prime-radiant 0.1.0__tar.gz

prime_radiant-0.1.0/.gitignore

@@ -0,0 +1,23 @@
+# Secrets — never commit
+.env
+
+# Data and run artifacts — root-anchored: src/prime_radiant/epi/data is CODE
+/data/
+/model-output/
+*.duckdb
+logs/
+.coverage
+
+# Python
+.venv/
+__pycache__/
+*.pyc
+.pytest_cache/
+.ruff_cache/
+dist/
+
+# OS
+.DS_Store
+
+# MkDocs build output
+site/

prime_radiant-0.1.0/CHANGELOG.md

@@ -0,0 +1,32 @@
+# Changelog
+
+All notable changes to this project are documented in this file.
+
+The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+Releases below the insertion marker are generated by python-semantic-release
+from the conventional commit history.
+
+<!-- version list -->
+
+## v0.1.0 (2026-09-01)
+
+- Initial Release
+
+## Pre-release history (unversioned, 2026-08)
+
+The project was built in adversarially-verified phases before its first tagged
+release; the full record lives in `NOTES/restart.md` and the `[claude]`-prefixed
+commit history.
+
+- **Phase 2A–2B** — FluSight data layer on git-vintage truth; WIS scorer
+  cross-validated against the official pipeline (relative WIS 0.999989).
+- **Phase 2C** — pooled LightGBM quantile model; beats FluSight-baseline.
+- **Phase 2D** — three-season league tables; LightGBM wins 2025-26 (relative
+  WIS 0.609 vs UMass-flusion 0.625), loses the two earlier seasons — printed.
+- **Phase 2E** — submission automation: `prime-radiant epi forecast|validate`,
+  hub-valid CSV writer, registration metadata, five-gate inert live path.
+- **Phase F** — Gradio dashboard on a precomputed serve bundle; deployed to
+  Hugging Face Spaces; FluSight registration PR opened.
+- **Phase G** — release hardening: CI, pre-commit, Dependabot, release and
+  publish automation, Docker, docs, and the public flip.

prime_radiant-0.1.0/CITATION.cff

@@ -0,0 +1,24 @@
+cff-version: 1.2.0
+message: "If you use this software, please cite it as below."
+title: "Prime Radiant"
+type: software
+authors:
+  - family-names: Gracey
+    given-names: Jeremy
+    email: jeremy.a.gracey@gmail.com
+version: 0.1.0
+date-released: 2026-08-31
+license: MIT
+repository-code: "https://github.com/JeremyGracey-AI/prime-radiant"
+url: "https://huggingface.co/spaces/jeremygracey-ai/prime-radiant"
+abstract: >-
+  Calibrated forecasting: a CDC FluSight quantile forecaster (pooled LightGBM
+  quantile regression ensembled with a FluSight-baseline replica, WIS-scored on
+  vintage-honest rolling-origin backtests) and a parked Metaculus LLM
+  forecasting thread.
+keywords:
+  - forecasting
+  - calibration
+  - flusight
+  - epidemiology
+  - quantile-regression

prime_radiant-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Jeremy Gracey
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

