proofpath 0.0.1__tar.gz

proofpath-0.0.1/.editorconfig

@@ -0,0 +1,15 @@
+root = true
+
+[*]
+charset = utf-8
+end_of_line = lf
+insert_final_newline = true
+trim_trailing_whitespace = true
+indent_style = space
+indent_size = 4
+
+[*.{yml,yaml,json,toml}]
+indent_size = 2
+
+[*.md]
+trim_trailing_whitespace = false

proofpath-0.0.1/.github/ISSUE_TEMPLATE/bug_report.yml

@@ -0,0 +1,31 @@
+name: Bug report
+description: Something behaves differently than documented
+labels: ["bug"]
+body:
+  - type: textarea
+    id: what
+    attributes:
+      label: What happened
+      description: Include the exact command and the output.
+    validations:
+      required: true
+  - type: textarea
+    id: expected
+    attributes:
+      label: What you expected
+    validations:
+      required: true
+  - type: input
+    id: version
+    attributes:
+      label: proofpath version
+      placeholder: "output of: proofpath --version"
+    validations:
+      required: true
+  - type: dropdown
+    id: os
+    attributes:
+      label: Operating system
+      options: [Linux, macOS, Windows]
+    validations:
+      required: true

proofpath-0.0.1/.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1 @@
+blank_issues_enabled: true

proofpath-0.0.1/.github/ISSUE_TEMPLATE/feature_request.yml

@@ -0,0 +1,23 @@
+name: Feature request
+description: Suggest a capability
+labels: ["enhancement"]
+body:
+  - type: textarea
+    id: problem
+    attributes:
+      label: The problem
+      description: What are you unable to do today?
+    validations:
+      required: true
+  - type: textarea
+    id: proposal
+    attributes:
+      label: What you propose
+    validations:
+      required: true
+  - type: checkboxes
+    id: spec
+    attributes:
+      label: Design check
+      options:
+        - label: I have read the design spec and this is not already covered or explicitly ruled out

proofpath-0.0.1/.github/ISSUE_TEMPLATE/wrong_verdict.yml

@@ -0,0 +1,39 @@
+name: Wrong verdict
+description: A claim was judged incorrectly
+labels: ["accuracy"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Accuracy reports are the most valuable issues this project receives.
+        They become regression cases in the evaluation set.
+  - type: textarea
+    id: claim
+    attributes:
+      label: The claim
+    validations:
+      required: true
+  - type: input
+    id: source
+    attributes:
+      label: The cited source (DOI or URL)
+    validations:
+      required: true
+  - type: dropdown
+    id: reported
+    attributes:
+      label: What proofpath reported
+      options: [SUPPORTED, REFUTED, NEI, GHOST REFERENCE, AMBIGUOUS, RETRACTED, UNVERIFIED]
+    validations:
+      required: true
+  - type: dropdown
+    id: correct
+    attributes:
+      label: What it should have reported
+      options: [SUPPORTED, REFUTED, NEI, GHOST REFERENCE, AMBIGUOUS, RETRACTED, UNVERIFIED]
+    validations:
+      required: true
+  - type: textarea
+    id: passage
+    attributes:
+      label: The passage that settles it

proofpath-0.0.1/.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,16 @@
+## What this changes
+
+<!-- One or two sentences. -->
+
+## Why
+
+<!-- Link the issue, or explain the problem this solves. -->
+
+## Checklist
+
+- [ ] `ruff check` and `ruff format --check` are clean
+- [ ] `pytest` passes
+- [ ] Type annotations on every new function signature
+- [ ] No network access in unit tests
+- [ ] If a new reported state was added, spec section 15 is updated too
+- [ ] This change does not make the tool sound more confident than its evidence

proofpath-0.0.1/.github/workflows/ci.yml

@@ -0,0 +1,41 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  lint:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: astral-sh/setup-uv@v5
+        with:
+          enable-cache: true
+      - run: uv sync --extra dev
+      - run: uv run ruff check
+      - run: uv run ruff format --check
+
+  test:
+    # The tool promises identical behaviour on all three platforms, so all three
+    # are tested on every change.
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-latest]
+        python-version: ["3.10", "3.13"]
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: astral-sh/setup-uv@v5
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - run: uv sync --extra dev
+      - run: uv run pytest

