The instrument
WebAssembly build: loading…
Pick a format and an operation, type operands as bit patterns or decimal values, and run. The hex bits are what actually execute; the decimal fields are a display convenience that converts through the format. Or deal a real line out of the repository's committed test vectors and watch the build reproduce it, or deal a curated edge case to see a corner where exact semantics show, computed live.
Result
bits
The operation list is the repository's generated dispatch table
(tests/msf_dispatch.h) — the same table the host test runner drives. The twelve
character-conversion operations and six flag-management operations are not hex-drivable and
are omitted here. The nine Clause 9.5 augmented operations
(augmentedAddition/Subtraction/Multiplication at the
three formats) return a head+tail pair under clause-fixed roundTiesTowardZero; this
instrument's single-result shape cannot hold two results honestly, so they are omitted here
too. All omitted operations are gated on the host (and, for the pairs, replayed on the
board) like everything else. Decimal entry for
binary16 rounds ties-to-even in the page before the bits run.
The numbers
Generated at build time by site/build_data.py from the files named in the
right-hand column — nothing here is typed by hand. Rebuild with site/build.py
and the table follows the tree.
| Measure | Value | Source |
|---|---|---|
| Loading… | ||
Planted-copy verification, by wave
Before each parcel of work merged, an independent verifier re-derived its claims against a copy of the tree carrying deliberate, pre-ledgered defects (“plants”). Caught plants are evidence the verifier was actually looking.
| Wave | Caught | Plants | |
|---|---|---|---|
| Loading… | |||
On silicon
Three workloads run bit-identically on RP2040 silicon through this library (the genuine Pico; IAS15 also across five RP2040-Zero clones — six chips in all). The figures below are recorded facts from docs/RECORD.md, with the board protocol in docs/BOARD.md — quoted with their units and context, not as speed claims.
IAS15 — REBOUND's integrator on the Pico
- What runs
- REBOUND's IAS15 gravitational integrator (GPL, scratch only — it never enters this repository), its binary64 arithmetic served by this library over the AEABI compat surface.
- Agreement
- 2,001 of 2,001 per-step state hashes match the unmodified desktop build exactly
(step-2000 state hash begins
4d4530a6) — on each of six chips: the genuine Pico and five RP2040-Zero clones, zero failures in the sweep. - Recorded rate
- 1.60 accepted steps per second at 125 MHz on every chip in the six-board sweep (2,000-step timed run). A full figure-8 period takes about ten minutes at that rate. The firmware blinks the onboard LED while working (Pico GP25; Zero WS2812 on GP16, outside every timed region); that LED build re-passed 968 of 968 harness vectors on all six boards, so it is hash-verified on silicon.
The sketch is an illustration of the figure-8 three-body choreography, drawn by this page. It is not simulation output and carries no data.
Interval geometry
- What runs
- Guaranteed ray classification (sphere, box, torus) over directed-rounded intervals, against a frozen 710-row exact-oracle catalogue.
- Agreement
- The full catalogue streamed to the board matches the host bit-for-bit: 710 of 710 rows,
stream hash beginning
ce81d66eon both sides; 7.92 s of workload-call time at 125 MHz.
KissFFT with runtime twiddles
- What runs
- KissFFT 131.2.0 (BSD-3, scratch only) with one named change: its setup phases hand exact
cycle fractions to this library's
sinPi/cosPiinstead of calling host trigonometry. It is never described as unmodified KissFFT. - Agreement
- All 12 configuration hashes bit-identical between host and board at the gated sizes N = 256/1024/4096 (complex and real, forward and inverse).
How this was verified
- Named authorities, never self-consistency. Correctness is decided
against Berkeley SoftFloat/TestFloat 3e (pinned by archive SHA-256 in
docs/metrics/fetches.jsonl, executed in scratch, never vendored), MPFR, or an exact-integer reference — per instantiation, in docs/COVERAGE.tsv. The plans of record are the SF-0/SF-1/SF-2 studies in docs/studies/. - Planted copies. Independent verifiers re-derive each parcel's evidence against a tree carrying pre-ledgered defects; the score to date is in the table above, misses included. A gate that has not been watched to fail is not treated as a gate.
- Board hashes. Silicon results name their firmware hash, commit and board, or they are not evidence; the three workloads above carry theirs into RECORD.md.
- This page. The WebAssembly build is proven at build time by replaying
the bundled sample of committed vectors through it under Node
(
site/replay.mjs, log insite/replay-log.txt); the count is in the table above. If WebAssembly is unavailable in your browser, the instrument says so and dealt vectors show their recorded results, labelled as recorded.
Status at SF-2 close
Tier 1 (the full Clause 5 surface) and Tier 2 are complete and gated
(counts in the table above, generated from docs/COVERAGE.tsv). That includes
roundTiesToAway as the fifth rounding attribute; correctly rounded
rSqrt/hypot, exp2/exp2m1 and
log2/log2p1 at all three formats; and Clause 9.5 augmented
add/subtract/multiply (head+tail pairs under clause-fixed rounding — excluded from the
instrument above by shape, gated on host and board). Nothing beyond the SF-2 study's list
is claimed; the generated counts are stamped as of the commit in the footer.