prime_radiant-0.1.0/PKG-INFO

@@ -0,0 +1,179 @@
+Metadata-Version: 2.5
+Name: prime-radiant
+Version: 0.1.0
+Summary: Calibrated forecasting: Metaculus LLM bot (Brier) and CDC FluSight quantile forecaster (WIS).
+Project-URL: Homepage, https://github.com/JeremyGracey-AI/prime-radiant
+Project-URL: Repository, https://github.com/JeremyGracey-AI/prime-radiant
+Author-email: Jeremy Gracey <jeremy.a.gracey@gmail.com>
+License-Expression: MIT
+License-File: LICENSE
+Keywords: calibration,epidemiology,flusight,forecasting,metaculus,quantile
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering
+Requires-Python: >=3.11
+Requires-Dist: epiweeks>=2.3
+Requires-Dist: forecasting-tools>=0.2.92
+Requires-Dist: httpx>=0.27
+Requires-Dist: lightgbm==4.7.0
+Requires-Dist: matplotlib<3.12,>=3.11
+Requires-Dist: pandas>=2.2
+Requires-Dist: pandera>=0.20
+Requires-Dist: pyarrow>=17
+Requires-Dist: pydantic-settings>=2.5
+Requires-Dist: pydantic>=2.9
+Requires-Dist: python-dotenv>=1.0
+Requires-Dist: pyyaml>=6.0.3
+Description-Content-Type: text/markdown
+
+# Prime Radiant
+
+[![ci](https://github.com/JeremyGracey-AI/prime-radiant/actions/workflows/ci.yml/badge.svg)](https://github.com/JeremyGracey-AI/prime-radiant/actions/workflows/ci.yml)
+[![AI-USE: declared](https://img.shields.io/badge/AI--USE-declared-2ea44f)](./AI-USE.md)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](./LICENSE)
+[![Dashboard](https://img.shields.io/badge/%F0%9F%A4%97%20Space-live-blue)](https://huggingface.co/spaces/jeremygracey-ai/prime-radiant)
+[![docs](https://img.shields.io/badge/docs-live-2ea44f)](https://jeremygracey-ai.github.io/prime-radiant/)
+<!-- At first release / first coverage upload these join (kept out until true):
+     codecov, PyPI version — see NOTES/go-live-runbook.md -->
+
+Calibrated forecasting, measured honestly. Two threads share this repo:
+
+- **FluSight epi forecaster** (`prime_radiant.epi`) — CDC FluSight-format quantile
+  forecasts of weekly US influenza hospital admissions: LightGBM quantile regression
+  in the shape of the FluSight-2023/24-winning flusion model, ensembled with a
+  validated replica of CDC's own baseline. Pure statistical pipeline; no LLM calls.
+- **Metaculus bot** (`prime_radiant.metaculus`) — LLM forecasting on binary
+  questions (parked at its retrieval phase; client + tests remain green).
+
+## Backtest results (three seasons, vintage-honest)
+
+Rolling-origin backtests over 85 weekly origins across the 2023-24, 2024-25 and
+2025-26 seasons. Every forecast is trained only on the data snapshot a live
+forecaster would have held on the Wednesday submission evening (the hub's git
+history is the vintage store); scored against final truth as of 2026-07-09,
+horizons 0-3, all 53 jurisdictions. `wis_rel` = mean WIS relative to the official
+FluSight-baseline on the common task set across all six models (lower is better;
+the official "scaled relative skill" collapses to this ratio on identical sets).
+Coverage columns are computed on each model's full scored set; relative-skill
+columns on the common set — `n` and `n_relative` in the CSVs disclose both.
+
+**2025-26** — our model leads the table on the natural scale (on the log(x+1)
+scale UMass-flusion edges it, 0.583 vs 0.585 — both columns are in the CSVs):
+
+| model | wis_rel | 50% cov | 95% cov |
+|---|---|---|---|
+| **prime-radiant-lgbm** | **0.609** | 0.398 | 0.818 |
+| UMass-flusion | 0.625 | 0.353 | 0.823 |
+| FluSight-ensemble | 0.666 | 0.526 | 0.903 |
+| prime-radiant-ensemble | 0.764 | 0.464 | 0.880 |
+| FluSight-baseline | 1.000 | 0.433 | 0.864 |
+
+**2024-25** — the multi-model ensembles beat us; we beat the baseline:
+
+| model | wis_rel | 50% cov | 95% cov |
+|---|---|---|---|
+| UMass-flusion | 0.669 | 0.397 | 0.820 |
+| FluSight-ensemble | 0.675 | 0.519 | 0.818 |
+| **prime-radiant-lgbm** | **0.796** | 0.338 | 0.738 |
+| prime-radiant-ensemble | 0.900 | 0.371 | 0.744 |
+| FluSight-baseline | 1.000 | 0.317 | 0.719 |
+
+**2023-24** — flusion dominates; our ensemble edges FluSight-ensemble:
+
+| model | wis_rel | 50% cov | 95% cov |
+|---|---|---|---|
+| UMass-flusion | 0.569 | 0.569 | 0.964 |
+| **prime-radiant-ensemble** | **0.716** | 0.422 | 0.909 |
+| FluSight-ensemble | 0.730 | 0.488 | 0.920 |
+| prime-radiant-lgbm | 0.841 | 0.369 | 0.831 |
+| FluSight-baseline | 1.000 | 0.282 | 0.897 |
+
+Full tables (per-horizon rows, log-scale variants, AE-median, task counts):
