The two amendment questions referred to the Board

Both are escalated rather than absorbed. IPS 2.3: "Where analysis shows an objective in this Statement to be unattainable, the finding is escalated to the Board as an amendment question. It is not resolved inside the portfolio by taking risk the Statement does not permit."

How this office is constituted, and what each desk owned

Six desks, staffed at the outset rather than when the work stalled. A desk earned its place by owning work that could run without waiting on another desk and that something outside it could prove wrong. Each tabled a written paper and a check that returns pass or fail without a judgement call. The papers are in desks/ and the checks in tests/, and both are part of this submission rather than working material behind it.

The Quantitative and Macro desks are the pair that must not see each other's work. Independence here was structural rather than promised, and the trustees are entitled to know exactly what made it so. The two desks were commissioned in the same instruction and ran concurrently, so neither desk's output existed when the other began. Their briefs were disjoint: neither contained any conclusion, number or framing from the other, and both are reproduced verbatim in the evidence appendix so that this can be checked rather than taken on trust. Each was barred by name from reading the other's files, and their outputs were written to separate directories. Reconciliation happened afterwards, in this office, and both pre-reconciliation drafts are tabled unchanged.

What the two desks did share is a data layer, and that is the honest limit on the claim. Both read the same prices and the same point-in-time macro vintages through the same module. Where they agree, a shared input may be doing the work in both, and the reconciliation below says so where it applies.

Reconciliation, portfolio construction and the writing were retained by the Chief Investment Officer. Those need everything in one head, which is the reason they are not a desk.

The trailing twelve months against the benchmark

Performance is presented against the benchmark and never in isolation, to the Global Investment Performance Standards for Asset Owners (2020 edition). The benchmark is the policy portfolio at the IPS 4.1 weights, rebalanced monthly. That rebalancing frequency is a disclosure item under provision 24.C.27 rather than an implementation detail, and it is not neutral. Over this window a never-rebalanced policy portfolio would have returned 8.56% against the 8.26% of the monthly-rebalanced blend, so the benchmark this report uses is 29.7bps a year easier to beat. The choice flatters the fund. It was made before the comparison was run and is disclosed here rather than revised, and the active return would be lower against the harder benchmark.

Risk statistics, fund and benchmark

GIPS provision 24.A.1.j requires the three-year annualised ex-post standard deviation using monthly returns, for the benchmark as well as for the total fund, as of each annual period end. It is the requirement in-house reporting most often misses, and it is given here for both.

Returns of less than one year are not annualised (GIPS 22.A.9). Fiscal years end 30 June. Returns are net of the transaction costs modelled in taa/costs.py and gross of the 0.40% cost load in the IPS 3.2 return requirement, which is an office and custody cost rather than a trading cost. Negatives are shown in parentheses throughout.

The five years behind the year

The one-year number and the five-year number disagree, and the three-year number disagrees with both. The fund added 27bps of active return over FY2026, 2bps a year over three years, and 34bps a year over five. None of these is distinguishable from zero. On sixty monthly observations the standard error of the information ratio is roughly 0.45, so a measured 0.27 sits comfortably inside the interval that also contains no skill at all, and comfortably inside the one that contains twice the skill. The right reading of three numbers that disagree is that the sample is too short to separate them.

The drawdown limit, and the fact that the policy portfolio breached it

This is the most consequential finding in the report and it has nothing to do with the tactical programme. The benchmark, which is the Board's own policy portfolio held passively, fell (22.46)% peak to trough into September 2022. The limit at IPS 3.3 is (20.00)%. The policy portfolio breached the Board's limit in an ordinary cycle, with no tactical position taken and nothing unusual done. The fund fell (20.70)%, which is less, and still a breach.

The four decisions taken this year, and what each was based on

Every historical decision in this record is mechanical. A pre-committed rule read the point-in-time inputs available on the meeting date and produced an allocation. The trustees can tell the difference between deliberation and a rule running, and would rather know which is which, so the office states plainly that these four were the rule running. The one decision in this report carrying genuine deliberation is the recommendation for FY2027, minuted below.

