robustrep 0.1.0__py3-none-any.whl

robustrep/__init__.py

@@ -0,0 +1,6 @@
+"""robustrep: robust, transport-agnostic reputation scoring for AI agents."""
+from .config import Config
+from .pipeline import score
+
+__all__ = ["Config", "score"]
+__version__ = "0.1.0"

robustrep/aggregate.py

@@ -0,0 +1,262 @@
+"""Aggregators turn (values, weights) into (score, ci_low, ci_high).
+
+The arithmetic mean has breakdown point zero: a single extreme rating can move
+it arbitrarily far. The weighted median has breakdown point ~0.5 (roughly half
+the total weight must be adversarial to move it), so it is the default
+aggregator here. `Aggregator` is an abstract base so Bayesian and graph-based
+aggregators can be added later behind the same interface.
+"""
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+
+import numpy as np
+
+from .config import Config
+
+# Half-weight threshold is nudged down by this relative fraction so that
+# float64 cumsum drift at an exact tie doesn't step one element too far
+# (which would break scale invariance and bias the result upward).
+_TIE_EPS = 1e-9
+
+# Minimum fraction of bootstrap resamples that must carry non-zero total
+# weight for the CI to be considered meaningful.
+MIN_VALID_BOOT_FRACTION = 0.1
+
+# Cap on the number of (resample x vote) float64 cells materialized at once
+# by `bootstrap_ci`'s chunked multinomial draw, so peak memory stays bounded
+# even for a ratee with thousands of votes: a chunk holds at most this many
+# cells regardless of how large `n_boot` is.
+_MAX_BOOT_CHUNK_CELLS = 2_000_000
+
+
+def _validate(values: np.ndarray, weights: np.ndarray, name: str) -> tuple[np.ndarray, np.ndarray]:
+    """Shared input validation for `weighted_median` and `bootstrap_ci`.
+
+    Coerces both to 1-D float64 arrays and validates: 1-D, non-empty, equal
+    length, all-finite (no NaN/+/-inf), weights non-negative, and weights
+    summing to > 0. Returns the coerced `(values, weights)`. Raises
+    ValueError (messages prefixed with `name`, the caller's own name) on any
+    violation.
+    """
+    values = np.asarray(values, dtype=float)
+    weights = np.asarray(weights, dtype=float)
+    if values.ndim != 1 or weights.ndim != 1:
+        raise ValueError(f"{name}: values/weights must be 1-D")
+    if values.shape[0] == 0:
+        raise ValueError(f"{name}: values/weights must not be empty")
+    if values.shape != weights.shape:
+        raise ValueError(f"{name}: values and weights must have the same length")
+    if not np.isfinite(values).all() or not np.isfinite(weights).all():
+        raise ValueError(f"{name}: values/weights must not contain non-finite entries (NaN/inf)")
+    if (weights < 0).any():
+        raise ValueError(f"{name}: weights must be non-negative")
+    if weights.sum() <= 0:
+        raise ValueError(f"{name}: weights must sum to > 0")
+    return values, weights
+
+
+def _weighted_median_pos(cum: np.ndarray) -> int:
+    """Position in an ascending cumulative-weight array selected by the lower-weighted-
+    median convention: the first index at which cumulative weight reaches >= half the
+    total (`np.searchsorted(..., side="left")`), with the half-weight threshold nudged
+    down by `_TIE_EPS` to absorb float64 cumsum drift at exact ties. Shared by
+    `_weighted_median_sorted` and `weighted_median_index` so this tie-tolerance logic
+    lives in exactly one place. `cum` must be non-empty and non-decreasing (a cumsum).
+    """
+    half = 0.5 * cum[-1]
+    return int(np.searchsorted(cum, half * (1 - _TIE_EPS), side="left"))
+
+
+def _weighted_median_sorted(v: np.ndarray, w: np.ndarray) -> float:
+    """Weighted median of already-sorted, already-validated inputs.
+
+    Assumes `v` is sorted ascending and `w` is finite, non-negative, and sums
+    to > 0. Not for external use (no validation) — see `weighted_median`.
+    """
+    return float(v[_weighted_median_pos(np.cumsum(w))])
+
+
+def weighted_median(values: np.ndarray, weights: np.ndarray) -> float:
+    """Return the weighted median of `values` under `weights`.