+[`reports/backtest_<season>.csv`](reports/). Calibration:
+
+![Calibration curves](reports/calibration.png)
+
+## Honest framing
+
+- **The backtest is net biased against us, and we keep it that way.** Adversarial
+  verification proved that on three 2024-25 holiday weeks the official baseline's
+  run saw data committed Thursday+, which our live-Wednesday vintage discipline
+  refuses; on the 24 information-equal origins our relative WIS improves (lgbm
+  ~0.73). It is not one-way: at three October-2023 origins our vintages were
+  fresher than what the official run used. One 2023-24 origin (2024-04-13) had a
+  week-stale vintage — for our models only; the officials ran on fresh data there.
+- **Our intervals are too narrow.** The lgbm's 50% intervals cover 34-40% and its
+  95% intervals 74-83% — under-dispersed, visible in the calibration curves.
+  FluSight-ensemble is better calibrated even where we beat it on WIS. This is
+  the main modeling debt; season-level bagging (flusion's stabilizer) is the
+  known lever.
+- **Wins and losses both stand.** We lead 2025-26 outright; UMass-flusion beats
+  us clearly in 2023-24 and 2024-25. The 2023-24 lgbm ran on ~1.2 seasons of
+  training history.
+- **The scorer and baseline replica are self-validating — for the season they
+  were validated on.** The replica reproduces the official 2024-25 baseline to
+  relative WIS 0.99999 on fingerprint-matched vintages (Phase B). In 2023-24 the
+  replica scores 0.976 vs the official baseline — a spread-construction
+  divergence at matched anchors that Phase B's validation (2024-25-era official
+  code) does not cover; adversarially checked: substituting the actual official
+  baseline into our ensemble changes no table ordering.
+
+## Methodology in one paragraph
+
+NHSN weekly admissions (hub target data, git-vintaged) → per-100k 4th-root
+transform with per-location scale/center fitted per origin → pooled LightGBM
+quantile regression (23 levels, horizon-as-feature, deterministic, exact-pinned
+4.7.0) → monotone sort, inversion, integer rounding at the hub boundary →
+per-quantile-median ensemble with the baseline replica → WIS/coverage scoring on
+natural and log(x+1) scales, pinned-truth, common-task relative skill. Every
+stage boundary carries a pandera contract; leakage invariants (vintage as-of,
+feature cut, scaler fit) are hypothesis-tested properties.
+
+## Setup
+
+```sh
+uv sync --all-groups
+make check            # offline gates: ruff + format + pyright + pytest-cov
+make test-integration # network: real hub clone, S3 benchmarks, gate + reports
+```
+
+Requires Homebrew `libomp` on macOS for LightGBM. Built with
+[Claude Code](https://claude.com/claude-code) (agentic coding assistant by
+Anthropic); every phase was adversarially verified
+by refuter agents before being declared done — see `NOTES/restart.md` and the
+`[claude]`-prefixed commit history.
+
+## Dashboard
+
+**Live: https://huggingface.co/spaces/jeremygracey-ai/prime-radiant**
+
+A Gradio dashboard (US choropleth of predicted 3-week change, per-state fan
+charts, reliability curves, model-vs-baseline league tables) serves the frozen
+backtest record from `serve_data/`, a ~1.7MB precomputed bundle:
+
+```sh
+make bundle                        # offline: rebuilds serve_data/ deterministically
+uv run python dashboard/app.py     # local: http://127.0.0.1:7860
+```
+
+The Space (`jeremygracey-ai/prime-radiant`, CPU-basic) installs only
+gradio/plotly/pandas/pyarrow — never this package, so no LightGBM/libomp and no
+hub clone at serve time. `.github/workflows/space-deploy.yml` stages and
+validates the Space tree on every dispatch; the actual push is triple-gated
+(manual `deploy=true` + `SPACE_LIVE=1` repo var + `HF_TOKEN` secret) and stays
+inert until go-live.
+
+## Limitations
+
+- Single data source (NHSN), single target (`wk inc flu hosp`); no rate-change /
+  peak / ED-visit targets yet.
+- Interval under-dispersion as above; no season-bagging yet.
+- 2022-23 is not backtestable (the hub's vintage history begins Oct 2023).
+- Live submission machinery exists (Phase E) but is structurally inert: the
+  weekly workflow's live path is triple-gated and unimplemented past its gates,
+  and nothing submits anywhere without the go-live runbook being executed by
+  hand. The dashboard serves a frozen bundle, not a live feed.

