This matrix covers only the increment in the associated pull request. It is not a validation claim for the pre-existing differential methods and it is not independent scientific review.
| Operation | Bounded domain / explicit unsuccessful condition | Independent or known-answer check | Adversarial and property evidence | Status in this increment |
|---|---|---|---|---|
| Integer count validation | Non-negative integer; booleans, fractions, NaN/Inf and unsafe float history rejected | Literal integer and unit-rational fixtures | Negative, fractional, nonfinite, Boolean, 2^54 float, ragged JSON and exponent-token tests |
Implemented; final CI evidence attached to PR |
| Checked totals | Fixed-width overflow throws unless widening was explicitly selected | typemax(Int) boundary |
Fixed-width overflow and mixed BigInt cases |
Implemented |
| Exact sample proportions | Non-negative counts and positive total; zero total gives undefined values | [1,1,1] gives three 1/3; Python fractions.Fraction compares full sample results |
Values beyond 2^53, zero-total samples, scale invariance, feature permutations, random integer matrices with proportions summing to one |
Implemented |
| Exact group proportions | Same count contract; members and group labels are bound to the input fingerprint | Hand-computed two-group fixture | Empty groups and aligned-label failures; group zero-total state | Implemented |
| Observed richness | Number of strictly positive exact counts; empty sample richness is zero | Uniform three-taxon and single-taxon fixtures | Empty and extreme-scale samples | Implemented |
| High-precision Shannon | Positive total; scoped MPFR precision above 53 bits; undefined at zero total | Python decimal references for log(2), log(3), and counts [1,2,3,4] at a 220-digit context |
Extreme rare-taxon scale that underflows Float64; precision round-trip and concurrent 128/256/512/1024-bit scopes |
Implemented; no exactness or sampling-error claim |
| High-precision Gini–Simpson | Positive total; exact (N²-Σc²)/N² intermediate within the configured rational budget; undefined at zero total |
Exact known answers 2/3, 0, and 7/10, converted once to MPFR |
Extreme scale, budget exhaustion, serialization and concurrency tests | Implemented |
| Fixed-decimal display | Display only; exact/storage values remain separate | 1/3, 1/2, and negative half-way fixtures |
Huge BigFloat display avoids Float64 overflow; ties-away-from-zero check |
Implemented |
| Direct JSON boundary | JSON object, explicit validated policy, features-by-samples matrix, 16 MiB / 100,000-cell limits | In-process and real-listener request fixtures | Deterministic malformed-input fuzz, wrong shapes, unsafe integers, float representations and unknown policies | Implemented |
| Run/DuckDB boundary | Read-only database; supported integer column schemas; no missing/negative counts; bounded grouped matrix | Taxon totals computed by hand | HUGEINT pair whose sum overflows SQL SUM, float schema, NULL and negative controls, unsupported sample-read filter |
Implemented |
| Immutable derived analysis | Canonical JSON-safe record; content ID recomputed on load; no replacement of an existing destination | Repeat save/load and canonical-byte round trip | Concurrent publishers, same-ID tampering, invalid record, failed publication cleanup and source/run ownership checks | Implemented |
The high-precision reference generator is
test/references/check_high_precision_references.py; it uses Python's decimal
module and does not read output from the Julia implementation. The CI job checks
that the constants embedded in the Julia tests still match that independent
generator.
- Exact values larger than the I-JSON range, all rationals, and MPFR decimals are
strings in canonical records. JSON response tests cover a value above
2^53. - A stored MPFR decimal is parsed directly into a new
BigFloatat its recorded precision and checked for value equality, without aFloat64intermediate. - Concurrent tests vary precision in spawned tasks and publish the same derived
ID repeatedly. Exactly one publisher may report
created=true; all must see the same complete record. - A modified same-ID record fails its content check and is not overwritten. Failure before publication leaves no final or temporary file.
- The run route uses the existing read-only opener. Derived files live outside the merged results database directory.
bench/validated_statistics.jl measures exact summaries at three table sizes,
high-precision summaries at 64/256/1024 bits, and sequential/concurrent batches.
It reports every trial, median runtime, median allocation, software versions,
platform and thread count. CI retains the resulting JSON artifact for 30 days.
There is intentionally no cross-platform timing threshold and no claim that
higher precision is computationally free.
The following cannot be inferred from the tests above:
- Representative biological validation. Synthetic boundary fixtures are not biological certification. Applicable comparisons with real or curated biological data and any pre-registered reference analyses remain required.
- UI/workflow acceptance. The new modes are exposed through the API in this increment. Browser presentation, accessibility, warnings, provenance workflows and their tests remain an acceptance task; no application TypeScript was added.
- Simulation-based estimator evaluation. These operations are deterministic descriptive transformations, not newly proposed estimators with a claimed confidence/coverage property. A simulation cannot certify the separate pre-existing inferential methods; method-specific simulation and sensitivity studies remain necessary wherever issue #1 requires them.
- Mutation and broader recovery/soak testing. Negative controls and metamorphic properties are present, but a retained mutation-testing campaign, expensive-operation cancellation tests, crash-durability tests and long soak runs are not claimed.
- Independent review and release approval. The implementation author and CI cannot supply independent statistical/scientific review or the project's explicit release approval. Those records must come from the designated reviewers.
- Deferred methods. Exact Fisher/permutation/NB tests and symbolic mathematics are not implemented or unblocked.
Consequently, this matrix supports review of a bounded increment; it does not justify administrative closure of issue #1 on its own.