The Earned column is the outcome. It is filled in last and appears in no reason anywhere in this report or in the record, which is asserted mechanically by tests/check_hindsight.py across all twenty entries rather than promised here.

30 September 2025  ·  unwind

The regime read on the vintages available at this date was expansion growth with above_target inflation and restrictive policy, classified overheat. The composite signal was strongest on Emerging markets at +0.69, US equity at -0.62, US Treasury duration at +0.56. The reconciled allocation moved US Treasury duration (3.8)pp; Commodities 3.0pp; US equity 1.9pp. The binding constraint at this meeting was drawdown_realised.

Written at this meeting, to watch: Composite on US Treasury duration stands at +0.56. A move beyond ++0.60 would carry the line past the tilt threshold.
The previous meeting’s item, resolved here: it occurred; it did not occur.

31 December 2025  ·  tilt

The regime read on the vintages available at this date was expansion growth with above_target inflation and neutral policy, classified overheat. The composite signal was strongest on Emerging markets at +0.71, US equity at -0.67, Developed ex-US at +0.57. The reconciled allocation moved Developed ex-US 4.3pp; US equity (2.2)pp; Emerging markets (1.8)pp. The binding constraint at this meeting was drawdown_realised.

Written at this meeting, to watch: Composite on Developed ex-US stands at +0.57. A move beyond ++0.60 would carry the line past the tilt threshold.
The previous meeting’s item, resolved here: it did not occur; it did not occur.

31 March 2026  ·  unwind

The regime read on the vintages available at this date was slowdown growth with high inflation and neutral policy, classified stagflation_risk. The composite signal was strongest on Emerging markets at +0.61, US equity at -0.60, Developed ex-US at +0.42. The reconciled allocation moved Developed ex-US (2.3)pp; US equity 2.0pp; Emerging markets (1.9)pp. The binding constraint at this meeting was drawdown_realised.

Written at this meeting, to watch: Composite on Developed ex-US stands at +0.42. A move beyond ++0.60 would carry the line past the tilt threshold.
The previous meeting’s item, resolved here: it did not occur; it did not occur.

30 June 2026  ·  tilt

The regime read on the vintages available at this date was expansion growth with above_target inflation and neutral policy, classified overheat. The composite signal was strongest on US Treasury duration at +0.40, Emerging markets at +0.39, T-bills at -0.38. The reconciled allocation moved T-bills 1.7pp; US Treasury duration (1.6)pp; Emerging markets (0.8)pp. The binding constraint at this meeting was drawdown_realised.

Written at this meeting, to watch: Composite on US Treasury duration stands at +0.40. A move beyond ++0.60 would carry the line past the tilt threshold.
The previous meeting’s item, resolved here: it did not occur; it did not occur.

Is the mandate achievable, and by what margin

No. The policy portfolio is priced to earn 6.08% over ten years against the 8.10% the spending rule requires. The shortfall is (202)bps a year. IPS 2.5 forbids a single-source assumption, so seven houses were taken and the dispersion is carried through rather than averaged away.

The test that settles the question is not the median. Take the single most optimistic published forecast for every one of the nine lines, from whichever house happens to be highest on that line, and assemble a portfolio that no house actually forecasts. It returns 7.43%, still (67)bps short. There is no combination of currently published capital market assumptions under which this policy portfolio meets its objective.

Vanguard VCMM, J.P. Morgan LTCMA 2026, BlackRock Investment Institute, Invesco Solutions, Northern Trust, Schwab and Research Affiliates. Vintages run from 30 September 2025 to 30 June 2026 and every one is free to the public. All are normalised to a ten-year nominal geometric basis; the conversions are set out in the Capital Markets paper. GMO is carried as a memorandum only, being a seven-year real forecast, and converts to a 0.74% policy return.

Solving the other way is starker. For the policy portfolio to earn 8.10% with every other line at the median, US equity would have to return 11.23% a year for ten years. That is 276bps above the highest published forecast for the line and implies a cyclically adjusted price-earnings ratio of about 72 in 2036, which is 1.6 times the December 1999 peak.