prime_radiant-0.1.0/README.md

@@ -0,0 +1,147 @@
+# Prime Radiant
+
+[![ci](https://github.com/JeremyGracey-AI/prime-radiant/actions/workflows/ci.yml/badge.svg)](https://github.com/JeremyGracey-AI/prime-radiant/actions/workflows/ci.yml)
+[![AI-USE: declared](https://img.shields.io/badge/AI--USE-declared-2ea44f)](./AI-USE.md)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](./LICENSE)
+[![Dashboard](https://img.shields.io/badge/%F0%9F%A4%97%20Space-live-blue)](https://huggingface.co/spaces/jeremygracey-ai/prime-radiant)
+[![docs](https://img.shields.io/badge/docs-live-2ea44f)](https://jeremygracey-ai.github.io/prime-radiant/)
+<!-- At first release / first coverage upload these join (kept out until true):
+     codecov, PyPI version — see NOTES/go-live-runbook.md -->
+
+Calibrated forecasting, measured honestly. Two threads share this repo:
+
+- **FluSight epi forecaster** (`prime_radiant.epi`) — CDC FluSight-format quantile
+  forecasts of weekly US influenza hospital admissions: LightGBM quantile regression
+  in the shape of the FluSight-2023/24-winning flusion model, ensembled with a
+  validated replica of CDC's own baseline. Pure statistical pipeline; no LLM calls.
+- **Metaculus bot** (`prime_radiant.metaculus`) — LLM forecasting on binary
+  questions (parked at its retrieval phase; client + tests remain green).
+
+## Backtest results (three seasons, vintage-honest)
+
+Rolling-origin backtests over 85 weekly origins across the 2023-24, 2024-25 and
+2025-26 seasons. Every forecast is trained only on the data snapshot a live
+forecaster would have held on the Wednesday submission evening (the hub's git
+history is the vintage store); scored against final truth as of 2026-07-09,
+horizons 0-3, all 53 jurisdictions. `wis_rel` = mean WIS relative to the official
+FluSight-baseline on the common task set across all six models (lower is better;
+the official "scaled relative skill" collapses to this ratio on identical sets).
+Coverage columns are computed on each model's full scored set; relative-skill
+columns on the common set — `n` and `n_relative` in the CSVs disclose both.
+
+**2025-26** — our model leads the table on the natural scale (on the log(x+1)
+scale UMass-flusion edges it, 0.583 vs 0.585 — both columns are in the CSVs):
+
+| model | wis_rel | 50% cov | 95% cov |
+|---|---|---|---|
+| **prime-radiant-lgbm** | **0.609** | 0.398 | 0.818 |
+| UMass-flusion | 0.625 | 0.353 | 0.823 |
+| FluSight-ensemble | 0.666 | 0.526 | 0.903 |
+| prime-radiant-ensemble | 0.764 | 0.464 | 0.880 |
+| FluSight-baseline | 1.000 | 0.433 | 0.864 |
+
+**2024-25** — the multi-model ensembles beat us; we beat the baseline:
+
+| model | wis_rel | 50% cov | 95% cov |
+|---|---|---|---|
+| UMass-flusion | 0.669 | 0.397 | 0.820 |
+| FluSight-ensemble | 0.675 | 0.519 | 0.818 |
+| **prime-radiant-lgbm** | **0.796** | 0.338 | 0.738 |
+| prime-radiant-ensemble | 0.900 | 0.371 | 0.744 |
+| FluSight-baseline | 1.000 | 0.317 | 0.719 |
+
+**2023-24** — flusion dominates; our ensemble edges FluSight-ensemble:
+
+| model | wis_rel | 50% cov | 95% cov |
+|---|---|---|---|
+| UMass-flusion | 0.569 | 0.569 | 0.964 |
+| **prime-radiant-ensemble** | **0.716** | 0.422 | 0.909 |
+| FluSight-ensemble | 0.730 | 0.488 | 0.920 |
+| prime-radiant-lgbm | 0.841 | 0.369 | 0.831 |
+| FluSight-baseline | 1.000 | 0.282 | 0.897 |
+
+Full tables (per-horizon rows, log-scale variants, AE-median, task counts):
+[`reports/backtest_<season>.csv`](reports/). Calibration:
+
+![Calibration curves](reports/calibration.png)
+
+## Honest framing
+
+- **The backtest is net biased against us, and we keep it that way.** Adversarial
+  verification proved that on three 2024-25 holiday weeks the official baseline's
+  run saw data committed Thursday+, which our live-Wednesday vintage discipline
+  refuses; on the 24 information-equal origins our relative WIS improves (lgbm