proofpath-0.0.1/.github/workflows/release.yml

@@ -0,0 +1,35 @@
+name: Release
+
+# Publishes to PyPI via Trusted Publishing (OIDC). No API token is stored
+# anywhere in this repository or its secrets.
+on:
+  push:
+    tags: ["v*"]
+  workflow_dispatch:
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: astral-sh/setup-uv@v5
+        with:
+          enable-cache: true
+      - run: uv build
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist
+          path: dist/
+
+  publish:
+    needs: build
+    runs-on: ubuntu-latest
+    environment: pypi
+    permissions:
+      id-token: write   # required for Trusted Publishing
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: dist
+          path: dist/
+      - uses: pypa/gh-action-pypi-publish@release/v1

proofpath-0.0.1/.gitignore

@@ -0,0 +1,33 @@
+# Python
+__pycache__/
+*.py[cod]
+.venv/
+venv/
+*.egg-info/
+dist/
+build/
+.pytest_cache/
+.ruff_cache/
+.mypy_cache/
+
+# Secrets
+.env
+.env.*
+!.env.example
+*.pem
+*.key
+secrets/
+config.local.*
+
+# proofpath local state
+.proofpath/
+*.sqlite
+*.sqlite3
+data/cache/
+models/
+
+# OS / editors
+.DS_Store
+Thumbs.db
+.idea/
+.vscode/

proofpath-0.0.1/CHANGELOG.md

@@ -0,0 +1,23 @@
+# Changelog
+
+All notable changes to this project are documented here. 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).
+
+## [Unreleased]
+
+## [0.0.1] - 2026-09-10
+
+First release. The verification pipeline is not implemented; this reserves the name
+and establishes the interface, packaging and CI that later phases build on.
+
+### Added
+- Design specification with measured source-access data (spec §6), the fetch ladder
+  and its permission model (§7), and corrected reference resolution (§8).
+- Phased implementation plan, ordered by risk retired rather than user-visible
+  progress.
+- `proofpath` command. A bare invocation is a first-class entry point rather than a
+  help screen, which is where the TUI will attach. Exit codes are fixed: `0` clean,
+  `1` findings, `2` the run itself failed.
+- Cross-platform CI on Linux, macOS and Windows, and PyPI publishing through trusted
+  publishing rather than a stored API token.

proofpath-0.0.1/CLAUDE.md

@@ -0,0 +1,52 @@
+# proofpath — project instructions
+
+`proofpath` verifies that the sources behind a claim actually say what the claim
+says. Read `docs/superpowers/specs/2026-09-10-proofpath-design.md` before making
+design decisions; it is the source of truth and records *why* things are the way
+they are.
+
+## Non-negotiable product rules
+
+These are not style preferences. Breaking one breaks the product's reason to exist.
+
+1. **Never assert without a passage.** A `SUPPORTED` or `REFUTED` verdict must carry
+   the quoted evidence it is based on. No passage, no verdict.
+2. **Never present absence of evidence as evidence of absence.** Unreachable,
+   blocked, paywalled and ambiguous are distinct reported states (spec §15), never
+   silently collapsed into a verdict.
+3. **Never call a real reference a ghost.** Uncertainty resolves to `AMBIGUOUS`, not
+   `GHOST REFERENCE` (spec §8). The false-ghost rate is a release gate.
+4. **Never prompt without a TTY.** In CI or when piped, an `ask` permission is
+   treated as `deny` and reported (spec §7.1).
+5. **Never install anything large without explicit consent** (spec §7.1).
+6. **Every report states its own coverage.** A low-coverage run must not look like a
+   clean one.
+
+## Stack
+
+| Concern | Choice | Why |
+|---|---|---|
+| CLI / TUI | `typer` + `textual` | bare `proofpath` opens the TUI via `invoke_without_command=True` |
+| HTTP | `httpx`, then `curl_cffi`, then browser on consent | fetch ladder, spec §7 |
+| PDF | `pymupdf` | pure Python, no Java/Docker |
+| Vector store | `sqlite-vec` | per-document corpora are small; no server |
+| Embeddings / NLI | ONNX runtime by default, `torch` only via the `[gpu]` extra | keeps base install small |
+| Config / cache paths | `platformdirs` | Windows/macOS/Linux parity |
+
+## Conventions
+
+- Python 3.10+. **Type annotations are mandatory** on every function signature.
+- `ruff check` and `ruff format --check` must be clean.
+- `pytest` must pass on Linux, macOS and Windows.
+- Device preference order is always **CUDA → MPS → CPU** (and the ONNX equivalent:
+  CUDA → CoreML → CPU).
+- `pathlib` everywhere. Every file read/write passes `encoding="utf-8"` explicitly.
+- Code, identifiers and comments in English.
+
+## Hard constraints
+
+- **No Docker or Java requirement** in any default path. GROBID stays opt-in.
+- **No hardcoded secrets, tokens or emails.** The Crossref/OpenAlex contact address
+  comes from config or `PROOFPATH_CONTACT_EMAIL` and is optional.
+- No network access in unit tests. Provider responses are fixtures.
+- `cli` and `tui` contain no logic; both call the same `verify()` entry point.