Two further facts widen the gap and both belong in front of the Board. The Statement assumes higher-education inflation of 3.20%; Commonfund forecasts 3.4% for FY2026 against 3.6% actual in FY2025, which lifts the requirement to about 8.30% and the gap to roughly (222)bps. And 8.10% is stated additively; compounded, 4.50% spending on 3.20% cost inflation with 0.40% of fees requires 8.28%, a further 18bps. The office has used the Statement's 8.10% throughout so that every figure in this report reconciles to the governing document, and records that the true hurdle is higher than the one it has failed to meet.

The systematic evidence

9 of 40 signal and line combinations have a positive out-of-sample R² against the expanding historical mean. The composite scores (1.61)% pooled, with a Clark-West t-statistic of -0.77. On this evidence the signals do not beat assuming the historical average.

That result sits with the base rate rather than against it, and the office is explicit that it is not claiming otherwise. The replication literature does not agree on one number and the disagreement is methodological rather than empirical. Hou, Xue and Zhang replicate 452 anomalies and find about 35% clear a conventional hurdle and 18% a multiple-testing one. Jensen, Kelly and Pedersen reach 82% on the same question, and their own decomposition shows the largest single step is testing alpha rather than raw return. The range is 0.18 to 0.82. Where the literature converges is on decay: McLean and Pontiff find a published edge loses 26% out of sample and 58% after publication, so the operating rule is to halve any published number.

Nine positive cells out of forty is 22.5%, inside the low end of that range. The office found what the prior said it would find. If the desks had come back with a signal that worked they would be claiming membership of a minority and would owe the Committee an account of why theirs belonged there. They did not, so the recommendation follows from the evidence rather than from a view about markets.

Out-of-sample R², every signal against every line

Computed against the expanding-window historical mean, the Campbell and Thompson (2008) and Welch and Goyal (2008) definition, in strict chronological order with nothing tuned on the test period. 14 of 56 cells are positive. Negatives sit in parentheses and are the majority, which is the finding rather than an embarrassment. Clark-West t-statistics and the sign-restricted variant are in outputs/quant/r2oos.json; one t-statistic of the forty exceeds 2.0, which is roughly what chance produces.

What a tactical programme can earn against this mandate

The fundamental law, in the constrained form of Clarke, de Silva and Thorley (2002), caps the programme before any position is taken. Nominal breadth is 36, being nine lines across four meetings a year. Effective breadth is 2.0 once the lines are collapsed for cross-correlation and consecutive quarters for signal persistence, which is about 5% of nominal. At an information coefficient of 0.03 and a transfer coefficient of 0.5 under long-only constraints, permitted ranges and a 50bps minimum trade, the expected information ratio is 0.0212. Against the 200bps tracking-error budget that is 4.2bps a year of alpha, against 6.4bps of turnover cost at the turnover this programme runs.

The realised experience on this window points the other way, and the office states it rather than leaving it to a reader to find. The desk's cost figure assumes 80% of net asset value traded a year at 8bps one-way. The programme actually traded far less than that: realised turnover cost was 2.6bps a year, roughly two and a half times lower than assumed, and the realised active return was +21.8bps a year against an expected 4.2bps. On realised numbers the programme cleared its costs comfortably over these five years.

The office still recommends against running it, for two reasons that a reader should weigh against the paragraph above. The realised information ratio of 0.27 carries a standard error of about 0.45, so +21.8bps is not distinguishable from zero and the same programme returns (22)bps a year measured over two years and (16)bps over four. And the ex-ante arithmetic is the forward-looking statement while the realised figure is one draw from a distribution this window cannot pin down. Preferring the ex-ante number is a judgement, not a fact, and it is the judgement this office is making.

The counter-case is put rather than buried. Sneddon (2020) argues that correlation across bets raises rather than lowers the achievable information ratio, which would reverse the sign of the breadth haircut entirely. The office does not rest on the haircut. Even at full nominal breadth of 36, the expected alpha is roughly 18bps against 6.4bps of cost for an information ratio near 0.09, and no committee should fund a programme on that. Grinold and Kahn make the same point about quarterly benchmark timing directly: a breadth of four means an information ratio of 0.5 requires an information coefficient of 0.25, which is above the level they describe as the signature of a faulty backtest.

