# Risk Register **Living document. Solo-owner release-decision aid, not compliance theater.** The consolidated threat table for the project — likelihood × impact, mitigation, residual, owner (always 0thernes). Ad-hoc risks were previously scattered across AUDIT-LOG entries and inspection WARNs; this is the one place. Newest material risks first. Likelihood/impact: **L** low · **M** med · **H** high. Residual = after mitigation. | # | Risk | L | I | Mitigation (in place) | Residual | Status | | --- | ------------------------------------------------------------------------------------------------------------------------------------------------ | --- | --- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------- | ------------------------------------------------------------ | | R1 | **Faculty coupling stays weak** — `meanAbsCoupling ≈ 0.18`; the "emergence = coupling > count" thesis is the product's core bet and is unproven. | H | H | GWT bind-gate lifted baseline 0.158→0.197; measured every run by `coupling-audit.ts`; [NHSI-PROGRESS-DASHBOARD](./NHSI-PROGRESS-DASHBOARD-2026-06-26.md) tracks it. | M–H | **Open — top research risk** | | R2 | **No external validation** — no third party has reproduced any claim; the self-scored A-Life row is the biggest comparative uncertainty. | H | M | **Repro artifact shipped 2026-07-02:** [`scripts/reproduce.ts`](../scripts/reproduce.ts) — one command replicates the determinism fingerprint + gate-band coupling on any machine; matrix source-cited. | M | **Open** — artifact ships; a citable third-party run remains | | R3 | **Overclaim drift** — consciousness/AGI/quantum language creeps past what's wired. | M | H | Adversarial honesty audits; `verify:facts` + sync gates; Butlin framing pinned "8/14 + 6/14, NOT sentient"; SSOT receipt blind-spot closed 2026-07-01. | L | Controlled | | R4 | **Frame-budget ceiling** — `5× think()` ≈ 9.77 ms (~58% of a frame); GOAL5 `<2%` target unmet. | M | M | Staggered `driveSuper` cadence + 20 cheap light-echoes; `dt`≤0.05 clamp; RenderGovernor FX-shed. See [BENCHMARKS](./BENCHMARKS-2026-06-26.md). | M | Managed, remediation open | | R5 | **GPU VRAM leak recurrence** — per-instance `THREE.*` outside cache without `dispose()`. | M | M | `World.dispose()` sweep + spy-verified tests; recurring-class memory. | L | Controlled (recurs — watch) | | R6 | **Local↔GitHub drift** — primary checkout silently falls behind while worktrees push. | M | L | `sync-guard.ts` (predev) + `autocrlf=input` + post-commit autopush; `git reset --hard origin/main` recovery. | L | Controlled | | R7 | **Tsotchke corpus availability / license** — 3 repos fenced (LLM/proprietary); upstream is external. | L | M | Real MIT ports vendored source-level + credited; fenced set documented; never called fake. | L | Controlled | | R8 | **Deploy exposure** — static Pages build could leak a secret or serve a broken bundle. | L | H | No frontend secrets (public/private boundary); keyless LLM chat; golden-hash + smoke in CI. See [SECURITY-ARCHITECTURE](./SECURITY-ARCHITECTURE-2026-07-02.md). | L | Controlled | | R9 | **AI-sandbox abuse** — user-driven code/prompt paths (copilot, sandbox) could DoS or escape. | L | M | Recursion/allocation caps; secret-leak grep bypass closed 2026-07-01; no eval of untrusted code in sim. | L | Controlled (verify each release) | | R10 | **Doc sprawl / staleness** — 54+ docs; a receipt or claim drifts unseen. | M | L | Living-docs law + `sync-surfaces` + `verify:facts`; the reports blind-spot was closed. | L | Controlled | ## Release gate on risk Do not tag a release with any **H residual** risk unmitigated and undocumented. R1 and R2 are the standing high-residual items; they are research/validation risks, not defects, and are acceptable for a research artifact **as long as they are stated honestly** on every public surface (which they are).