House Policy

House Policy — Investing v2.0

Committed parameters for the paper book. Every line is either [policy] — a binding requirement for compliant paper-book behavior — or [agent] — guidance honored by judgment. Neither label claims machine enforcement. The current _shared/executor is an LLM-executed broker-routing skill, not deterministic policy-as-code, and it does not implement the book-level limits below. Until a paper-book gate enforces them, automated actions must not be described as policy-compliant merely because this skill was loaded.

1. Scope

• Governs the Archon $1M paper book only. Real accounts are outside this document's binding scope as of v2.0. [agent]
• Claude Code is analyst-only; live placement never originates here, in any mode. [policy]
• Factor targets (§5) describe intended allocation policy; enforcement applies to the paper book. Real-account adherence is by intention, not gate. [agent]

2. Authority

Archon never overrides Risk Architecture to chase returns. [agent]

Thesis-book entries (§3) are decided by Nirav alone: Archon and strategists may analyze or draft the thesis note, but the entry intent is authored and signed by Nirav — no other actor may originate it. [policy]

3. Position-cap hierarchy

Two layers bind simultaneously; the tighter always wins. A downstream layer may be stricter than this policy, never looser. [policy]

effective_cap = min( book_max_position_pct ,
                     sleeve_share × per_position_max_pct )

• book_max_position_pct = 15% of total book [policy]
• max_concurrent_positions (book) = 8 [policy]
• max_gross_exposure = 100% of book value — no leverage; cash is the residual with no minimum. Position caps alone permit 8 × 15% = 120%, so this constraint is what actually bounds total risk. [policy]
• Default stop-loss = −5% from entry unless the intent states a thesis-invalidation level [agent]

Sleeve allocation

Core sleeves run systematic processes — scheduled contribution and drift correction back to committed targets. Tactical sleeves open short-horizon discretionary positions on a view. Thesis sleeves hold long-horizon conviction positions against a written thesis note (below). The class distinction is load-bearing in §6. A discretionary-on-a-view entry may never route through a core sleeve to obtain size; if no non-core sleeve can host it at workable size, the answer is "no position", not reclassification. [policy]

Kill-listed strategists have sleeve 0 and do_not_promote: true; both conditions are AND-gated at the executor. [policy]

Thesis-book sleeve

The thesis-book sleeve exists because conviction positions are neither systematic (core) nor short-horizon (tactical). It is funded by reducing core shares (dca-investor 55→45%, rebalancer 30→25%); total sleeve shares still sum to 100%.

• Max 3 concurrent thesis positions (they also count toward the book's max_concurrent_positions = 8). [policy]
• Every entry requires a written thesis note filed before the intent, containing: the market-implied expectation being disputed, at least one pre-registered falsifier that would force exit, a catalyst or review window, and a review date. No note → no entry. [policy]
• A fired falsifier forces exit or an explicit written amendment of the thesis — silence is not an option. [policy]
• Stop-loss default in §3 does not apply; the falsifier is the thesis position's invalidation level, and it must be stated in the note. [policy]
• No strategist runs this sleeve; entries are Nirav-signed (§2). [policy]

4. Sizing basis

Conviction tiers persist, but size is gated on edge, not confidence. [agent]

Tier size is a fraction of the §3 effective cap for the sleeve the position lives in — never a fraction of the book directly, and never a fraction of a fraction. This supersedes the v1.1 table, which read two ways an order of magnitude apart ("7.5%" of book vs. of sleeve-adjusted cap); neither ambiguous reading survives. [policy]

Worked example. A MEDIUM swing-trader position: effective cap 1.25% of book (§3), so size = 50% × 1.25% = 0.625% of book. A MEDIUM thesis-book position: effective cap 5.0%, so size = 2.5% of book. A HIGH dca-investor allocation may use the full 15% book cap. The tier never overrides the cap hierarchy — §3's tighter-wins rule already bound before the tier applied.

