vizier-trading-skill

What Vizier is (and isn't)

• One skill, natural-language-driven — there is no /analyze vs /invest. Behavior is read from intent: an empty/vague call → a read-only market sweep; "research MU" → a thesis, no execution; "buy $3 of AAPL" → an order it just executes. It pauses to ask only on real ambiguity ("invest a little"), never double-confirms the obvious, and treats "I think I should sell" as deliberation, not an order.
• Multi-horizon — every analysis yields a long-term read (quality/fundamentals) and a short-term read (technicals/catalyst/flow), and presents both; divergence is a feature. Each thesis/position carries a core vs tactical tag so anti-churn applies correctly per horizon.
• Multi-venue — US equities/ETFs (the ibkr server) and crypto spot (the crypto server), routed by the asset; risk limits and NAV are kept strictly per-account / per-venue.
• Manager / breadth discovery — a broad request ("analyze the market and bring me recommendations") makes Vizier the manager of a research team: it partitions the market into coverage areas, fans out research-only envoy subagents across them in parallel, then prunes the funnel by potential, risk/reward and correlation-based diversification — so the shortlist spans the market instead of three of the same bet. The envoys can only read (Scout); only the main thread ever executes. Note the cost split: a plain "what's going on in the market?" is the cheap, fixed, read-only sweep, while "bring me recommendations" triggers the expensive multi-agent fan-out (N parallel envoys) — Vizier never auto-escalates the cheap call into the expensive one. Breadth defaults to a US equities/ETFs + crypto spot universe (risk kept per venue).
• Confirmation by default; autonomy strictly opt-in — a complete, explicit order (asset + amount, e.g. "buy $50 of AAPL") is itself the confirmation: it executes after the safety gates without an extra prompt. "Confirmation by default" is what governs under-specified or skill-derived trades ("invest a little", a thesis Vizier proposed) — there it shows its reasoning and waits for your OK before any live order. Autonomy is a separate, conscious choice with hard, code-enforced prerequisites.

How it thinks

Vizier is a hybrid: the judgment lives in the skill (SKILL.md + references/, loaded on demand), while every money-sensitive calculation lives in a deterministic Python core the skill calls and reads back as a {"ok": bool, "data": ...} envelope. Code does the math that must not be re-derived by an LLM "in its head" between rounds — that is exactly the forget-between-rounds failure the split prevents.

flowchart LR
    U["User request<br/>(natural language)"] --> V["VIZIER skill<br/>(judgment, orchestration)"]
    V -->|reads data| S["Scout MCP<br/>(senses)"]
    V -->|money-math, state| C["Deterministic core<br/>python -m vizier"]
    V -->|places orders| X["Valet MCP<br/>(hands: ibkr / crypto)"]
    C --> M["Private memory<br/>theses · decision log · NAV"]

The reasoning runs as a fan-out of subagents, adapting depth to the request:

flowchart LR
    A["Analysts (parallel)<br/>Fundamental · Technical · News · Macro"] --> BB["Bull x Bear<br/>debate"]
    BB --> T["Trader<br/>proposes thesis + trade<br/>(horizon tag, conviction 1-5)"]
    T --> DS["Data-sufficiency gate<br/>proceed | downsize | abstain"]
    DS --> R["Risk gate + PM<br/>sizing within limits"]
    R --> PM["Pre-mortem / red-team<br/>why is THIS trade, NOW, wrong?"]
    PM --> E["Execute via Valet"]

Research mode stops at "Trader proposes"; execution mode continues through the gates to Valet, then journals the thesis and the fills.

MODE: PAPER · venue: ibkr

TL;DR — Buy $60 NVDA (core) + $40 XLE (tactical); breaker clear, book has room in both.

