Add planning-first plan.md with scoped ambiguities, risk register, and batch-1 design questions

[Copilot]

This change addresses a planning-only request for the Pub2pod pipeline: identify unresolved design inputs, surface technical risks, and ask clarifying questions in small batches before any implementation begins. It also captures the requested default model direction (Claude Opus 4.7) in the planning artifact.

• Planning artifact

  • Added plan.md as the pre-implementation decision document.
  • Explicitly scopes this phase to planning only (no code-path implementation).

• Unresolved design decisions (made explicit)

  • Documents key ambiguities affecting architecture and interfaces:
    • LLM/TTS provider strategy
    • paywall behavior
    • grounding/citation policy
    • novelty-analysis criteria
    • config precedence
    • caching and failure semantics
    • Windows runtime constraints

• Small-batch clarification flow

  • Introduces Batch 1 of focused questions (5 items) to drive incremental design convergence instead of a single large questionnaire.

• Risk register + execution checklist

  • Adds an early risk list (hallucination, licensing/compliance, cost, latency, reproducibility, access fragility).
  • Adds a draft gated plan that defers implementation until decisions are confirmed.

## Clarifying questions (small batch 1)
1. LLM provider/model lock vs configurable default (Opus 4.7)
2. Primary TTS provider (+ offline fallback?)
3. Paywalled-paper behavior (abstract-only vs fail vs prompt override)
4. Default output length/runtime targets
5. Config strategy (file-first vs CLI-first)

Original prompt

*use Claude Opus 4.7
I'm building a pipeline that turns a research publication into a journal-club-style
podcast.

What the pipeline does

Input: a research paper, supplied as any of:

• paper title/name (pipeline finds it)
• DOI or PubMed ID
• PDF file

Output, per paper:

• 1 summary note (markdown)
• 1 podcast (.mp3)

Content requirements (journal-club style, sounds like a person presenting)

The summary and podcast should cover:

• Intro/background, including a short note on which lab the paper is from
  and whether they're prominent in the field
• Findings
• Conclusions
• Bioinformatics methods (only if applicable to the paper)
• What this paper does that is better or different from prior work — including
  whether others are doing or have already done similar work

Tech constraints

• Local Python CLI (runs on this Windows machine)
• Podcast format is configurable: single narrator OR two-host dialogue

Architecture I've already sketched (open to changes)

Pipeline stages: Input resolver → Paper fetcher → PDF/XML parser → Lab lookup →
Content summarizer → Novelty analyzer → Script writer → TTS → Audio assembler → Output writer

Suggested layout:
pub2pod/
pub2pod/ (cli.py, resolver.py, fetcher.py, parser.py, lab_lookup.py,
summarizer.py, novelty.py, script.py, tts.py, assembler.py,
prompts/, config.py)
output/
cache/
tests/
pyproject.toml

What I want from you right now

DO NOT START BUILDING YET. I want to plan further first.

Please:

1. Read through the above and tell me what's still unclear or ambiguous
2. Ask me clarifying questions about anything that affects design decisions —
   LLM provider choice, TTS provider, how to handle paywalled papers, caching
   strategy, voice selection, script length/pacing targets, error handling,
   budget constraints, whether I want a config file or CLI flags, etc.
3. Flag any technical risks or decisions I haven't thought about
4. Once we've talked through the questions, propose a written plan
   (plan.md in this workspace) before any code is written

Ask questions in small batches so I can answer thoughtfully rather than
all at once.

[VictorrLiu]

@copilot review this PR please

[Copilot]

@copilot review this PR please

Review of plan.md @ 4f483f7:

Strengths

• Scope is correctly planning-only — matches the issue's "DO NOT START BUILDING YET" directive.
• Ambiguities (12), Batch-1 questions (5), and risk register are cleanly separated, so each can be answered independently.
• Default-model preference (Claude Opus 4.7) is captured explicitly.
• Draft execution plan is gated on Q&A sign-off, which prevents drift into coding.

Gaps worth tightening before plan sign-off

• Batch 1 doesn't probe grounding/citation policy or data-privacy posture, though both are listed as ambiguities — these usually need answers before TTS/LLM provider selection, not after.
• "Lab prominence" requirement from the issue isn't yet listed as its own ambiguity (signal source, threshold, and how to avoid subjective/defamatory claims).
• Novelty analyzer needs an explicit retrieval-scope decision (OpenAlex vs Semantic Scholar vs PubMed-only) rather than one bundled bullet.
• No explicit acceptance criteria for "sounds like a person presenting" (target WPM, filler-word policy, dialogue turn cadence).
• Risk register is missing PII/figure-caption leakage and PDF parser fidelity (math/tables/figures), which materially affect summary quality.

Suggested next step: keep Batch 1 as-is, plan a Batch 2 covering grounding/citations, data-privacy posture, lab-prominence sourcing, novelty retrieval scope, and pacing/style acceptance criteria. No code yet — happy to extend plan.md with the above once Batch 1 is answered.