proofpath-0.0.1/CONTRIBUTING.md

@@ -0,0 +1,34 @@
+# Contributing
+
+Thanks for considering a contribution.
+
+## Before you start
+
+Read `docs/superpowers/specs/2026-09-10-proofpath-design.md`. It records the design
+decisions and, more importantly, the measurements behind them. Several obvious-looking
+approaches were tested and rejected — the spec says which and why.
+
+For anything beyond a small fix, open an issue first so we can agree on the approach.
+
+## Setup
+
+```bash
+uv sync --all-extras
+uv run pytest
+uv run ruff check
+uv run ruff format --check
+```
+
+## Standards
+
+- Type annotations on every function signature.
+- `ruff check` and `ruff format --check` clean.
+- Tests pass on Linux, macOS and Windows.
+- No network access in unit tests — use fixtures.
+- New reported states go in the table in spec §15, not just in code.
+
+## The rules that matter most
+
+A pull request that makes the tool sound more confident than the evidence supports
+will be rejected, however good the code is. See the non-negotiable rules in
+`CLAUDE.md`.

proofpath-0.0.1/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Yigit Erdogan
+
+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.

proofpath-0.0.1/PKG-INFO

@@ -0,0 +1,112 @@
+Metadata-Version: 2.5
+Name: proofpath
+Version: 0.0.1
+Summary: Check whether the sources behind a claim actually say what the claim says.
+Project-URL: Homepage, https://github.com/Yigtwxx/proofpath
+Project-URL: Repository, https://github.com/Yigtwxx/proofpath
+Project-URL: Issues, https://github.com/Yigtwxx/proofpath/issues
+Project-URL: Changelog, https://github.com/Yigtwxx/proofpath/blob/main/CHANGELOG.md
+Author: Yigit Erdogan
+License: MIT License
+        
+        Copyright (c) 2026 Yigit Erdogan
+        
+        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.
+License-File: LICENSE
+Keywords: citation,cli,entailment,fact-checking,nli,rag,research-integrity,retrieval-augmented-generation
+Classifier: Development Status :: 2 - Pre-Alpha
+Classifier: Environment :: Console
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering
+Classifier: Topic :: Text Processing :: Linguistic
+Classifier: Typing :: Typed
+Requires-Python: >=3.10
+Requires-Dist: curl-cffi>=0.16.1
+Requires-Dist: fastembed>=0.5
+Requires-Dist: httpx>=0.28
+Requires-Dist: orjson>=3.11
+Requires-Dist: platformdirs>=4.3
+Requires-Dist: pymupdf>=1.25
+Requires-Dist: python-docx>=1.1
+Requires-Dist: scrapling>=0.4.15
+Requires-Dist: sqlite-vec>=0.1.6
+Requires-Dist: textual>=1.0
+Requires-Dist: typer>=0.15
+Provides-Extra: browser
+Requires-Dist: scrapling[fetchers]>=0.4.15; extra == 'browser'
+Provides-Extra: dev
+Requires-Dist: mypy>=1.14; extra == 'dev'
+Requires-Dist: pytest-cov>=6.0; extra == 'dev'
+Requires-Dist: pytest>=8.3; extra == 'dev'
+Requires-Dist: respx>=0.22; extra == 'dev'
+Requires-Dist: ruff>=0.9; extra == 'dev'
+Provides-Extra: gpu
+Requires-Dist: sentence-transformers>=3.3; extra == 'gpu'
+Requires-Dist: torch>=2.5; extra == 'gpu'
+Description-Content-Type: text/markdown
+
+# proofpath
+
+> Don't guess. Show the evidence.
+
+`proofpath` checks whether the sources behind a claim actually say what the claim