+  ~0.73). It is not one-way: at three October-2023 origins our vintages were
+  fresher than what the official run used. One 2023-24 origin (2024-04-13) had a
+  week-stale vintage — for our models only; the officials ran on fresh data there.
+- **Our intervals are too narrow.** The lgbm's 50% intervals cover 34-40% and its
+  95% intervals 74-83% — under-dispersed, visible in the calibration curves.
+  FluSight-ensemble is better calibrated even where we beat it on WIS. This is
+  the main modeling debt; season-level bagging (flusion's stabilizer) is the
+  known lever.
+- **Wins and losses both stand.** We lead 2025-26 outright; UMass-flusion beats
+  us clearly in 2023-24 and 2024-25. The 2023-24 lgbm ran on ~1.2 seasons of
+  training history.
+- **The scorer and baseline replica are self-validating — for the season they
+  were validated on.** The replica reproduces the official 2024-25 baseline to
+  relative WIS 0.99999 on fingerprint-matched vintages (Phase B). In 2023-24 the
+  replica scores 0.976 vs the official baseline — a spread-construction
+  divergence at matched anchors that Phase B's validation (2024-25-era official
+  code) does not cover; adversarially checked: substituting the actual official
+  baseline into our ensemble changes no table ordering.
+
+## Methodology in one paragraph
+
+NHSN weekly admissions (hub target data, git-vintaged) → per-100k 4th-root
+transform with per-location scale/center fitted per origin → pooled LightGBM
+quantile regression (23 levels, horizon-as-feature, deterministic, exact-pinned
+4.7.0) → monotone sort, inversion, integer rounding at the hub boundary →
+per-quantile-median ensemble with the baseline replica → WIS/coverage scoring on
+natural and log(x+1) scales, pinned-truth, common-task relative skill. Every
+stage boundary carries a pandera contract; leakage invariants (vintage as-of,
+feature cut, scaler fit) are hypothesis-tested properties.
+
+## Setup
+
+```sh
+uv sync --all-groups
+make check            # offline gates: ruff + format + pyright + pytest-cov
+make test-integration # network: real hub clone, S3 benchmarks, gate + reports
+```
+
+Requires Homebrew `libomp` on macOS for LightGBM. Built with
+[Claude Code](https://claude.com/claude-code) (agentic coding assistant by
+Anthropic); every phase was adversarially verified
+by refuter agents before being declared done — see `NOTES/restart.md` and the
+`[claude]`-prefixed commit history.
+
+## Dashboard
+
+**Live: https://huggingface.co/spaces/jeremygracey-ai/prime-radiant**
+
+A Gradio dashboard (US choropleth of predicted 3-week change, per-state fan
+charts, reliability curves, model-vs-baseline league tables) serves the frozen
+backtest record from `serve_data/`, a ~1.7MB precomputed bundle:
+
+```sh
+make bundle                        # offline: rebuilds serve_data/ deterministically
+uv run python dashboard/app.py     # local: http://127.0.0.1:7860
+```
+
+The Space (`jeremygracey-ai/prime-radiant`, CPU-basic) installs only
+gradio/plotly/pandas/pyarrow — never this package, so no LightGBM/libomp and no
+hub clone at serve time. `.github/workflows/space-deploy.yml` stages and
+validates the Space tree on every dispatch; the actual push is triple-gated
+(manual `deploy=true` + `SPACE_LIVE=1` repo var + `HF_TOKEN` secret) and stays
+inert until go-live.
+
+## Limitations
+
+- Single data source (NHSN), single target (`wk inc flu hosp`); no rate-change /
+  peak / ED-visit targets yet.
+- Interval under-dispersion as above; no season-bagging yet.
+- 2022-23 is not backtestable (the hub's vintage history begins Oct 2023).
+- Live submission machinery exists (Phase E) but is structurally inert: the
+  weekly workflow's live path is triple-gated and unimplemented past its gates,
+  and nothing submits anywhere without the go-live runbook being executed by
+  hand. The dashboard serves a frozen bundle, not a live feed.