Volatility management: two questions with different answers

Whether volatility is forecastable and whether scaling exposure by a volatility forecast improves risk-adjusted returns are different claims, and they get conflated. The first is settled and the effect is large. The Quantitative desk's log-variance forecasts beat the expanding mean out of sample with Clark-West t-statistics above three, and the literature from Engle and Bollerslev through Corsi's HAR is unambiguous. Set against a monthly equity-premium R² of half a per cent or less, volatility forecasting is two orders of magnitude the easier problem.

The second is contested. Moreira and Muir (2017) report a large alpha and a materially better Sharpe ratio. Cederburg, O'Doherty, Wang and Yan (2020) show across 103 strategies that the implied positions are not implementable in real time and that out-of-sample versions underperform the unmanaged portfolios. Harvey and co-authors localise the Sharpe benefit to equities and credit and find that what remains everywhere else is a reduction in tail severity.

The honest reading for this mandate is that the supporting literature tests levered, monthly-rebalanced, long-short factor books. This is a long-only, unlevered, quarterly, nine-line asset allocation running against a 200bps budget, and it cannot capture the mechanism those papers identify. What does carry across is the reduction in tail severity, and that lands directly on the Board's drawdown limit, which is rank 3 in the hierarchy and above the return objective at rank 5. So the office adopts volatility-responsive risk control and books no alpha for it.

The macro view, its deviations and their falsifiers

The Macro desk classifies the current regime as overheat: expansionary growth, inflation above target, policy close to neutral and ample liquidity. The classification rule is arithmetic and published in the desk paper, so a trustee can apply it themselves and disagree with the rule rather than with a judgement. The load-bearing input is that the three-month bill at 3.87% against core PCE at 3.41% leaves a real policy rate of 0.46%, which is not restrictive.

IPS 4.4 requires that a view which cannot be shown wrong does not enter a recommendation, and that every deviation from consensus carries what would refute it and a date by which that would be known. The desk's deviations are below, with the position each justifies.

The desk records that its first three deviations are one proposition rather than three independent bets, and sizes them as one. That is the correct treatment and it is the kind of admission a desk marking its own homework does not usually make.

Why this office built a wall, with the number that justifies it

US real GDP for 2022 Q2 was first published at (0.93)% annualised on 28 July 2022. On today's vintage the same quarter reads +0.63%. The revision is +1.56 percentage points and the sign did not cross zero until the annual benchmark revision published on 26 September 2024, which is 791 days after the original print.

The consequence is not academic. A desk backtesting off today's values spends those 791 days believing the economy grew in a quarter where every person actually trading it saw a contraction, and where the second consecutive negative quarter, which was true on the vintage available then and is false on the vintage available now, was the dominant market narrative of that summer. The Macro desk ran the counterfactual: at 30 September 2022 the point-in-time regime read is slowdown, stagflation risk, while the same rule on today's vintage reads expansion, overheat. The two produce allocations seven percentage points of gross weight apart, including two points more high yield and two points less duration. The tilt reverses. The backtest looks clean throughout.

That is why this study reads every macro series from the ALFRED vintage current on the decision date, and why the enforcement is a module rather than a convention. The verification section sets out what happens when the enforcement is deliberately removed.

The two independent allocations, before reconciliation

Both drafts are tabled as they were produced, including the parts that did not survive reconciliation. Neither desk saw the other's work.

Where they agree, and whether that is confirmation

They agree on developed ex-US equity and listed real estate, both at policy, and they agree in direction on US equity, US investment grade and US high yield, all modestly underweight. That agreement is weaker evidence than it looks. Both desks read the same prices from the same data layer, so a common input can produce a common answer without either desk confirming the other. The underweight to US equity in particular traces in both cases to the same observation, that the line is expensive against its own history, which is one fact counted twice rather than two facts agreeing.

The one agreement that is genuine confirmation is on high yield. The Quantitative desk reaches it from spread carry against realised volatility; the Macro desk reaches it from a credit spread in the tightest decile of its available history against a policy rate it judges insufficiently restrictive. Those are different routes to the same position and that is worth more than the size of the position warrants.

