0006 — A named-fund reference portfolio may be published without promoting a sleeve

Date: 2026-08-12. Status: accepted.

Context

Decision 0004 states that “the deliverable of this research programme is a design map and a control, not an allocation”, and that “an allocation becomes appropriate only after the investor policy is defined”. That condition is not met: framework open decision 1 is still open, and no experiment has declared a horizon, liability model or drawdown tolerance.

Read literally, that forbids the thing the project exists to produce. Read as intended, it forbids something narrower — presenting an allocation as though the research had validated it. The two readings need separating, because the cost of the literal one is now visible:

• The design map records, per candidate, what would have to become true before it could be held. It does not say what to hold now, so a reader who wants an answer must assemble one themselves from nine research pages, and will assemble it wrong.
• The largest measured result in the repository — a contractual ~109 bp/yr against the investor’s own counterfactual, 99% confident in about twelve months (structural and tax-aware edges) — is bought entirely by decisions a design map cannot express: which fund, which wrapper, which account, which lot. Withholding the construction withholds the only near-certain edge the repository has found.
• The absence has already caused a specific error. Experiment 004’s index verdict was repeated to the project owner as though it applied to KMLM, DBMF and CTA, which were never tested. That error is recorded on its own page. A page that names products and states exactly what was measured on each is the structural fix.

Decision

A reference portfolio naming concrete funds, weights and accounts may be published, provided it makes no promotion claim. It lives in docs/research/portfolio-recommendation.md and nowhere else.

“Recommended” in that page means, and may only mean, the best-supported construction given the evidence. It does not mean the construction is expected to beat an index, and it does not advance any sleeve’s status.

Four constraints make that separation enforceable rather than rhetorical.

1. Every holding carries the status of what it buys, in the closed vocabulary, with the experiment that set it. A holding whose status is exploratory must say so beside its weight.
2. Every holding carries its evidence class — contractual, risk premium, risk control, or “nothing better exists” — and the class governs how the line may be described. A risk premium may never be described as an edge.
3. Every tilt is priced in confidence terms, not in expected return alone: edge, tracking error, probability of being ahead at a stated horizon, and the horizon to 90% confidence, from studies/outperformance_horizon.py. A tilt quoted as “+X pp/yr” without its tracking error is not reportable.
4. The investor-policy inputs that remain missing are stated on the page, and the one parameter the evidence cannot set — the equity/bond split — is presented as the investor’s choice with the others held fixed.

The page is subordinate to the research framework: it may not state a premium, a status or a cost that the framework or a linked experiment page does not.

What this does not change

• No sleeve is promoted. Decision 0004 stands in full, including zero leverage, rebalancing as risk control only, and the exploratory products being usable as implementation proxies and for nothing else — forty-eight on the US shelf and twelve ex-US, of which eight survive a fair comparator.
• No number from research/ may appear in the shipped application as a finding Amended by decision 0007: the application may render a finding provided its status, date, interval and source travel with it. The ban on shipping price data is separate and still stands (decision 0002).
• The cheap broad-market portfolio remains the control (decision 0003). The reference portfolio is that control plus placement discipline plus two optional exploratory sleeves; it is not a competitor to it.

Alternatives considered

Publish nothing until the investor policy is defined. Rejected. The policy inputs that are missing bear on one parameter — the equity/bond split — and on the size of the optional sleeves. They do not bear on fund choice, account placement, lot discipline or turnover, which is where the measured edge is. Withholding all four because one is unsettled is a worse error than stating the fourth as a range.

Publish a portfolio without naming funds. Rejected. “Hold a cheap broad-market fund” is unfalsifiable and unimplementable; naming VTI at 3 bp against a measured 0.55 bp spread, and stating the date, is both. The cost is that fund-specific facts decay, which the page handles with dated facts and a stated review trigger rather than by staying vague.

Add the reference portfolio to the framework instead of a new page. Rejected under the one-canonical-place rule, and for the opposite reason to usual: the framework answers whether a return source is real, and mixing “what to hold” into it would let a construction choice inherit the framework’s evidentiary authority. Two pages with sharply different questions is the correct split; a third page on the same question would not be.

Let the page state a probability of beating the market. Rejected. Against a cheap index the honest budget is ~5.4 bp against 313 bp of tracking error and a 0.538 thirty-year probability, which is not a claim worth making and is trivially misreadable as one that is.

Consequences

• The repository now has two audiences and two entry points: the framework for whether a return source is real, and the recommendation page for what to hold. docs/README.md must keep both visible and must not let a reader mistake the second for a finding.
• Any status change in any experiment obliges a check of the recommendation page, since every holding on it cites a status. That coupling is deliberate and is the cost of naming funds.
• Fund-specific facts on that page decay on a timescale of months. It carries a review trigger — the SEC multi-class order count and the per-fund fee table — and a standing instruction to re-check rather than re-quote.
• Supersede this record, do not amend it, if a sleeve is ever promoted: the recommendation page’s meaning of “recommended” would then have to change.