prime_radiant-0.1.0/psr_parser.py

@@ -0,0 +1,40 @@
+"""Commit parser for python-semantic-release tolerating `[actor] type(scope):`.
+
+The house convention prefixes agent-authored commits with `[claude] ` before a
+conventional-commits subject. psr's stock ConventionalCommitParser anchors its
+regex at string start, so every prefixed commit fails to parse (verified
+empirically against psr 10.6.2). This subclass strips exactly one leading
+`[word] ` token, then defers to the stock parser.
+
+Loaded by psr via the FILE-PATH form in pyproject
+(`commit_parser = "psr_parser.py:ActorPrefixConventionalParser"`) — the bare
+module form fails because the repo root is not on sys.path when psr imports.
+psr itself never lives in the project env (its tomlkit floor conflicts with
+gradio's cap), so the semantic_release import below only resolves under
+`uvx python-semantic-release==10.6.2` or the release workflow's action; the
+pure strip logic stays importable (and unit-tested) without it.
+"""
+
+import re
+
+_ACTOR_PREFIX = re.compile(r"^\[[\w.-]+\]\s+")
+
+
+def strip_actor_prefix(message: str) -> str:
+    """Remove one leading `[actor] ` token from a commit message, if present."""
+    return _ACTOR_PREFIX.sub("", message, count=1)
+
+
+try:
+    # psr is never in the project env (tomlkit conflict with gradio) — resolved
+    # only under uvx / the release workflow, hence the scoped pyright ignore.
+    from semantic_release.commit_parser.conventional import (  # pyright: ignore[reportMissingImports]
+        ConventionalCommitParser,
+    )
+except ImportError:  # pragma: no cover — project env; psr runs isolated
+    ConventionalCommitParser = None  # type: ignore[assignment, misc]
+else:
+
+    class ActorPrefixConventionalParser(ConventionalCommitParser):
+        def parse_message(self, message: str):  # noqa: ANN201 — psr's own signature
+            return super().parse_message(strip_actor_prefix(message))