+
+    Tie/even-count convention: this is the *lower* weighted median. Sorting
+    values ascending and walking cumulative weight, the returned value is the
+    first one at which cumulative weight reaches >= half the total weight
+    (`np.searchsorted(..., side="left")`, with the half-weight threshold
+    nudged down by a small epsilon to absorb float64 cumsum drift at exact
+    ties). For an even count with equal weights this picks the smaller of the
+    two middle values (e.g. [0.2, 0.8] with equal weights returns 0.2),
+    matching a robust, deterministic, order-independent convention rather
+    than numpy's mean-of-two-middles.
+
+    Raises ValueError if `values`/`weights` are not 1-D, are empty,
+    mismatched in length, contain NaN or +/-inf, or if any weight is
+    negative or all weights are zero (see `_validate`).
+    """
+    values, weights = _validate(values, weights, "weighted_median")
+    order = np.argsort(values, kind="stable")
+    return _weighted_median_sorted(values[order], weights[order])
+
+
+def weighted_median_index(values: np.ndarray, weights: np.ndarray) -> int:
+    """Return the *original* index of the element `weighted_median` would return.
+
+    Same validation, stable sort, and tie tolerance as `weighted_median` (see
+    `_validate` and `_weighted_median_sorted`) -- `values[weighted_median_index(values,
+    weights)] == weighted_median(values, weights)` always holds. Useful when a caller
+    needs to know *which* input element was selected (e.g. `explain()` marking the
+    vote/tag chosen by the weighted-median chain), not just its value.
+    """
+    values, weights = _validate(values, weights, "weighted_median_index")
+    order = np.argsort(values, kind="stable")
+    cum = np.cumsum(weights[order])
+    return int(order[_weighted_median_pos(cum)])
+
+
+def _bootstrap_draws(
+    v_sorted: np.ndarray,
+    w_sorted: np.ndarray,
+    n_boot: int,
+    rng: np.random.Generator,
+    chunk_size: int,
+) -> np.ndarray:
+    """The array of valid weighted-median bootstrap draws (length <= n_boot).
+
+    `v_sorted`/`w_sorted` must already be sorted ascending by value. Draws
+    are generated `chunk_size` resamples at a time via
+    `rng.multinomial(n, [1/n]*n, size=chunk_size)` (see `bootstrap_ci` for
+    why a multinomial count draw is equivalent to classic index resampling
+    with replacement) and accumulated, so peak memory is bounded by
+    `chunk_size * n` regardless of `n_boot`. A resample whose weights sum to
+    zero has an undefined weighted median and is dropped from the result
+    rather than counted.
+    """
+    n = len(v_sorted)
+    pvals = np.full(n, 1.0 / n)
+    boots = []
+    remaining = n_boot
+    while remaining > 0:
+        take = min(chunk_size, remaining)
+        remaining -= take
+        counts = rng.multinomial(n, pvals, size=take)
+        cum = np.cumsum(counts * w_sorted, axis=1)
+        valid = cum[:, -1] > 0
+        cum_valid = cum[valid]
+        half = 0.5 * cum_valid[:, -1:]
+        idx = (cum_valid >= half * (1 - _TIE_EPS)).argmax(axis=1)
+        boots.append(v_sorted[idx])
+    return np.concatenate(boots) if boots else np.array([], dtype=float)
+
+
+def bootstrap_ci(
+    values: np.ndarray,
+    weights: np.ndarray,
+    n_boot: int,
+    rng: np.random.Generator,
+    ci_level: float,
+) -> tuple[float, float]:
+    """Vectorized bootstrap (lo, hi) quantiles of the weighted median.
+
+    Equivalent to resampling n `(value, weight)` pairs with replacement
+    `n_boot` times and taking the weighted median of each resample, but
+    without a Python-level loop over resamples. A resample's weighted
+    median depends only on how many times each *position* was drawn (its
+    resample count), not on the order of the draws, so after sorting
+    `values` once (stable), `rng.multinomial(n, [1/n]*n, size=chunk)`
+    produces a batch of resamples' per-position counts in a single call --
+    equivalent to drawing n indices with replacement, once per resample --
+    and the cumulative sum of `counts * sorted_weights` along each row gives
+    every resample's weighted median in one vectorized pass (see
+    `_bootstrap_draws`). Processed in chunks of at most
+    `max(1, 2_000_000 // n)` resamples so peak memory stays bounded even for
+    a ratee with thousands of votes; this is what keeps the per-ratee
+    bootstrap affordable at ~28k ratees.
