Mandate

Living document. Update when rules change; do not silently violate.

Purpose

Manage capital for absolute return with a bias toward high-growth thematic opportunities in AI infrastructure, energy, robotics/AI, and defence — while prioritizing capital preservation and avoiding late entry after parabolic, crowded moves.

Scope

Universe preference: liquid names where exits are realistic. Illiquid microcaps only with explicit size caps and a written exception.

Return & risk posture

• Objective: grow capital meaningfully by capturing structural themes early and correctly sized.
• Constraint: drawdowns and thesis failure must be survivable; no “double down to get back to even” without new evidence.
• Style: long-biased thematic; hedges and shorts optional later, not required for MVP.

Benchmarks

Success is measured against public total-return indices, not against a homemade blend and not against the two-year doubling scenario.

Judge both NAV (cash included — this grades the cash decision) and the deployed sleeve (stock picking only). Prefer rolling three-year windows once the history exists; until then, report since inception and since each review.

Do not introduce a blended S&P/QQQ policy portfolio. Do not change the primary benchmark to make a period look better.

Risk rules (non-negotiable)

Exact percentages can be tuned; the existence of hard caps cannot.

1. Max single-position weight — cap % of portfolio NAV at cost and/or at market (define one primary rule and stick to it). Unratified: the buy gate has enforced market value vs NAV since 2026-08-13. The code made the choice this rule asks for and the mandate has not confirmed it; until it does, the enforced rule and the written rule are not the same sentence.
2. Max theme concentration — cap combined weight in any one theme (AI infra, Energy, Robotics, Defence, Other).
3. Cash buffer — maintain a minimum cash (or cash-like) percentage for opportunity and stress.
4. Thesis invalidation — every position has written kill criteria recorded before or at fill time; a fill without invalidation on the book is a mandate violation, not a TODO. Hit → exit or reduce, do not renegotiate emotionally.
5. No average-down without new evidence — adding requires incremental information, not price alone.
6. Parabolic / crowded filter — prefer not initiating full size into vertical, highly crowded moves without a fresh asymmetric catalyst. Crowding is measured, not felt: check valuation percentile vs the name's own 5-year history, short interest, extension above the 200-day average, and consensus-revision breadth before entry. "Skip if it feels like chasing" notes are binding — if the note exists and the tape gaps up, the answer is skip.
7. Liquidity — position size must respect average volume and expected exit horizon.
8. Kill-switch — measure both (a) drawdown on the unitized deployed sleeve (time-weighted, so a new fill at cost is not a loss) and (b) NAV drawdown (cash included). The sleeve answers “are we selecting or timing badly?” NAV answers “is fund capital actually impaired?” A 90%-cash book can never trip an NAV switch, which is why the diagnostic stays on deployed capital. Do not treat a 15% loss on an $18k seed book as the same event as a 15% loss on $200k deployed.

−15% unitized deployed sleeve = mandatory diagnostic review, not an automatic trim. Classify the move before acting:

Per-name invalidation (rule 4) still forces exit or reduce immediately. Do not wait for a book-level trigger.

Capital Phase 1 (invested cost at or below $75k): cash is already the risk-management lever. A 15% sleeve drawdown does not halt thesis-intact buys and does not require raising cash. The buy gate matches this: it will not refuse a fill solely on the 15% flag while still under the capital Phase-1 cap.

The 15% condition stays on Portfolio → Mandate while the sleeve is still below its high-water mark. Briefing Due only asks for the diagnostic until a covering book-level write exists for this breach (a completed portfolio review task that names the sleeve diagnostic). It re-opens if the sleeve recovers then breaches again, if drawdown deepens by 5 percentage points from the diagnosed print, or after 14 days while still breached.

After capital Phase 1: the same 15% remains a diagnostic. Until a harder NAV-aware capital-preservation threshold is set (revise after the first live month), new buys still need a written override while the flag is on. That is a temporary software halt on new risk, not an order to sell the book — and since 2026-09-19 the gate knows the difference: a reduce or sell skips the caps and the kill-switch entirely, so the halt can no longer block the exit a diagnostic recommends.

