% core and 96.9% Intl402; f1997ad8 reaches 99.91%. Temporal ended up with fifteen calendars: twelve closed-form, Chinese and Dangi by astronomical calculation, and Umm al-Qura from the required month data.

6fc39171 (98.1%) is the new front end built for correct early errors. 321c328b reaches 99.0% (95,091 of 96,029). 8cec7647 then honours YAML list-form flags: metadata, which removes 181 phantom executions and changes the denominator to 95,848; the number was republished as 99.2% rather than left inflated.

f6f4a804 adds zipp-wasm; 768287f8 adds the persistent-VM embedding API, "compile once, re-enter later", which is the model the browser Engine class still follows. 694886b8 turns on fat LTO and codegen-units = 1 (2.80× → 2.77×, "free"). 44c4c738 validates regular-expression literals at compile time, the single largest conformance jump of the week at +712 executions.

1217d683: "Fix the test262 runner; real pass rate is 96.97%, not 100%". The runner had reported 48,550 passes of 48,556 single-mode executions; the corpus required 96,029. 9ef187f9: "Focus the workspace on zipp-vm; rewrite the docs against measurements", deleting the AOT, Cranelift, LLVM, zero-knowledge and TypeScript crates from May. b865f9c6 adopts mimalloc as the global allocator, 4.20× → 3.34× Node on the honest suite. The roadmap's B10 entry the same day is "the measured backlog from the 25-agent speed hunt".

4e12a406 and f0d725a8 enable the Tier C function-level JIT by default. 62dbba15 adds PERF_ROADMAP.md, whose tickets (B1, B2, B3…) number every later performance commit. The day before, 7e29d967 had fixed six JIT soundness bugs found by an adversarial audit of the session's own codegen. June ends with 1,039 commits.

c43b17ce: WTF-8 so that a lone surrogate survives a round trip, as ECMAScript requires. 7f4e58b6: BigInt with an i128 fast tier backed by num-bigint. f00c2010: "Add bench/real: ten-workload five-engine suite (the honest pack)", the ancestor of today's real13 suite. The day before, b2c04ace had given Atomics a real blocking wait with a global FIFO waiter registry and cross-thread notify, and 247238ca had added proper tail calls.

5fba4d7d: "Run ES modules: top-level-await module mode + test262 module runner". The module graph, [[CycleRoot]], deferred evaluation ordering and errored-cycle re-import would each get their own fix over the following three months; today's zipp mjs is the descendant of this commit.

76169428: "implement eval (indirect / global-scope) — the #1 conformance lever". 97e07a67 adds ShadowRealm. Both are cheap in an explicit-frame VM: a new realm is a new set of intrinsics and a new global, not a new native stack.

The busiest day in the log. 8520b6a8 wires regular expressions through the regress crate (later forked for UTF-16 execution and a regex JIT). 65e33aed: ArrayBuffer and eleven TypedArray constructors. 48633f27: Proxy traps. 4a3f85b6 through 00df2adc: Temporal, from Duration to PlainDate to Instant to Now. ebf7dca3: "implement the Intl namespace + all 11 services (0 -> ~340 tests)". a9b1d901: "split the 17.4k-line vm.rs into a vm/ module folder". b5d691e1: a mark-sweep collector for the new VM.

79d10480: "tools: add test262 conformance runner". dd820da7: "async generators (async function*) — the biggest conformance lever". From this day on, features were measured in executions passed, though it would take until 25 July for the runner to count them correctly.

In one day: 59032ebd "native x86-64 JIT for hot integer functions (4.1x where it fires)"; b882eea5 "OSR loop-region JIT (double/SSE2)"; 050fe88d "call-free inline cache for GetProp/SetProp — object.js 14ms → 3ms"; bca59c77 scalar replacement, "object 31ms → 8ms"; e6e63a0a "rope strings + cached length/index — kill the O(n²) in string concat & scan"; 0db0d5af "Remove the old JS engines; zipp js now runs zipp-vm"; a5cf5a3e promises and a microtask event loop; and d7d9d5e0, "cheap 2^53 guard (range trick) — BEAT V8 on loop.js (26.5 vs 28.0ms)".

262ae969: "prototypes, new, instanceof, in (the ES5 object model)". e2440724 wires zipp js <file.js>. A fuzzer campaign follows, fixing Number::toString and bugs numbered 1 through 32 one commit at a time. ecc96a94 is the first bytecode compiler and VM; 4c826c54 absorbs an earlier engine and rebrands it; 50634ccd is "FIX broken build + RETRACT the false perf claims in 4c826c5". Then 87174f64: "zipp-vm: clean-sheet JS engine v2 — explicit-frame register VM", the design every later page on this site describes.

a71a1bdc, "Initial commit - ZIPP language v0", 13:38 on 29 May 2026. d0fe82fa: "Add native JIT (Cranelift) for the integer subset — beats V8 on a hot loop". c30bb934: "Add LLVM release tier (clang -O3) — matches V8 on dense f64". 924ed4a1: a conservative mark-sweep GC. fe045b2b: "the WASM contract profile: deterministic, gas-metered WebAssembly". b1847a67: "zk hardening: bind proofs to a program hash". 5d1c1fab: a TypeScript front end via oxc. All of it was deleted on 25 July; the habit of naming the measurement in the commit subject survived.