+
+    Inputs are validated exactly like `weighted_median` (see `_validate`).
+    A resample whose weights sum to zero has an undefined weighted median
+    and is skipped rather than counted. If fewer than `max(2, n_boot // 10)`
+    resamples remain valid, raises ValueError instead of silently reporting
+    a CI built from too few (or zero) samples. Skipping all-zero resamples
+    conditions the CI on non-zero weight; unreachable in-pipeline since
+    Config forbids zero weights.
+    """
+    values, weights = _validate(values, weights, "bootstrap_ci")
+    n = len(values)
+    order = np.argsort(values, kind="stable")
+    v_sorted, w_sorted = values[order], weights[order]
+
+    chunk_size = max(1, _MAX_BOOT_CHUNK_CELLS // n)
+    boots = _bootstrap_draws(v_sorted, w_sorted, n_boot, rng, chunk_size)
+
+    min_valid = max(2, int(n_boot * MIN_VALID_BOOT_FRACTION))
+    if boots.size < min_valid:
+        raise ValueError("bootstrap degenerate: too many zero-weight resamples")
+
+    alpha = (1 - ci_level) / 2
+    lo, hi = np.quantile(boots, [alpha, 1 - alpha])
+    return float(lo), float(hi)
+
+
+def _bootstrap_ci(
+    values: np.ndarray,
+    weights: np.ndarray,
+    n_boot: int,
+    seed: int,
+    ci_level: float,
+) -> tuple[float, float]:
+    """Bootstrap (lo, hi) quantiles of the weighted median, seeded by a plain
+    integer `seed`. Thin wrapper around `bootstrap_ci` (the one bootstrap
+    implementation) for callers that don't manage their own
+    `numpy.random.Generator`.
+    """
+    return bootstrap_ci(values, weights, n_boot, np.random.default_rng(seed), ci_level)
+
+
+class Aggregator(ABC):
+    """Abstract base for turning per-ratee (values, weights) into a score with a CI.
+
+    Concrete implementations (e.g. `WeightedMedian`, and future Bayesian or
+    graph-based aggregators) all expose the same `aggregate` interface so the
+    pipeline can swap aggregation strategies without changing callers.
+    """
+
+    @abstractmethod
+    def aggregate(self, values: np.ndarray, weights: np.ndarray) -> tuple[float, float, float]:
+        """Return (score, ci_low, ci_high)."""
+
+
+class WeightedMedian(Aggregator):
+    """Weighted-median aggregator with a bootstrap confidence interval.
+
+    `aggregate` resamples (values, weights) pairs with replacement `n_boot`
+    times, computes the weighted median of each resample, and reports the
+    `ci_level` central interval of the bootstrap distribution (widened if
+    needed so it always contains the point estimate). Note: with tied or
+    coarse-grained values the percentile bootstrap can legitimately return a
+    zero-width interval (every resample lands on the same value) — consumers
+    should read the CI together with the underlying vote count rather than
+    treating a narrow interval alone as high confidence.
+    """
+
+    def __init__(self, n_boot: int = 1000, seed: int = 0, ci_level: float = 0.95):
+        if n_boot < 0:
+            raise ValueError("WeightedMedian: n_boot must be >= 0")
+        if not 0 < ci_level < 1:
+            raise ValueError("WeightedMedian: ci_level must be in (0, 1)")
+        self.n_boot, self.seed, self.ci_level = n_boot, seed, ci_level
+
+    @classmethod
+    def from_config(cls, cfg: Config) -> "WeightedMedian":
+        """Build a WeightedMedian aggregator from a pipeline Config."""
+        return cls(n_boot=cfg.bootstrap_n, seed=cfg.bootstrap_seed, ci_level=cfg.ci_level)
+
+    def aggregate(self, values: np.ndarray, weights: np.ndarray) -> tuple[float, float, float]:
+        values = np.asarray(values, dtype=float)
+        weights = np.asarray(weights, dtype=float)
+        point = weighted_median(values, weights)
+        if len(values) == 1 or self.n_boot == 0:
+            return point, point, point
+        lo, hi = _bootstrap_ci(values, weights, self.n_boot, self.seed, self.ci_level)
+        # defensive; not expected to fire
+        return point, float(min(lo, point)), float(max(hi, point))