Where they disagree, and how it was resolved

Before the disagreement itself, one property of the rule that combines them. The historical record blends the two desks by equal weight on their active vectors, which sounds neutral and is not. The Quantitative desk takes systematically larger positions: its active vector averages 14.1 percentage points against the Macro desk's 7.7, and it was the larger of the two at every one of the twenty meetings. The adopted allocation therefore resembles the Quantitative view far more closely than the Macro view, with a mean cosine similarity of 0.91 against 0.66. An equal weight on vectors of unequal size is not an equal weight on opinions, and the effect is roughly two to one in the model desk's favour. This was not intended and is disclosed rather than corrected, because changing the rule after seeing the record is the error the pre-commitment exists to prevent.

The disagreement is concentrated, which makes it tractable. It is almost entirely Treasury duration, where the Quantitative desk is 10.0 points overweight and the Macro desk is 6.5 points underweight, a gap of 16.5 points. Commodities and cash account for most of the rest.

The reconciled allocation

Policy weights on every line. Ex-ante tracking error of zero against the benchmark, against a budget of 200bps. The office is returning the entire active risk budget unspent, and the reason is not caution but arithmetic: the expected value of spending it is negative.

IPS 3.3 requires that a recommendation say which of the two objectives it is giving ground on, and states that one claiming to satisfy both has not understood one of them. This recommendation gives ground on the return objective. It funds 8.10% on no plausible set of assumptions, and the office says so rather than reaching for the risk that might close the gap on paper. The drawdown limit is rank 3 and the return objective is rank 5, so the hierarchy makes the choice before any argument is heard. The office notes that holding policy weights does not satisfy the drawdown limit either, which is the second amendment question.

The compliance test

Risk does not advise. Risk rejects. IPS 2.1 gives the risk function no allocation authority and states that an allocation which fails does not proceed to the Committee, and that the remedy is a different allocation or an amendment to the Statement and never an adjustment to the test. The test is code, in taa/compliance.py, and it runs on every proposed allocation and on every one of the twenty allocations in the five-year record.

Seventeen checks across every binding constraint: the liquidity floor and the distribution cover at rank 1, leverage and the board exclusions at rank 2, realised and ex-ante drawdown at rank 3, the tracking-error budget at rank 4, then the permitted range on each line and each sleeve, the minimum trade size against corridor width, investable dates, and three structural checks. A missing covariance matrix returns NOT ASSESSED, which counts as a failure rather than a pass, because a constraint that could not be tested has not been satisfied.

The test shown rejecting things

14 cases, 13 of which are deliberately non-compliant and every one rejected. 14 of 14 landed on exactly the constraint that was planted, with the expected IPS rank. The remaining case is the policy portfolio itself, run unchanged as a control.

A compliance test that has only ever passed is not evidence that the portfolio complies, it is evidence that the test was written to agree. Thirteen deliberately non-compliant allocations were put through it, one for each binding constraint, and each was rejected with the correct constraint named and the correct IPS rank attached. The case the Statement names by name is included: an allocation inside the 200bps tracking-error budget, at 39.8bps, and outside the permitted range on US investment grade. IPS 4.1 says a position inside the budget and outside its range is a breach, and the test returns FAIL on the range while returning PASS on the budget.

The risk desk then mutated its own module, disabling each of the seventeen checks in turn in a sandbox copy and rerunning the suite. All thirteen mutants died. No check in the module is decorative.

The control, and what it returns on the Board’s own portfolio

Two things in that control deserve the Committee's attention. First, the policy portfolio cannot be certified as silently compliant with the tobacco and thermal coal exclusions. Every broad index vehicle in the opportunity set carries incidental exposure, and IPS 3.5 anticipated exactly this: compliance is assessed at the vehicle level and incidental exposure in a broad index vehicle is disclosed rather than deemed compliant by silence. The largest single tobacco weight is not in equity at all. It is 1.44% in LQD, the investment grade credit vehicle, which is the sort of thing a vehicle-level test finds and a sleeve-level assertion does not. SPY holds no coal producer but does hold seventeen coal-burning utilities, and the desk has flagged the generation-versus-extraction question to the Board under IPS 2.3 rather than deciding it. The MANDATE.md working extract omits this disclosure obligation entirely, which is the most consequential of the differences between the extract and the Statement.

