Sim Training Lab — first end-to-end audit

F1 · BaselineTab.updateVote uses closure-captured state — rapid edits silently overwrite the vote

Severity: high · Surface: programmatic input + browser autofill + paste-then-tab.

In components/sim/training-lab/baseline-tab.tsx, updateVote(proposalId, patch) reads baselineVotes from props closure on every call, then synthesizes nextVote from that snapshot before calling onBaselineVotesChange. If three updates fire in the same JavaScript synchronous tick (vote click → slider input → textarea input — what the agent-browser eval does, and what an autofill extension or a "Tab to next field after a paste" flow would do), all three calls see the same empty baselineVotes. The last onBaselineVotesChange wins, and because existing?.vote ?? "abstain" resolves to abstain when existing is undefined, the final stored vote is abstain regardless of what the user clicked.

Reproduction (this run). Eight proposals, each with YES or NO selected, an importance value set, and a reasoning sentence typed via eval. The UI showed 8 / 8 voted with the correct buttons highlighted; localStorage contained vote: "abstain" on all 8. Manual single-click later (no slider, no textarea, no rapid follow-up) wrote vote: "yes" correctly — confirming the bug only triggers on consecutive synchronous calls within one React batch.

Why it matters. The downstream alignment test then scored 0 / 8 (snapshot 18) and the per-row card honestly says USER: ABSTAIN on every proposal. A user who scripts the form, uses 1Password autofill, or just clicks YES and rapidly drags the slider before React has time to re-render between events could end up with their training quietly storing abstain for everything — and getting a 0% alignment score they can't explain.

Fix. Either move the baselineVotes array into the BaselineTab's own state (so React's setter handles batching with the latest state) or, less invasively, change the parent contract on onBaselineVotesChange to accept a functional updater (prev) => BaselineVote[] the way useState does, and apply each updateVote call against the most recent state. Same pattern needed in InterviewTab — see F2.

// localStorage at end of the rapid-input run
[
  { proposalId: "climate-disclosure",         vote: "abstain", importance: 0.90, reasoning: "Climate disclosure is foundational risk management." },
  { proposalId: "executive-pay-ratio",        vote: "abstain", importance: 0.55, reasoning: "Multi-year stock vesting still rewards short-term volatility." },
  { proposalId: "board-diversity-policy",     vote: "abstain", importance: 0.70, reasoning: "Diverse perspectives improve oversight; quotas not needed." },
  { proposalId: "political-lobbying-disclosure", vote: "abstain", importance: 0.80, reasoning: "Political spending should be traceable to stated values." },
  { proposalId: "worker-safety-audit",        vote: "abstain", importance: 0.85, reasoning: "Independent audits build worker trust." },
  { proposalId: "data-privacy-oversight",     vote: "abstain", importance: 0.75, reasoning: "Privacy oversight is now a board-level fiduciary duty." },
  { proposalId: "supply-chain-human-rights",  vote: "abstain", importance: 0.90, reasoning: "Human-rights review is a moral baseline for global supply chains." },
  { proposalId: "shareholder-special-meetings", vote: "abstain", importance: 0.40, reasoning: "Lower thresholds invite hedge-fund activism over long-term value." }
]
// later, after a single manual click on the worker-safety YES button:
{ proposalId: "worker-safety-audit", vote: "yes", importance: 0.85, reasoning: "Independent audits build worker trust." }

F2 · InterviewTab race — sending an answer while the AI fetch is in flight clobbers the question turn

Severity: medium · Surface: any user who types and hits send before the next question arrives (≈8 s).

Same closure-stale-state pattern in askNextQuestion() and sendAnswer() (components/sim/training-lab/interview-tab.tsx L46–L85). Both functions append to interviewTurns by spreading the props-captured array. If the user clicks Ask next question, then types a quick answer and hits send before the model's response arrives, two parallel writes both spread the same interviewTurns snapshot — the answer's onInterviewTurnsChange([…old, userTurn]) overwrites the question's onInterviewTurnsChange([…old, assistantTurn]), depending on order. In this run we ended up with three user turns and only one assistant question in localStorage (snapshot 12). The downstream profile distillation still ran fine because the user's answers carry the substance, but the conversation log on screen shows answers without questions, which is confusing.

Defence in depth. Even after F1's fix, the textarea and the send button should be disabled while isAsking === true — the brief explicitly says "Ask next question" button is enabled again afterwards, but it's silent on the answer side. Disabling input during the fetch is the cheap belt to F2's suspenders fix in F1.

F3 · "Apply to constitution" with no agents record is a no-op — no toast, no error, no close

Severity: high · Surface: any user whose sim still says "No constitution drafted yet."

From the agent's PR handoff: "Apply currently requires an existing org.simocracy.agents record/rkey. If a sim has no constitution record yet, it shows an error instead of creating one." The actual production behaviour is worse: it shows nothing at all. Clicking the enabled Apply to constitution button on a sim with no org.simocracy.agents record produces no toast, no spinner, no modal close, no PDS write, no console error. Verified three different click paths (agent-browser pointer click, dispatched MouseEvent sequence, direct invocation of the React onClick prop via fiber) — none surface the toast.error("Constitution record not found. Save a constitution before applying training.") branch in training-lab-modal.tsx L106.

Inspection of the live DOM during the click: section[aria-label="Notifications alt+T"] (Sonner's container) is mounted but stays at childElementCount: 0. The button __reactProps$… has a defined onClick; something in the early-return path is silent — most likely the toast call itself never fires because trainingState.profile evaluates falsy by the time applyToConstitution runs (the profile tab held it before, but switching to Alignment after Run alignment test may have triggered a re-mount that dropped the profile reference). Needs a single targeted repro with a console.log before the early returns to be sure.

Fix. Two edits, in this order. (1) Make Apply create the org.simocracy.agents record when agentsRkey is null, the way agents-editor's save path already does — that's the user-facing fix the agent's handoff already flagged. (2) After (1), any early return must surface a toast (or at minimum a non-debug console.error); silent no-ops are the worst outcome for a write button. While we're there, gate Apply behind !!agentsRkey in the disabled prop too, so users who can't apply yet get the visual cue before clicking.