1. AI memory cycle discipline — HBM/DRAM/NAND names are the ai_memory sleeve inside AI infrastructure (not a separate core theme). They count fully toward the rule-10 AI-capex cap. Do not treat peak-cycle EPS or trough trailing multiples as “cheap” without contract coverage, mix shift, and normalized-earnings evidence. Prefer starter sizes; add only on new information. Soft guide: memory/storage sleeve ≤ 15% NAV until it earns a larger allocation in review.
2. Factor concentration (correlation-aware) — theme labels are not diversification. The mandate map is a unit-sum allocation, not a stress-beta model: each name has explicit weights, a one-line rationale, and a review date. Unknown names are unclassified and flagged. Cap weighted AI-capex + AI-memory as one position-like risk vs NAV (cash included). DoD autonomy, commercial aerospace, and surgical-procedure growth are not hyperscaler capex. Maintain a standing "hyperscaler capex guidance −20%" stress (haircut × mapped AI-capex/memory weight). Cash is a diversifier versus this factor; the kill-switch (rule 8) stays on deployed capital.

Initial numeric defaults (revise after first live month):

Capital & deployment

PowerFund NAV is the equity book only: cash + marked stocks. Starting allocated capital: $250,000. Bitcoin and gold are a separate sleeve and must not be counted as PowerFund cash.

NAV = cash + market value of open positions. Adding a fill debits cash by cost basis.

New risk is planned in the deployment queue (dollars + window + why) and only hits the book when a fill is confirmed. Do not treat a weekly dollar target as a quota.

Principle: PowerFund does not allocate capital because capital is available. Capital is released in stages as the investment process earns greater trust.

These capital phases are not the software phases in plan.md. Software work supports the PM; it does not authorize more invested cost. Do not wait for software Phase N to enter capital Phase N. The operator process that implements this ladder is gpt-agent-process.md.

Capital deployment phases

Only the $75k Phase-1 cap is live in the buy gate. Continuing past an authorized cap without the written review is a mandate violation, not a rounding error.

Capital Phase 1 — Prove the process with limited live capital

Deploy up to $75k invested cost of the $250k allocated NAV. Baseline ~$10k/month, starter positions first, acceleration only when weakness is thesis-intact.

The objective is not primarily to maximize returns on $75k. It is to answer: does the PowerFund investment process actually work when real money is involved?

Gate: can we research → decide → size → deploy → monitor → invalidate → review, repeatedly, without breaking our own rules?

Evidence we want before crossing $75k:

1. Selection — we can distinguish genuine opportunities from fashionable thematic exposure.
2. Sizing — starters, adds, factor caps, and cash prevent a bad idea from becoming a portfolio-level problem.
3. Anti-chase — we do not deploy simply because an asset or theme is moving.
4. Volatility — a correction leads to reassessment and selective acceleration, not panic or indiscriminate averaging down.
5. Journal — we can tell good process / bad outcome from bad process / good outcome. Clocked grades are the evidence for this, not the claim: read them by decision class, and do not count repeated holds on one name, or several entries into one factor, as independent observations.
6. Operations — the weekly PM workflow runs through PowerFund, not partly in the operator’s head.

Before any buy that would take invested cost through $75k, write the Phase-1 → Phase-2 review (gpt-agent-process.md ritual 13):

1. All four core themes represented, or a hole explicit and accepted.
2. Factor mix (especially AI-capex) acceptable vs the −20% stress — ticker-count is not diversification.
3. Starters have had at least one evidence cycle (print, backlog, or guidance — not just price).
4. The deployment ladder behaved, or skips were written down.
5. Scenario / valuation calibration is credible, not systematically too optimistic.
6. What the next invested cap and cash target are. Default proposal to confirm or revise: Phase-2 cap $150k invested cost.

Starter stubs first. Add on thesis-intact weakness or confirmation. Do not treat a weekly dollar target as a quota.

Capital Phase 2 — Increase capital only after the process earns it

There is no live Phase-2 dollar cap. Phase 1 proves we can operate the machine. Phase 2 should prove the machine scales — that the process still works when positions and portfolio interactions are economically meaningful.