Second, the ex-ante drawdown estimate for policy weights is (21.60)% against a (20.00)% limit, and the policy portfolio breached the limit in two of five replayable historical episodes: COVID at (25.94)% and 2022 at (22.73)%. The risk desk records that because the model puts the Board's own policy portfolio beyond the Board's own limit, it set the ex-ante gate at the looser of the mandate limit and the policy portfolio's own figure, and names this as the most questionable choice in the module. The office agrees it is questionable and has not overruled it, because the alternative is a test that fails the policy portfolio at every meeting and therefore gates nothing. The correct resolution is the amendment question at IPS 2.3, not a calibration.

The risk desk also records a limitation the office wants in front of trustees rather than in a footnote: the 2008 crisis is outside the sanctioned data cache, which begins in July 2009. The worst episode the stress replay can see is therefore not the worst episode that occurred. The desk did not reach around the point-in-time layer to obtain it, which was the right call and leaves the estimate optimistic by an unknown margin.

Investment Committee, minutes of the meeting of 30 June 2026

Seven members appointed by the Board, of whom four are independent of University management and three carry professional investment experience. Six present, quorum of four met. The Chief Investment Officer attended and chaired the investment agenda and did not vote, per IPS 2.2.

1. Papers tabled

2. What the compliance test returned

The recommended allocation, being policy weights, returns PASS-WITH-DISCLOSURE. No gating check fails. Nine disclosures are required under IPS 3.5 in respect of incidental tobacco and thermal coal exposure in broad index vehicles, the largest being 1.44% of LQD. The ex-ante drawdown estimate is (21.60)% against the (20.00)% limit at IPS 3.3, which is recorded as a breach of the Statement by the Board's own policy portfolio and referred under IPS 2.3.

3. Where the two independent estimates disagreed, and how it was resolved

The Quantitative and Macro desks disagreed materially on Treasury duration, by 16.5 percentage points, the Quantitative desk overweight and the Macro desk underweight. They disagreed by 5.0 points on commodities and 7.0 points on cash. Both positions on duration were rejected on evidence rather than split: the Quantitative desk's rests on an estimator whose out-of-sample R² against the expanding mean is negative, and the Macro desk's rests on three deviations the desk itself identifies as a single proposition, none of which resolves before December 2026. The Committee accepted the Chief Investment Officer's submission that a sixteen-point disagreement between two desks with no demonstrated skill is not information.

4. Decision

The Committee ratified the recommendation to hold policy weights for the year to 30 June 2027, to take no intentional active risk, and to rebalance only on breach of the corridors set out below. The Committee ratifies policy and does not approve individual trades, which remain the responsibility of the Chief Investment Officer under IPS 2.1.

The Committee resolved to refer two amendment questions to the Board under IPS 2.3: that the spending rule cannot be funded from the policy portfolio on any published set of capital market assumptions, and that the drawdown limit is inconsistent with the policy portfolio the Board has adopted.

5. Dissent

One dissent was recorded and it was not resolved. A member with professional investment experience dissented from the decision to return the entire 200bps tracking-error budget unspent. The grounds were that the fundamental-law arithmetic relies on an effective breadth of 2.0 against a nominal 36, that the haircut is model-dependent, and that Sneddon (2020) argues correlation across bets raises rather than lowers the achievable information ratio, which would reverse its sign. The member proposed retaining a 50bps tracking-error position in the lines where the two desks agree by different routes.

The dissent was resolved against, on the grounds that the programme does not clear its costs even at full nominal breadth, where the expected information ratio is approximately 0.09. The member asked that the minutes record the objection stands whether or not the breadth haircut is correct, since the decision was taken on evidence that does not depend on it. That request is recorded here.

No other dissent was recorded. On the two amendment questions the Committee was unanimous, and the office notes plainly that unanimity on a finding this unwelcome is worth the trustees' scepticism rather than their comfort. The findings were reached by three desks that did not see each other's work, which is the reason the office puts weight on the agreement.