prime_radiant-0.1.0/pyproject.toml

@@ -0,0 +1,160 @@
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[project]
+name = "prime-radiant"
+version = "0.1.0"
+description = "Calibrated forecasting: Metaculus LLM bot (Brier) and CDC FluSight quantile forecaster (WIS)."
+readme = "README.md"
+requires-python = ">=3.11"
+license = "MIT"
+authors = [{ name = "Jeremy Gracey", email = "jeremy.a.gracey@gmail.com" }]
+keywords = ["forecasting", "calibration", "flusight", "metaculus", "quantile", "epidemiology"]
+classifiers = [
+    "Development Status :: 3 - Alpha",
+    "Intended Audience :: Science/Research",
+    "License :: OSI Approved :: MIT License",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
+    "Topic :: Scientific/Engineering",
+]
+dependencies = [
+    # Metaculus thread
+    "forecasting-tools>=0.2.92",
+    "pydantic>=2.9",
+    "pydantic-settings>=2.5",
+    "python-dotenv>=1.0",
+    # FluSight epi thread (data layer)
+    "pandas>=2.2",
+    "pandera>=0.20",
+    "epiweeks>=2.3",
+    "httpx>=0.27",
+    "pyarrow>=17",
+    # EXACT pin: LightGBM determinism holds only within one compiled binary —
+    # docs state results differ across versions/compilers/systems.
+    "lightgbm==4.7.0",
+    # <3.12 cap: report PNGs are committed artifacts; hold the rendering series
+    # stable within a phase so regeneration diffs stay meaningful.
+    "matplotlib>=3.11,<3.12",
+    "pyyaml>=6.0.3",
+]
+
+[project.scripts]
+prime-radiant = "prime_radiant.epi.cli:main"
+
+[project.urls]
+Homepage = "https://github.com/JeremyGracey-AI/prime-radiant"
+Repository = "https://github.com/JeremyGracey-AI/prime-radiant"
+
+[dependency-groups]
+dev = [
+    "pytest>=8",
+    "pytest-cov>=5",
+    "hypothesis>=6.100",
+    "pytest-recording>=0.13",
+    # floors match .pre-commit-config.yaml revs so hooks and `uv run` agree
+    "ruff>=0.16.5",
+    "pyright>=1.1.411",
+    "jsonschema>=4.26.0",
+    "types-pyyaml>=6.0.12.20260815",
+]
+# NOTE: python-semantic-release is deliberately NOT a dependency group — its
+# tomlkit>=0.15 floor conflicts with gradio 6.26's tomlkit<0.15 cap. It runs
+# isolated instead: `uvx python-semantic-release==10.6.2` locally, the
+# containerized psr action (same version) in release.yml.
+docs = [
+    "mkdocs-material>=9.7",
+]
+# Dev-only deps for the HF Space dashboard (dashboard/). The Space itself never
+# installs this package — its requirements.txt is generated by space-deploy.yml.
+dashboard = [
+    # ==6.26.*: match the Space's README sdk_version exactly — FetchMerck commit
+    # 8732e99 is the recorded cost of developing against a different Gradio than
+    # the Space runs.
+    "gradio==6.26.*",
+    # <7 cap: plotly 7 flips layout.geo.fitbounds default, removes
+    # figure_factory.create_choropleth and *mapbox traces, and Gradio's bundled
+    # plotly.js has no verified plotly-7 support. Declared directly per the
+    # Phase F brief (transitive-only via forecasting-tools before this).
+    "plotly>=6.9,<7",
+]
+
+[tool.hatch.build.targets.wheel]
+packages = ["src/prime_radiant"]
+
+# Default hatchling sdists ship the entire working tree (NOTES, handoff briefs,