Evidence before committing substantially more of the $250k (and before authorizing a Phase-3 cap):

1. Repeatability — decisions are not one or two lucky trades.
2. Portfolio-level risk — correlation, AI-capex factor, theme concentration, and stress scenarios actually change incremental capital decisions.
3. Allocation — the system chooses not only what to own, but where the next dollar goes.
4. Evidence-driven adds — larger positions are earned by thesis confirmation or a genuine valuation dislocation, not price anchoring.
5. Drawdown behavior — enough volatility to see whether the process behaves under stress.
6. Calibration — expected-return scenarios, bear cases, valuation zones, and invalidation thresholds have some relationship to reality.
7. Signal usefulness — at least some research / data machinery produces information that improves decisions, not just more research. Desired evidence; not a lock that software Phase 2 must be “done.”

At the Phase-2 → Phase-3 review, set the next invested cap and cash target. Default proposal to confirm or revise: up to ~$225k invested cost (10% min cash on $250k NAV).

Capital Phase 3 — Earn the right to deploy most of the allocated capital

Would we trust this process with the entire proprietary allocation through ordinary stress — not “did we beat the S&P this quarter”?

A full multi-year market cycle is the capital Phase 4 bar, not a reason to freeze the last proprietary dollars for years if the Phase-2 proofs are real. Before authorizing something approaching full proprietary deployment, require:

1. Multiple decision types — entries, adds, holds, reductions, and exits — not only buys.
2. Actual thesis failures — we exited or reduced when evidence broke, rather than rewriting the thesis.
3. At least one meaningful factor stress — we saw what happens when a major common factor fell sharply.
4. Risk-before-capital — new risk is evaluated against the whole book before it is committed.
5. Theme rotation process — capital can move across AI infrastructure, power, robotics, and defence on expected return, not narrative loyalty.
6. Benchmark reporting — NAV and the deployed sleeve are judged vs S&P 500 and QQQ as this mandate requires (reporting discipline, not a hurdle that must already be cleared).
7. Process alpha — decisions where the system caused us not to chase, not to average down, to exit a broken thesis, or to buy more in a legitimate dislocation. Errors avoided count.

Min cash still applies. ~$225k invested is the theoretical ceiling under the current 10% cash rule, not a quota.

Capital Phase 4 — Outside capital is a different gate

Only after the proprietary $250k book produces evidence should we consider someone else’s money. Proof is qualitative, not “a few months of winners”:

1. Track record — not a backtest.
2. Repeatable attribution — we can explain where performance came from.
3. Process evidence — contemporaneous theses are preserved; no hindsight reconstruction.
4. Risk evidence — drawdowns, concentration, and losing names were handled per this mandate.
5. Operational reliability — research, data, book, and reviews survive institutional scrutiny.
6. Compliance readiness — legal structure, reporting, custody, investor communications, and regulation are solved before money is solicited.

This mandate does not authorize that activity. Software Phase 4 in plan.md is the product workstream; both gates are required.

Deployment ladder

Pre-committed, not mood-based. "Keeping ammunition for volatility" without pre-commitment is market timing, and the −25% day will not get bought on discretion. Cash is a call option on future dislocations — but maximizing cash because a crash is expected will hurt CAGR if markets keep compounding. The book should rarely reach a 20–30% scare with nothing left to deploy. That can mean cash, or fully valued positions that can be recycled.

Deploy via:

1. A baseline tranche per month up to the authorized phase invested cap. Current: ~$10k/month (set 2026-08-13), which reaches the $75k capital Phase-1 cap around January 2027. Continuing past an authorized cap is the phase-transition review, not creep.
2. Acceleration tranches triggered by theme drawdowns (e.g. −10% and −20% from entry), gated by a thesis-intact checklist — not by price alone. A book-level 15% sleeve drawdown that classifies as a valuation shock (rule 8) is the same kind of event: diagnose, then deploy per this ladder if the checklist passes. Do not freeze the ladder because starters did what volatile growth starters do.
3. Cash level is a decision recorded at review, not drift. If cash exceeds plan for two consecutive reviews, either deploy per ladder or write down why not.

Out of scope for this book: BTC DCA, gold as BTC reserve, and any capital not explicitly moved into PowerFund cash.