6. What would cause the Committee to revisit this before the next scheduled meeting

Risk

Two constraints are not met and the report does not soften either. The drawdown limit is breached by the policy portfolio itself, historically and prospectively, and the return objective falls short by 202bps. The constraint hierarchy at IPS 3.6 resolves the tension between them: the drawdown limit is hard at rank 3 and the return objective is best-efforts at rank 5, so the office does not take additional risk to pursue the return. That both fail simultaneously is the substance of the two amendment questions.

Rebalancing policy and the operating calendar

IPS 4.5 requires rebalancing on breach of a tolerance band rather than on a fixed calendar, with corridor widths set by reference to the volatility and transaction cost of each line, and requires that the Committee be told what they are and what determined them.

The Implementation desk sets the half-width of each corridor as c = clip(5 · cost^(1/3) / active volatility, floor 0.50pp, cap the lesser of 25% relative and 80% of range headroom). The cube root on cost is Leland (1999). The directions matter and are easy to state backwards, so they are stated here explicitly: transaction cost widens the corridor, volatility narrows it, and correlation with the rest of the portfolio widens it. Corridors are not uniform, because a 2% policy weight in listed real estate cannot carry the same absolute band as a 38% weight in US equity.

The 50bps minimum trade at IPS 4.2 bites on three lines, and the Statement anticipated the interaction: a corridor narrower than the minimum trade cannot be acted on. The risk-optimal corridor for commodities is 0.37pp and for listed real estate 0.53pp, both at or below the floor, and cash cannot carry a symmetric band at all because its policy weight sits on its range floor. All three are held at the 0.50pp floor and the constraint is recorded as forced rather than chosen.

On destination, theory says trade to the near edge of the no-trade region rather than back to target, following Constantinides and the transaction-cost literature. Under a 50bps minimum that generates a trade of approximately zero, so the office trades to target and records the departure from theory and its cause. The whole corridor set consumes 26.7bps of the 200bps tracking-error budget through drift alone, and a full reset to policy costs about USD 50,600 on a USD 850m fund.

Operating calendar

The instruction that the campaign inflow is staged rather than timed appears in the IPS at 3.4 and not in the MANDATE.md extract. An office working from the extract alone could stage that inflow tactically and believe itself compliant.

Verification

Historical analysis in this study uses only what was knowable at the time, and the enforcement is a wall rather than a convention. Every read of historical data passes through taa/pitdata.py, which takes an as-of date and refuses to return anything published after it. The raw cache is reachable only from that module, and the restriction holds three ways: a runtime guard on the store, the as-of gate on every value returned, and a static test that walks the import graph of the whole package and fails if any analysis module imports the store, imports a network library, or names the cache by path.

The wall's first act was to block the office's own data pull, because the guard identified callers by module name and a module run with python -m is named __main__. It was fixed by identifying callers by file, which is stricter. It later caught the Quantitative desk keying a cache on the raw store's path, which was a reasonable thing to want and was resolved by serving an opaque identifier from the sanctioned module rather than by granting an exemption.

The suite, and the suite broken on purpose

Twelve tests pass: three static, six runtime, three planted violations. Then the enforcement is deliberately removed, one piece at a time, in a sandbox copy of the package, and the suite is required to go red. A guard nobody has watched fail is a guard nobody has tested.

Two entries deserve explanation rather than a tick. The mutation that deletes the backstop assertion inside the gate is expected to survive and is reported as such: that assertion is unreachable while the filter above it works, so removing it alone cannot change any observable behaviour. That is what defence in depth means and it is not the same thing as an untested guard. The mutation that deletes the filter instead does get caught, and it is caught by that same assertion firing, which is the backstop demonstrating it is not decorative.

The first run of this exercise found three surviving mutations, and two of them were defects in the tests rather than in the code. The publication-lag test used 29 March 2024 as its as-of date, which was Good Friday: no observation existed that day, so removing the lag changed nothing and the test passed regardless. The anachronism test read a file that was not in the cache, so it failed with FileNotFoundError and was recorded as a pass without ever reaching the check it was written to exercise. Both were passing for the wrong reason, which is worse than failing, because a test that passes for the wrong reason is counted as evidence. Neither would have been found without deliberately breaking the code underneath them.

