BigBounce is an end-to-end research orchestration stack for six peer-reviewed cosmology papers (P1A, P1B, P2, P3, P4, P5). It consists of three layers:

1. Convex DB — canonical source of truth for paper versions, findings, R-rounds, novelty, notables, external reviews, pods, tasks, caveats.
2. bigbounce-mcp — a local MCP server (stdio transport) that exposes typed tools for the agents (Claude Code, Codex, Cursor) to read and mutate Convex state with guardrails (truth-audit-before-close, novelty cap at N3, no caveat-as-closure).
3. bigbounce.hubify.app — read-only static export that subscribes to Convex via React useQuery. No write paths, no API keys in the bundle.

Everything you see at bigbounce.hubify.app/papers/<slug> reflects the live state of the Convex tables documented below. When a paper version bumps, the site re-renders within seconds.

  .tex source  ──►  pdflatex (revtex4-2, 4 passes)  ──►  paper.pdf
       │                                                      │
       │                                                      ▼
       │                                          site/public/papers/<slug>_v<ver>.pdf
       │                                                      │
       ▼                                                      ▼
  bigbounce-mcp tool  ────►  Convex mutation  ────►  papers + paper_versions
  (e.g. paperVersions:bump)         │                         │
                                    ▼                         ▼
                            findings, notables,        React useQuery
                            externalReviews            (real-time subs)
                                    │                         │
                                    └─────────►   bigbounce.hubify.app
                                                  ├── /              (dashboard)
                                                  ├── /papers/<slug> (detail + dynamic review prompt)
                                                  ├── /contributions (N1–N4 novelty)
                                                  ├── /figures       (auto-updated per .tex)
                                                  ├── /activity      (event feed)
                                                  └── /docs          (you are here)

The deployment URL is https://brilliant-panther-471.convex.cloud (public read-only). Mutations require the deployment access token, which lives in ~/.convex/config.jsonon Houston's machine.

Skills are markdown files with YAML frontmatter. Agents auto-load them on session start. There is ONE canonical science skill stack — ~/.claude/scistack/ — with two namespaces inside it: astrostack/ (bigbounce-specific) and hubstack/ (cross-paper learning-loop, publishing, research, infra). Entries under ~/.claude/skills/are symlinks back into scistack. No project-scoped duplicates; “AstroStack” and “HubStack” are namespaces, not separate stacks.

Every paper carries a papers.novelty field. The Convex schema validator rejects N4 at the type level:

novelty: v.optional(v.union(
  v.literal("N1"),
  v.literal("N2"),
  v.literal("N3"),
  // N4 intentionally excluded — Houston standing directive 2026-06-03
))

The canonical tier definitions:

Each paper passes through three stages on the current version before it can be released:

1. Stage 1 — CCAI same-vendor sweep. Mechanical same-vendor pass. Necessary but never sufficient.
2. Stage 2 — Cross-vendor R-round. Direct vendor calls to Grok-4, GPT-5, Gemini-2.5-Pro, Perplexity-Sonar-Pro. Anthropic excluded per the no-echo-chamber rule. tools/cross_vendor_review_direct.py.
3. Stage 3 — External journal-style review. Houston pastes the dynamically generated prompt into ChatGPT / Gemini / Claude. Results logged via externalReviews:add.

Every finding gets a truth-audit verdict before close. Falsified or stale findings are not treated as "closed via fix" — they are closed-via-falsification with evidence.

The exit criterion for cascaded R-rounds: three consecutive convergent-silence rounds (zero new BLOCKERs / MAJORs).

• No public mutation endpoints. The Convex deployment URL is referenced from the site for read queries only (papers, findings, caveats, activity rollup). Mutations require the deployment access token in ~/.convex/config.json.
• API keys never exposed in build output. Vendor keys (Anthropic, OpenAI, Gemini, Grok, Perplexity) live in local .env.local files only — they never reach the static export. RunPod, HuggingFace, Vercel OIDC keys: same.
• MCP server is stdio-only. bigbounce-mcp runs as a subprocess of Claude Code or another local agent. There is no HTTP listener.
• Site is static. next.config.ts sets output: "export". No server-side routes, no API handlers, no auth flow, no session.
• External-review prompts are read-only. The /papers/<slug> page generates a copy-pasteable text prompt. Houston manually pastes into ChatGPT / Gemini / Claude. No agent invocation happens through the site.
• Schema validator hard gates. N4 novelty is rejected at the Convex validator. closureMethod enum includes text-only-no-real-action as an explicit ⚠️ flag.

• github.com/Hubify-Projects/bigbounce — full source.
• project-context/SSOT/index.md — cross-paper dashboard SSOT.
• project-context/SSOT/paper-N/status.md — per-paper close-the-gap.
• project-context/peer-reviews/findings-archive/— every R-round's raw findings.
• project-context/review-patterns/ — extracted pattern catalog (36 patterns and counting).
• ~/.claude/scistack/ — master skill stack (hubstack + astrostack + extensions).