№ 01fol. 1 / 12
Statistics
- stats.median
- stats.quantile
- regression.ols_multiple
The compute plane
445 modules.
One pinned runtime.
SQAI's compute plane is a library of compiled kernels: statistics to option pricing to LLM infrastructure, every function typed in the 4,778-capability contract. An agent names a function. The engine returns the number — warm, in under a millisecond.
445modules
100+applied packs
4,778capabilities in the contract
0.83–0.93 mswarm compute, measured
IIThe specimen book
Twelve families, set in type.
Read the library like a type specimen: each family a page, each function a face. Every name on these pages is real — typed in the contract, checked before it runs, hashed after.
№ 02fol. 2 / 12
Hypothesis testing
- stat_tests.pearson_r
- ab_testing.sample_size_proportion
№ 03fol. 3 / 12
Probability
- distributions.sample_normal
- distributions.sample_poisson
№ 04fol. 4 / 12
Time series
- timeseries_stats.holt_linear
- timeseries_stats.acf
№ 05fol. 5 / 12
Linear algebra
- linalg.mat_inv
- matrix_decomp.cholesky
№ 06fol. 6 / 12
Optimization
- root_finding.newton_raphson
- integrate.simpson
- operations_research.eoq_quantity
№ 07fol. 7 / 12
Finance & risk
- finance.npv
- finance.irr
- finance.cagr
- finance.max_drawdown
- risk.beta
- risk.conditional_var
№ 08fol. 8 / 12
Option pricing
- option_pricing.black_scholes_call
- option_pricing.black_scholes_put
- option_pricing.implied_volatility
№ 09fol. 9 / 12
Machine learning
- ml.classifiers.knn_predict
- ml.classifiers.logistic_predict
- ml.clustering.kmeans_fit
№ 10fol. 10 / 12
Signal
- fft.power_spectrum
- wavelet.denoise_haar
№ 11fol. 11 / 12
Information theory
- information_theory.entropy
- information_theory.kl_divergence
№ 12fol. 12 / 12
LLM infrastructure
- flash_attention.flash_block_output
- kv_cache.incremental_decode_kv_append
- rag_retrieval.bm25_term_score
The type case — applied packs
actuarial → viticulture
- 01actuarial
- 02aerospace
- 03epidemiology
- 04hydrology
- 05pharmacokinetics
- 06robotics
- 07seismology
- 08thermodynamics
- 09viticulture
- 100+applied packs in the case — nine drawers pulled here. Every one answers on the same contract: same policy checks, same provenance envelope.
IIIThe presswork
Compiled speed. Proven accuracy.
Behind every family: hand-optimized kernels, compiled ahead of time. Determinism costs nothing here — float64, single thread, and still faster than interpreted baselines by orders of magnitude.
Speed
91.86×median speedup vs interpreted baselines
- 111.24× mean · 1,593× maximum speedup
- measured across 4,207 compared functions on the public scorecard
- 0.83–0.93 ms warm compute · sub-ms query plane
Accuracy
0mismatches across 2,744 correctness checks
- 2,458 fuzz checks across 800+ datasets and 8 distributions
- 232 edge cases · 54 suite checks · bit-exact sorting
- maximum numeric error: 1.8e-15
0/16 → 16/16same model, engine computing
The accuracy A/B: the same model answers 16 quantitative questions. Computing alone, it scores 0/16. With the engine doing the computing, 16/16.
~110 seconds once.
0.83–0.93 ms after.
download · verify · pin — then it's just arithmetic
~110 sone-time provision — download, verify, pin
0.83–0.93 mswarm compute, measured
sub-msquery plane, every answer
VThe measured run
One call, receipt included.
One real computation, end to end. The same call returns the same value and the same hash in TypeScript and Python — and the runtime that executes it can be built down to exactly the modules you ship.
measured run — finance.npv
call
finance.npv(0.1, [-1000, 300, 420, 560, 680])value
505.020148896933identical in TypeScript and Pythonelapsed
0.83 mswarmcomputation_hash
b74f67d0d7a594aa7ac91f6291612452aa8ccdf603351ebc8d801a6fddd91bc8A number you can replay, not just read.
Ship only what you use.
One environment variable decides what the build service compiles.
- SQAI_RUNTIME_MODULES
- Name the modules your product needs. Only those enter the build.
- POST /v1/runtime/build
- The build service compiles a runtime to exactly that filter, on provision.
- filter-hash
- Bundles are keyed by it. The same filter always resolves to the same bundle.
Discovery stays in-process.
Agents don't guess the surface — they search it, locally.
- searchCapabilities
- Searches the contract embedded in the package. In-process — no network call.
- unsupported_operation
- An unknown name is refused with a structured error, plus nearest_matches to correct course.
VIQuestions
The fine points.
How fast is compute, really?
The first computation provisions the pinned runtime — about 110 seconds, once. After that, measured warm latency is 0.83–0.93 ms per computation, and the query plane answers in under a millisecond.
Can I ship a smaller runtime?
Yes. Set SQAI_RUNTIME_MODULES and the runtime is built to that filter on provision. Bundles are keyed by the filter's hash, so the same filter reproduces the same bundle.
Are results really identical across languages?
By construction. Kernels run in float64 on a single thread, and results are hashed over canonical JSON. finance.npv(0.1, [-1000, 300, 420, 560, 680]) returns 505.020148896933 with the same computation_hash in TypeScript and Python.
How does an agent know what it can compute?
It searches. searchCapabilities runs in-process over the contract embedded in the package, and an unknown function name returns unsupported_operation with nearest_matches — never a guess.