+# serve_data) — scope to the package + release-relevant files. LICENSE rides
+# along via the license-files mechanism.
+[tool.hatch.build.targets.sdist]
+only-include = [
+    "src/prime_radiant",
+    "psr_parser.py",
+    "README.md",
+    "CHANGELOG.md",
+    "CITATION.cff",
+    "uv.lock",
+]
+
+[tool.ruff]
+line-length = 100
+target-version = "py311"
+
+[tool.ruff.lint]
+select = ["E", "F", "W", "I", "UP", "B", "SIM"]
+
+[tool.ruff.lint.isort]
+# dashboard/ modules ship flat to the HF Space root and import as siblings.
+known-first-party = ["prime_radiant", "panel_data", "panel_plots"]
+
+[tool.pyright]
+pythonVersion = "3.11"
+typeCheckingMode = "basic"
+venvPath = "."
+venv = ".venv"
+
+# dashboard/ is not part of the package: it deploys flat to the Space root, so
+# its modules resolve as siblings there and need the same path here and in tests.
+[[tool.pyright.executionEnvironments]]
+root = "dashboard"
+extraPaths = ["dashboard"]
+
+[[tool.pyright.executionEnvironments]]
+root = "tests"
+extraPaths = ["dashboard"]
+
+[tool.semantic_release]
+# Runs isolated: `uvx python-semantic-release==10.6.2` locally, the psr action
+# (same version) in release.yml — see the dependency-group note above.
+version_toml = ["pyproject.toml:project.version"]
+# FILE-PATH form is load-bearing: the bare module form cannot import from cwd.
+commit_parser = "psr_parser.py:ActorPrefixConventionalParser"
+commit_message = "[claude] chore(release): v{version}"
+# House rule: one git identity everywhere, agent commits included.
+commit_author = "Jeremy Gracey <jeremy.a.gracey@gmail.com>"
+# uv.lock carries the project's own version; regenerate AFTER stamping and
+# ship it inside the release commit.
+build_command = "uv lock && uv build"
+assets = ["uv.lock"]
+allow_zero_version = true
+major_on_zero = false
+tag_format = "v{version}"
+
+[tool.semantic_release.changelog]
+mode = "update"
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+markers = [
+    "unit: fast, offline, fixture-backed",
+    "integration: touches network or the real hub clone",
+    "contract: schema/property invariants (offline)",
+    "e2e: full pipeline runs",
+]
+
+[tool.coverage.run]
+source = ["prime_radiant"]
+
+[tool.coverage.report]
+show_missing = true

prime_radiant-0.1.0/src/prime_radiant/__init__.py

@@ -0,0 +1 @@
+"""Prime Radiant: calibrated LLM forecasting bot for Metaculus binary questions."""

prime_radiant-0.1.0/src/prime_radiant/config.py

@@ -0,0 +1,18 @@
+"""Settings loaded from environment / .env — secrets and cost caps.
+
+Cost caps are hard limits enforced by the run loop (Phase C): the bot must
+stop spending when a cap is hit, and every run logs token spend against it.
+"""
+
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+
+class Settings(BaseSettings):
+    model_config = SettingsConfigDict(env_file=".env", env_file_encoding="utf-8", extra="ignore")
+
+    anthropic_api_key: str = ""
+    metaculus_token: str = ""
+    news_api_key: str = ""
+
+    per_question_budget_usd: float = 0.25
+    per_run_budget_usd: float = 2.50

prime_radiant-0.1.0/src/prime_radiant/epi/backtest/__init__.py

@@ -0,0 +1 @@
+"""Rolling-origin backtesting over target-data vintages."""