+says. Point it at a paper, a draft, or a link — it verifies every citation on three
+levels and shows you the passage behind each verdict.
+
+1. **Does the source exist?** — resolved against Crossref, OpenAlex, PubMed, arXiv.
+2. **Is it still valid?** — checked against Retraction Watch.
+3. **Does it support the claim?** — retrieval + entailment against the source text.
+
+Most tools stop at step 1. Step 3 is the point.
+
+Runs offline and free by default: no API key, no Docker, no server. Windows, Linux
+and macOS.
+
+## What it will not do
+
+It does not pretend to be certain. Full text is openly available for well under half
+of published citations, so many verdicts will honestly be *not enough information*,
+and every report ends with its own coverage figures. A tool that sounds sure about
+everything is the problem this one exists to fight.
+
+```bash
+uv tool install proofpath
+
+proofpath                        # interactive TUI
+proofpath check paper.pdf        # one-shot report
+proofpath check draft.md --format sarif
+```
+
+## Status
+
+🚧 Design stage — no working code yet.
+
+- [Design specification](docs/superpowers/specs/2026-09-10-proofpath-design.md) — what it does and the measurements behind each decision
+- [Implementation plan](docs/superpowers/plans/2026-09-10-proofpath-implementation-plan.md) — phases, ordered by risk retired
+- [Open items](docs/superpowers/OPEN-ITEMS.md) — what is unresolved, and what has not been verified yet
+
+## License
+
+MIT

proofpath-0.0.1/README.md

@@ -0,0 +1,43 @@
+# proofpath
+
+> Don't guess. Show the evidence.
+
+`proofpath` checks whether the sources behind a claim actually say what the claim
+says. Point it at a paper, a draft, or a link — it verifies every citation on three
+levels and shows you the passage behind each verdict.
+
+1. **Does the source exist?** — resolved against Crossref, OpenAlex, PubMed, arXiv.
+2. **Is it still valid?** — checked against Retraction Watch.
+3. **Does it support the claim?** — retrieval + entailment against the source text.
+
+Most tools stop at step 1. Step 3 is the point.
+
+Runs offline and free by default: no API key, no Docker, no server. Windows, Linux
+and macOS.
+
+## What it will not do
+
+It does not pretend to be certain. Full text is openly available for well under half
+of published citations, so many verdicts will honestly be *not enough information*,
+and every report ends with its own coverage figures. A tool that sounds sure about
+everything is the problem this one exists to fight.
+
+```bash
+uv tool install proofpath
+
+proofpath                        # interactive TUI
+proofpath check paper.pdf        # one-shot report
+proofpath check draft.md --format sarif
+```
+
+## Status
+
+🚧 Design stage — no working code yet.
+
+- [Design specification](docs/superpowers/specs/2026-09-10-proofpath-design.md) — what it does and the measurements behind each decision
+- [Implementation plan](docs/superpowers/plans/2026-09-10-proofpath-implementation-plan.md) — phases, ordered by risk retired
+- [Open items](docs/superpowers/OPEN-ITEMS.md) — what is unresolved, and what has not been verified yet
+
+## License
+
+MIT

