v0.3.0 — two_sample_test

dsh-probstat v0.3.0

Two independent groups, compared deterministically. New tool two_sample_test joins the six existing ones.

New: two_sample_test

Compare two groups without hand-waving — both estimates, the difference, the standard error, the statistic, the df, all three p-values, the tail-specific critical value and the decision come back together:

kind	when to use it	null distribution
mean_z	difference of means, population sigmas known	normal
mean_welch	difference of means, sigmas unknown, unequal variances	Student-t, Welch–Satterthwaite df
mean_pooled	difference of means, unknown sigmas, equal-variance assumption	Student-t, df = n₁ + n₂ − 2
proportion	difference of two proportions	normal, pooled proportion

• delta0 sets the null-hypothesis difference (default 0), so non-inferiority style questions work without reshaping the data.
• The Welch df is fractional and reported unrounded — rounding it to an integer silently moves the p-value.
• pooledVariance / pooledProportion are reported so the equal-variance / homogeneous-proportion assumption is checkable rather than assumed.
• A shaky normal approximation (smallest expected cell < 5) is flagged in note; a zero standard error (both groups all-success or all-fail) is an error, not a fake statistic.

kind: "mean_welch", mean1: 5.2, mean2: 4.6, sd1: 1.1, sd2: 0.9, n1: 12, n2: 15
→ t 1.5248215, df 21.166825, two-sided p 0.1421069, critical 2.0786163 → do not reject at 0.05

Verification

• 59 tests (was 50), coverage 86.97% → 88.12%.
• Every numeric expectation comes from the in-repo oracle test/oracle/anchors.py; the new section computes the t tail by Simpson quadrature on the t density instead of the continued-fraction incomplete beta used by the plugin — a shared mistake cannot pass both sides. Cross-checked against published t-table values (2.228138852 / 2.042272456 / 1.795884819).
• Oracle self-check covers invariants: p-values in [0, 1], two-sided = 2 × smaller tail, Welch df < pooled df on unequal variances, pooled variance = df-weighted mean of the two variances, pooled proportion = total successes over total n.

安装 / Install:

dsh plugin --profile web add github:TYEclipse/dsh-probstat#v0.3.0

中文摘要

新增 two_sample_test：双样本检验，一次给出两组估计、差值、标准误、统计量、自由度、三个 p 值与 α 判定。

• 四种口径：均值差（σ 已知 z / σ 未知 Welch / 等方差合并方差 t）+ 双比例差（合并比例正态近似）。
• Welch 自由度按 Welch–Satterthwaite 原值报告不取整（取整会静默改变 p 值）。
• 合并方差与合并比例一并给出，假设可核对；比例口径最小期望单元格 < 5 时提示近似不可靠；标准误为 0 时报错而不给无意义统计量。
• 测试 50 → 59，覆盖率 86.97% → 88.12%；期望值全部由仓内 oracle 生成，新区的 t 尾概率改用 Simpson 数值积分独立实现，并与公开 t 表值交叉验证。

v0.2.0

dsh-probstat v0.2.0

Two new tools — the "what does this data actually say?" half of statistics, next to the existing distribution/interval math.

EN

• hypothesis_test — one-sample tests with the decision spelled out: mean_z (known sigma), mean_t (unknown sigma, Student-t with df = n − 1) and proportion (against p0, normal approximation). Each run returns the statistic, the standard error, all three p-values (two-sided / left / right), the critical value for the requested tail, and whether the null is rejected at alpha (default 0.05). The proportion kind adds a note when n·p0 < 5 instead of quietly handing back a p-value.
• sample_size — required sample size for a target margin of error: mean_z closed form n = (z·σ/E)², proportion via the Wald closed form (method: "wald", default) or the smallest n whose Wilson half-width meets E (method: "wilson", binary search). p defaults to 0.5, the conservative worst case; impossible targets return valid: false with the reason.

中文

• hypothesis_test：单样本假设检验——σ 已知走 z、σ 未知走 t（df = n−1）、单比例检验（对 p0 的正态近似）。一次给出统计量、标准误、三个 p 值（双侧/左/右）、所请求尾部的临界值与 α 判定（默认 0.05）。比例检验在 n·p0 < 5 时用 note 明确提示近似不可靠。
• sample_size：目标误差限所需样本量——均值闭式 n = (z·σ/E)²；比例可选 Wald 闭式（默认）或 Wilson 口径（二分求 Wilson 半宽 ≤ E 的最小 n）；p 默认 0.5（最保守），目标不可达时报错并说明原因。

Quality

• Tests 37 → 50; coverage 83.70% → 86.97% (baseline ratcheted up).
• Oracle moved to test/oracle/anchors.py (independent Python implementation: math.erf, math.comb, bisection, independent t and Wilson code) with a --check mode cross-validating published z/t-table constants. Every numeric assertion in the test file is generated there — including all legacy anchors.
• The TS implementation and the Python oracle agree to the documented precision; the one place they differ in the 8th significant digit (a squared critical value) is pinned at 5 digits with the reason in the test.

Install:

dsh plugin --profile web add github:TYEclipse/dsh-probstat

v0.1.0

v0.1.0 — initial release

First release of dsh-probstat: deterministic probability & statistical-inference math for DeepSeek Harness, zero runtime dependencies.

Four tools:

• dist_calc — pdf / cdf / survival / quantile / stats for six distributions: normal, binomial, poisson, exponential, uniform, geometric
• z_score — standard-normal table math: z ↔ probability/percentile, two-sided p-values, interval probabilities
• confidence_interval — mean intervals (known σ → z; unknown σ → Student-t, df = n−1) and proportion intervals (Wilson score method)
• event_probability — union / intersection / conditional / Bayes / complement / at-least-one identities

Numerics: normal cdf via A&S 7.1.26 erf fit (|err| ≤ 1.5e-7), quantile via Acklam inverse normal (rel. err < 1.15e-9); binomial/poisson exact recurrence summation; Student-t via Lentz continued fraction + Lanczos log-gamma, cross-checked against published t-table values. 37 tests, anchors generated by an independent Python oracle (test/anchors-probstat.py) and cross-validated with published z/t-table values.

---

v0.1.0 — 首发版本

dsh-probstat 首发：DeepSeek Harness 概率与统计推断数学工具箱，零运行时依赖。

四个工具：

• dist_calc — 六种分布（正态/二项/泊松/指数/均匀/几何）的 pdf / cdf / 生存函数 / 分位数 / 统计量
• z_score — 标准正态表换算：z ↔ 概率/百分位、双侧 p 值、区间概率
• confidence_interval — 均值置信区间（σ 已知走 z；未知走 t，df = n−1）+ 比例 Wilson 区间
• event_probability — 并 / 交 / 条件 / 贝叶斯 / 补集 / 至少一次成功恒等式

数值方法：正态 CDF 用 A&S 7.1.26 erf 拟合（绝对误差 ≤ 1.5e-7），分位数用 Acklam 逆正态（相对误差 < 1.15e-9）；二项/泊松递推精确求和；t 分布用连分式不完全贝塔 + Lanczos log-Gamma，并与公开 t 表值交叉验证。37 项测试，锚点由独立 Python oracle（test/anchors-probstat.py）生成。

Install: dsh plugin --profile web add github:TYEclipse/dsh-probstat