Reference-probability rule — scope

No stated edge versus a reference → no size applies to tactical sleeves as written: a tactical recommendation that cannot name its reference probability is LOW by definition. Confidence alone never establishes attractiveness: a 90% event can be fully priced, a 35% event can be attractive. [policy]

The rule's intent is to force the question "what does the market already believe?" — not to zero every honest long-horizon view. Scope by sleeve class:

• Core — not applicable; systematic entries carry no per-position view.
• Tactical — applies in full. HIGH and MEDIUM both require a named reference probability; the tiers differ by strength of edge.
• Thesis — HIGH requires a named reference probability plus payoff asymmetry, same as tactical. MEDIUM is available without a point probability when the §3 thesis note is complete: it must still name the market-implied expectation it disputes (a qualitative reference), a pre-registered falsifier, and a review date. An incomplete note is LOW. Honest "I cannot name a point probability" therefore caps a thesis position at MEDIUM (2.5% of book); it no longer zeroes it. [policy]

5. Factor targets and rebalancing

Committed weights: Value 25% · Momentum 20% · Quality 25% · Low Vol 15% · Broad Market 15%. [agent]

Tilt envelope

Regime tilts move a factor by at most ±10 percentage points from baseline. A factor is never permitted outside this envelope, whatever any drift band would otherwise allow:

• Tilts must fund each other — weights sum to 100% at all times; a tilt up requires an offsetting tilt down.
• Unclear regime signal → revert to baseline.
• Tilts persist until reverted deliberately or by the unclear-signal rule; they do not expire.

Rebalancing

The active target is the current tilted weight (baseline if untilted). Rebalancing restores the active target — it does not silently undo a tilt.

Trigger fires on whichever comes first:

1. 25% relative drift from the active target — e.g. untilted Value (25%) outside 18.75–31.25%; Value tilted to 35% outside 26.25–43.75%; or
2. exit from the tilt envelope above.

Rule 2 clips rule 1: a factor tilted to its 35% ceiling has an effective band of 26.25–35%, because the envelope binds tighter than the drift band on the upside.

Cadence: quarterly check; trigger fires only on breach.

6. Drawdown gates

Measured peak-to-trough on book equity. [policy]

7. Negative constraints

Not authorized under v2.0, regardless of apparent opportunity: [policy]

• live placement from Claude Code, in any mode
• options intents outside an options-trader profile (which does not exist)
• any position from a kill-listed strategist
• any intent the executor did not receive from a strategist — except thesis-book intents, which it accepts only when Nirav-signed per §2
• any thesis-book entry without a filed thesis note, or any discretionary entry routed through a core sleeve

8. Amendment procedure

Asymmetric by design — pre-commitment binds only when changing the policy is slower than the impulse to change it. [agent]

• Tightening takes effect immediately.
• Loosening requires written rationale plus a 7-day cooling-off period.
• Loosening is forbidden while §6 Restricted or Halt is active, and within 30 days of a stop-out.
• Every change records date, rationale, and whether it tightened or loosened. MAJOR version bump on any loosening; MINOR on tightening or addition; PATCH on wording.
• No calibration or backtest result may mutate a parameter automatically. Human approval, always.

Reasons that are not grounds for amendment

Recent underperformance of a factor · a compelling single idea that exceeds a cap · a strategist requesting more room · short-horizon market moves.

9. Review cadence

Quarterly review against: book drawdown state, factor drift, gross exposure, and cap breaches. Until a deterministic paper-book gate exists, any breach is an enforcement gap to close, not evidence that the policy parameter should be loosened.

Change log

Connections

• _shared/executor — current broker-routing skill; it does not yet enforce the paper-book limits in §3 or §6 and is outside this policy's binding real-account scope
• archon/references/portfolio-rules.md — runtime profile; must not contradict §3
• Named references in §4 are recorded at decision time; no prediction runtime is required
• Track A inventory and Track B (IPS canon + policy-as-code) in vault skill-lab/