proofpath-0.0.1/SECURITY.md

@@ -0,0 +1,24 @@
+# Security Policy
+
+## Reporting a vulnerability
+
+Please **do not** open a public issue for a security problem.
+
+Use GitHub's private vulnerability reporting on this repository
+(Security → Report a vulnerability). You should get an initial response within a
+week.
+
+## Scope
+
+`proofpath` fetches remote content and can, with explicit consent, download and run a
+browser engine. Reports about the following are especially welcome:
+
+- Fetched content escaping its parsing boundary.
+- The permission prompt being bypassed, or a large install happening without consent.
+- Secrets or the configured contact address leaking into reports, logs or caches.
+- Path traversal through a crafted document or archive.
+
+## Out of scope
+
+- A wrong verdict on a claim. That is an accuracy issue — open a normal issue.
+- Rate limiting or availability of third-party APIs.

proofpath-0.0.1/docs/superpowers/OPEN-ITEMS.md

@@ -0,0 +1,120 @@
+# Open items
+
+Everything left unresolved as of **2026-09-10**, written down so none of it has to be
+reconstructed from memory. Each entry says enough to be picked up cold.
+
+Spec: `specs/2026-09-10-proofpath-design.md` · Plan: `plans/2026-09-10-proofpath-implementation-plan.md`
+
+---
+
+## 1. Blocked — needs an action outside this repo
+
+| # | Item | What to do |
+|---|---|---|
+| 1.1 | **CI workflow is not in the repo.** The file is written and working locally but the GitHub token lacks the `workflow` scope, so the push was rejected. | Run `gh auth refresh -h github.com -s workflow`, complete the device flow, then commit `.github/workflows/ci.yml` on its own. The file already exists on disk, untracked. |
+| 1.2 | PyPI name `proofpath` is free but **not reserved**. | Publish `0.0.1` once there is a working skeleton. Publishing an empty placeholder is against PyPI policy, so this waits for Phase 0 to finish. |
+
+---
+
+## 2. Decisions taken but not confirmed by the author
+
+Recorded with the recommendation that was made, so they can be accepted or reversed
+deliberately rather than by default.
+
+| # | Decision | Recommended | Why it might change |
+|---|---|---|---|
+| 2.1 | `install_browser` default (spec §7.1) | `ask` | `deny` is safer but hides from the user that a blocked source was recoverable |
+| 2.2 | Reddit in v0.4 | include, optional | needs a user-registered OAuth app; Bluesky and HN work without one, so Reddit is never blocking |
+| 2.3 | `spiyweb` keeps its name | yes | the spider-web-as-graph metaphor fits; every clean alternative on PyPI was worse |
+| 2.4 | `reasonhound` keeps its name | yes | already a good name |
+| 2.5 | Turkish sources | after v0.4 | OpenAlex/Crossref coverage is much weaker for Turkish; half-supporting it would damage trust |
+| 2.6 | `--summarize` in the TUI | off by default there too | keeping "offline unless asked" true in both front-ends |
+
+---
+
+## 3. Unvalidated assumptions — test before building on them
+
+These were asserted during design without being checked. Each one, if wrong, changes
+a phase.
+
+| # | Assumption | How to check | If wrong |
+|---|---|---|---|
+| 3.1 | **A usable NLI cross-encoder exists as ONNX.** The whole "no torch in the base install" decision rests on this and it was never verified. | Search the ONNX model zoo / HF for a DeBERTa-MNLI or similar exported model; measure it on SciFact dev | either export one with `optimum`, or make `[gpu]`/torch the default and accept the install size |
+| 3.2 | `sqlite-vec` ships working wheels for Windows and macOS arm64 | install on all three in CI during Phase 0 | fall back to numpy brute force; per-document corpora are small enough |
+| 3.3 | `textual` renders the §13.1 layout correctly in Windows Terminal at 80 columns | render a fixture screen in CI on Windows | simplify the box drawing to ASCII |
+| 3.4 | SciFact, AVeriTeC and PubHealth are still downloadable and pinnable | fetch each once, record the revision | substitute a comparable set and note it in the spec |
