Skip to content

Sley 2.1.8

Choose a tag to compare

View all tags
@GreyforgeLabs GreyforgeLabs released this 05 Nov 04:12
· 79 commits to main since this release
Immutable release. Only release title and notes can be modified.
v2.1.8
abce1d4

Sley 2.1.8 is a point release of work that is already built and measured. Python // and % on values the compiler can prove non-negative now lower to plain divide and remainder (#73), refusals name the fix (#75), help python-task is a shorter card for tasks whose tests come from elsewhere (#82, #83), and the repository carries a model-cost benchmark (#74, #76–#84). With process start included, Sley 2.1.8 is faster than CPython 3.13 on all eleven benchmark programs in one session (0.37–0.92x of CPython's time); digitsum is 2.83x faster than in 2.1.7. On Claude Haiku 5.5, Sley costs about 1.25x Python per trial (95% CI 1.10–1.45); it is not cheaper than Python yet. Fuel and instruction counts match call --reference.

Release notes · Benchmarks · Model cost · Quickstart

Cut from 4da66c0 on release/2.1.8. Work merged to main after that (lists, list results, native Text and str operations, bisect, task set v2) belongs to 2.2 and is not in this release.

Benchmarks (wall ms, hyperfine median [min-max], release-archive musl binaries, one session, 3 rotated rounds)

benchmark Sley 2.1.7 Sley 2.1.8 2.1.7 / 2.1.8 CPython 3.13.5 2.1.8 / CPython VM instructions 2.1.7 → 2.1.8
fib 30 78.3 [74.6-117.5] 77.0 [75.3-170.1] 1.02x 84.0 [80.8-173.2] 0.92x 14,808,950 → 14,808,950
loopsum 1e6 34.9 [34.0-83.3] 31.1 [29.6-48.0] 1.12x 58.8 [56.8-138.3] 0.53x 11,000,003 → 9,000,003
collatz 5000 27.5 [26.8-32.9] 14.4 [14.0-81.8] 1.90x 35.9 [34.9-58.6] 0.40x 7,019,817 → 4,171,757
primes 30000 28.5 [27.5-93.9] 15.5 [15.0-28.7] 1.83x 41.5 [40.6-47.1] 0.37x 7,995,140 → 3,803,716
gcdsum 400 26.0 [25.3-41.1] 26.1 [25.3-44.1] 0.99x 42.6 [41.7-93.1] 0.61x 4,361,274 → 4,361,274
tak 24 16 8 70.2 [68.0-95.8] 70.3 [68.3-111.7] 1.00x 80.3 [78.6-134.6] 0.88x 8,726,719 → 8,726,719
triples 120 12.4 [11.8-32.3] 12.9 [12.1-38.0] 0.96x 26.1 [24.8-40.2] 0.49x 2,283,157 → 2,283,157
digitsum 150000 54.3 [52.9-62.9] 19.2 [18.5-37.0] 2.83x 50.1 [48.0-63.3] 0.38x 13,372,227 → 6,272,226
isqrt 50000 40.5 [39.7-48.5] 28.8 [28.2-36.9] 1.41x 58.4 [56.8-93.1] 0.49x 9,771,330 → 6,918,614
inversions 1000 20.3 [19.6-26.8] 20.6 [19.9-32.2] 0.99x 30.6 [29.5-65.5] 0.67x 4,007,971 → 4,007,971
inversions 400 8.9 [8.3-79.0] 8.8 [8.3-45.6] 1.00x 17.3 [16.6-57.5] 0.51x 646,212 → 646,212

Development measurements on one 8-vCPU cloud VM pinned to one vCPU; not a GA claim. Other runtimes were not re-measured (see the 2.1.7 notes: C, Rust, Go, LuaJIT and Lua 5.4 are faster everywhere).

Disclosures

  • Model cost: Sley is about 20–25% more expensive than Python per trial on Claude Haiku 5.5 (#81: 1.20x, #84: 1.25x; raw API #90 on main: 1.27x, CI 0.91–1.78). The gap is hidden thinking, not retries. One model, 18 tasks; the set is no longer held out, since the card lines were tuned against two of its tasks. Tokens as reported by Claude Code; $ at API list price.
  • "Instead of" card lines (#83) ship without the planned same-run A/B; they coincided with better count_leq/parse_int and worse mode_value/second_largest/primes.
  • Refusal hints (#75) have no before-measurement (the 2.1.7 baseline was never taken).
  • Small differences within noise: triples 0.96x and fib 1.02x against 2.1.7, with identical instruction counts. The 2.1.6→2.1.7 fib(30) regression (~4%) is not fixed.
  • Soundness gate for #73 was run as a CPython differential (200 random programs, 8,000 assertions, 0 mismatches), not against a forced slow path.
  • Persistent-fuzz proofs were not refreshed (pinned clang 18.1.8 unavailable); the tracked proofs are 2.1.7's. Those lanes' code is unchanged since 2.1.7.
  • check_candidate_result_contract.py and check_transaction_contract.py were waived (carried forward from the 2.1.2 operator decision, not re-confirmed). This is not a pass. ga_claimed stays false; provenance unsigned; reproducibility single-host.
  • The token-loop numbers from 2.1.7 were not re-run; no token-loop claim is made for 2.1.8.

Linux x86_64

sley-2.1.8-linux-x86_64.tar.gz contains static musl sley and sley-agent binaries, built from 3c5b7119c34642abb49480a776bd527ed5f5bec0; two clean builds on the primary host produced identical archives.

  • Size: 7,699,874 bytes; 17 archive members
  • SHA-256: 399e2220b2da0c1561bd5b6e210eea6d06a2062fc3524322dc1a2b9f6ef5cf47
sha256sum -c SHA256SUMS
tar xzf sley-2.1.8-linux-x86_64.tar.gz
cd sley-2.1.8-linux-x86_64
./bin/sley version
Assets 5
Loading