By horizon
  • Long  — NVDA: datacenter capex still accelerating; quality metrics intact
  • Short — NVDA: RSI 71, extended — divergence flagged: scale-in, don't chase

Action
  • NVDA — ibkr, BUY $60, market (liquid large-cap)   [EXECUTED — order id 1247]
  • XLE  — ibkr, BUY $40, limit 92.10                 [EXECUTED — order id 1248]

Conviction — NVDA 4/5 (secular + earnings momentum) · XLE 3/5 (honest: a hedge, not a hero)

Scenarios + price target (NVDA)
  • Bear $95 · Base $128 (analyst consensus relay, n=42) · Bull $150
  (crude multiple cross-check flagged as crude — no independent valuation claimed)

Risks & caveats — news feed may be incomplete for corporate events (8-K check run: clean);
  data-sufficiency on XLE: thin positioning data — annotated, amount honored

Pre-mortem — buying an extended name near an earnings window (checked: 24 days out)

Post-trade portfolio — NVDA 8.2% · XLE 5.5% · cash 22% · 6 positions

Sources — SEC filings (as_of 2026-05-28), FRED, IBKR quotes 14:31 ET

The full format (mandatory MODE banner, per-name [RECOMMENDATION]/[EXECUTED] tags, honest-caveat rules) is specified in references/output-template.md.

Measuring edge — the scorecard

"Is this actually working?" is answered by a deterministic scorecard, not by vibes. The skill feeds it current prices and a benchmark price history (SPY for equities, BTC/USDT for crypto) and the core computes, over every thesis (open marked-to-market + closed on recorded outcomes):

• P&L and period return per thesis, with days_held reported so a 3-day tactical and a 6-month core are weighed by the reader — no annualization (annualizing both into one number is nonsense).
• Hit rate and win/loss profile over closed theses; aggregates overall, per horizon and per venue.
• Alpha vs the benchmark over the SAME window as each thesis — the number that separates "the brain has edge" from "it just rode the market".

The honesty rules are the point: a thesis that can't be scored is named in skipped (never guessed), an aggregate with no sample is null, not a fake zero, and a benchmark that doesn't cover the window yields a null or annotated alpha with the reason attached. If the ideas are not beating the benchmark, the scorecard says exactly that — that verdict is what the paper-first ladder exists to produce.

The safety model (code-enforced, fail-closed)

The money-safety rules are not prose the model is asked to remember — they are deterministic checks with exact, self-latching arithmetic. A precise word on bindingness: these checks are advisory/bookkeeping that bind when the skill calls them each candidate with an honest live NAV — Vizier holds no order pipe of its own, so the one hard, code-enforced dollar backstop is the Valet's MAX_DAILY_VALUE (enforced at the executor no matter what the skill does), which is why arming autonomy is forbidden until it is set. The dangerous mode (autonomy) is guarded by four composed legs plus a re-arm guard — a backstop for the robot running unattended, never a clamp on a human's confirmed explicit order:

• Cumulative daily ceiling — at most a configured % of the fixed start-of-day NAV (50% on the aggressive profile) can be deployed across a rolling 24h window, sourced from the persistent decision log so it survives restarts and /loops. A shrinking account never re-authorizes fresh slices — this is the drain fix (a per-round cap alone does not stop a loop from emptying the account).
• Per-run ceiling — a single autonomous round is capped at a smaller % (33%) or a max trade count, anchored to the same fixed baseline; it resets each round.
• Drawdown kill — if NAV falls past a kill threshold from the day baseline, the gate latches the kill and hard-blocks every further candidate (even if NAV recovers); the skill then disarms and a deliberate manual re-arm is required. (The latch is code; the disarm action is the skill obeying the block — the gate does not auto-disarm for you.)
• Re-arm guard — arm-autonomy refuses while a window is still active (re-arming would reset the baseline and wipe the day's spend — a drain vector); a legitimate re-arm requires an explicit disarm or window expiry. No force-override.
• Per-order discipline — re-verify the session/account_type before every order, re-check the circuit breaker, reconcile against your own sent-order log ∪ broker positions (never lagged positions alone), and run a mandatory data-sufficiency gate that downsizes or abstains when the evidence isn't there — even under an explicit order.

The Valet's own per-order guards (MAX_ORDER_VALUE, MAX_DAILY_VALUE, DUPLICATE_WINDOW_SECONDS) are set as an independent second line when arming. Live execution is venue-s