+| 3.5 | X Community Notes dumps are still published and parseable | download one day's file, inspect the columns | drop the X path entirely; Bluesky and HN already carry the social provider |
+| 3.6 | `fastembed` model quality is sufficient for passage ranking | compare recall@k against a sentence-transformers baseline in Phase 1 | move embeddings to the `[gpu]` extra as well |
+
+---
+
+## 4. Design details deliberately left to implementation
+
+Not oversights — they need real data to set, and guessing now would be false
+precision. Each must end up written into the spec once measured.
+
+| # | Detail | Set during |
+|---|---|---|
+| 4.1 | Title token-set Jaccard threshold for reference matching (spec §8) | Phase 3, tuned on the ghost test set |
+| 4.2 | Year tolerance beyond ±1 for online-first publications | Phase 3 |
+| 4.3 | Numeric comparison tolerance, and how to treat ranges vs point values (spec §10) | Phase 2 |
+| 4.4 | Confidence threshold that routes a verdict to the judge | Phase 1 calibration |
+| 4.5 | Chunk size and `k` for retrieval. Abstracts are trivial; a 12,000-word full text is not, and the two may need different settings | Phase 1, re-checked in Phase 4 |
+| 4.6 | Which embedding model and which NLI model, by name and revision | Phase 1 |
+
+---
+
+## 5. Genuinely unsolved problems
+
+Known gaps with no chosen answer yet. These are the ones worth thinking about away
+from the keyboard.
+
+**5.1 — A citation that supports a paragraph, not a sentence.**
+Spec §5.1 mentions a paragraph-level fallback but does not define when it triggers.
+Attaching a paragraph-wide claim to one sentence produces confident nonsense. Needs a
+rule, and probably a distinct reported state.
+
+**5.2 — Author-year citation styles.**
+`(Smith et al., 2020)` is much harder than `[12]`: the marker does not index the
+bibliography directly, several works share an author-year, and `ibid.`/`op. cit.`
+exist. Phase 5 currently treats both styles as one task; they are not.
+
+**5.3 — Coverage is the real product ceiling.**
+Measured: direct full text for well under half of sampled citations (spec §6.1). Every
+abstract-only verdict is a weak verdict. Worth investigating whether Semantic Scholar,
+CORE, or OpenAIRE meaningfully raise this before accepting the number.
+
+**5.4 — What "confidence" means to a user.**
+A `0.91` is a model score, not a probability of being right. Showing it as-is invites
+misreading, hiding it removes signal. No decision yet on how to present it.
+
+**5.5 — Caching fetched full text.**
+Spec §16 says cached locally, not redistributed, clearable. Retention period and
+whether publisher content should be cached at all are unresolved.
+
+---
+
+## 6. Deferred by choice
+
+Not problems, just not now. Recorded so they are not rediscovered as new ideas.
+
+| Item | Revisit when |
+|---|---|
+| Landing site: one Astro site, four pages (`evidencelab.dev/`, `/proofpath`, `/reasonhound`, `/spiyweb`) | after v0.1 ships something runnable |
+| GitHub org `evidencelab` (free at time of checking) | when the landing site is built |
+| `evidencelab.dev` domain (free at time of checking) | same |
+| `spiyweb` as an alternative retrieval backend for `proofpath` | after Phase 6; the real tie between the three projects |
+| GROBID parser as an opt-in `--parser` | if Phase 5 citation pairing accuracy proves inadequate |
+| Turkish sources | after v0.4 |
+
+---
+
+## 7. Next session starts here
+
+1. Finish item **1.1** — CI is a prerequisite for every "works on three platforms" claim.
+2. Check assumption **3.1** — if no usable ONNX NLI model exists, the packaging story
+   changes and it is better to know before Phase 1 than during it.
+3. Begin **Phase 1**: pin SciFact, build `scripts/eval.py`, get a first number.
+
+Phase 1 carries a kill criterion on purpose. Getting to a real number quickly is the
+point of the whole ordering.