Decision process

Every material idea is logged before or at action time:

1. Thesis — what must be true
2. Catalysts — what could reprice the asset
3. Risks — what breaks the thesis
4. Invalidation — observable conditions that force exit/reduce
5. Sizing — why this weight given volatility, liquidity, conviction
6. Outcome review — a grade on a clock, not only at exit. Every material decision is graded at 30 / 90 / 180 days from its own anchor: the fill for an enter or add, the decision date for a hold, since a hold buys nothing and is the judgement to keep owning the exposure from there. Four dimensions — thesis, timing, sizing, risk management — so a market outcome and a process grade stay distinct: a name can fall 15% with the thesis intact and the timing wrong, or rise 20% with the thesis wrong. Grades are immutable and name the horizon they are about, and each is judged on evidence available at that horizon, not on what was learned since. An exit-time grade is still recorded, as an off-clock observation.

Every high-quality dossier should answer not only “would we own this?” but at what price we become unusually eager. States are driven by scenario values vs price, not by a raw percentage drawdown:

A name falling from $200 to $150 is not automatically more attractive if base value fell from $220 to $140. A name whose base value rose while the stock fell has become the situation cash is for. Rank these states on the monthly opportunity pass (gpt-agent-process.md ritual 9).

Edge definition

We seek situations where fundamentals, CapEx, contracts, policy, or supply-chain evidence are improving (or about to) while price and narrative have not fully discounted that path.

That edge has two complementary expressions:

1. Volatility alpha — own and research high-quality structural winners; concentrate deployment when fear creates a valuation dislocation while the thesis, and our estimate of intrinsic value, is intact. We do not buy because something fell 25%. We buy more aggressively when something we already understand fell and intrinsic value did not.
2. Bottleneck discovery — continuously search further down the supply chain for less-followed companies whose earnings power is being transformed before valuations fully reflect it. Prefer control of a bottleneck, switching costs, and inflecting orders over “it is small.” The ideal discovery is an obscure mid-cap in a rapidly scaling system whose qualification creates switching costs, orders inflect, estimates rise repeatedly, and the multiple rerates.

We do not define edge as:

• trading every headline in a hot sector
• buying solely because a story is fashionable
• assuming “AI” or “defence” in the name is sufficient
• a permanent exclusion list of popular names

Crowding is a penalty to required margin of safety, not an absolute ban. A heavily owned consensus winner (NVIDIA, a hyperscaler, a future OpenAI listing) can still be owned after a large dislocation. An underfollowed name can earn capital earlier because expectations are lower. Crowding still creates higher overlap, violent unwind risk, and less asymmetry — measure it (rule 6); do not pretend it is a virtue.

Label trades honestly. A datapoint that appears in an IEA flagship report, a NATO communiqué, or hyperscaler guidance is consensus by definition — citing it is not evidence of earliness. When a position does not clear the earliness bar, call it what it is: momentum with fundamental support. That label is allowed, but it changes the exit posture from buy-and-hold to trend-following with theme-level exit indicators (see themes.md) and makes the crowded filter (rule 6) mandatory, not advisory.

Stretch targets never override risk rules. Return scenarios (e.g. doubling over N years) are scenarios, not objectives. Recurring 20–40% thematic corrections are part of the return engine if we have already done the research; they are not a reason to disable the 15% deployed diagnostic. Keep measuring it even if a growth book is likely to touch it. What capital Phase 1 changes is the response (diagnose, maybe deploy) — not the measurement, and not per-name invalidation. (See 2026-08-13 strategy second opinion, finding 3.)

Instruments & leverage

• Spot/long equity and liquid commodity proxies first.
• Options, leverage, and shorting: allowed only with explicit rules added to this mandate (size, max loss, purpose).
• Until then: no leverage beyond broker default cash/margin for settlement convenience; no speculative options book.

Review cadence

Open mandate decisions

Recorded here so they are visible as undecided rather than sitting only in a review. Each is the operator's call, not the code's, and none is resolved by anything shipped so far.

Factor concentration is deliberately not on this list as a proposed limit. The diagnostics attribute most of the drawdown to one AI-infrastructure factor across names filed under five themes, so theme caps are not constraining what is actually correlated — but measuring and reporting the factor split should come before any threshold is fixed to it, or the number gets fitted to one small sample.

Compliance note (future scale)

Personal capital today. Any third-party capital, advice, or published signals for others is capital Phase 4: separate legal/compliance work, and only after the proprietary book has produced the Phase-4 proofs above. This mandate does not authorize that activity.