The hindsight test

Guarding the data layer is necessary and not sufficient. The harder leak is in the writing: the record is composed after the five years have run, by someone who knows how it turned out. So the discipline is mechanical. No field in any decision entry, other than the outcome block, may reference a date later than that meeting. The outcome is written last and appears in no reason. Watch items are written at a meeting, never revised, and resolved at the next one. Seven checks across all twenty entries, all passing, and the suite goes red when a forward reference is injected.

One of those checks initially failed on a false positive, matching the string "0.1" from an outcome against an unrelated signal reading. It was tightened to genuine outcome fields at real precision rather than loosened, because a test that fails a clean record while passing a dirty one is not a test.

Where the inputs are anachronistic

Some inputs genuinely cannot be reconstructed as they stood. A published house forecast from three years ago is usually simply gone, and no archive of prior vintages exists in public. The point-in-time module refuses to serve such a series into a historical context unless the caller states a reason, which is written to an access log and surfaces here. No anachronistic input was admitted into any of the twenty historical decisions. The capital market assumptions in this report are current-vintage and are used only for the forward-looking question of what the policy portfolio is priced to earn, which is a statement about today and carries no point-in-time problem.

Two data limitations are recorded rather than worked around. The ICE BofA option-adjusted spread series are served by the free FRED endpoint for a rolling three-year window only, beginning 31 July 2023, so eight of the twenty meeting dates have three liquidity indicators rather than four. Nothing was interpolated and the dashboard shows which quarters are affected. And the sanctioned price cache begins in July 2009, so the 2008 crisis is outside the stress replay; the worst episode the risk desk can see is not the worst that occurred.

Assumptions, and where the extract differs from the Statement

MANDATE.md is the working extract this office keeps beside the Statement. The two agree on every number this study depends on. All nine policy weights, all nine permitted ranges, the three sleeve ranges and eleven scalar limits are identical, which is checked mechanically by tests/check_mandate.py by parsing the extract and comparing it to the transcription of the Statement. There are no numerical conflicts.

They differ in coverage. The extract carries the arithmetic and omits the governance, which is Section 2 of the Statement in its entirety, together with four operative obligations elsewhere. Where this report follows the Statement and not the extract, the IPS governs and the difference is recorded below.

Modelling assumptions

The design system used

This report is set in the Coldbrook Capital design system, the house system shipped in ds/ in this folder: ds/colors_and_type.css holds the tokens and ds/preview/ ships the component previews. Colours are the token values from that stylesheet rather than a remembered approximation, and the components follow their previews. Three departures are recorded. The stylesheet opens with a Google Fonts import, which is dropped so these documents render identically from disk with no network. The stylesheet sets --navy: #0C1E48 while every preview and the written brand document use #0A1B3D, and the stylesheet's own --up token is #0A1B3D described as "same as navy", which it no longer is; the tokens file governs and the drift is reported rather than silently resolved. And the wordmark is Ashcroft's, since Coldbrook is the fictional firm the system was authored for and putting its mark on another institution's document would be wrong. There is no red or green anywhere in these documents: direction is carried by navy against clay, by parentheses on negatives, and by position relative to a zero axis.

Evidence appendix

Every claim taken from the literature, with a source a reader can open, marked according to whether the desk verified it in session or recalled it. The distinction matters: a training cutoff makes a recalled number a hypothesis rather than a fact, and several of these have moved.

40 claims from the Systematic desk, 32 verified in session and 8 recalled (80% verified).

Capital market assumptions, houses and vintages

The two pre-reconciliation allocations, as tabled

Quantitative desk rationale, as submitted: The desk recommends the policy portfolio. The composite signal has a pooled out-of-sample R2 of -1.61 per cent against the expanding historical mean, negative on 6 of 8 lines, with a Clark-West t of -0.77. Across the five signals and eight risky lines, 9 of 40 cells are positive. On that evidence the estimator does not earn active risk, and the size of a tilt should follow the evidence for it rather than the size of the budget available to express it. The constrained allocation reported here is the mechanical output of the optimiser run at the full budget; it is the model's answer and it is wh

Macro desk rationale, as submitted: Overheat. Growth is expansion on the 30 June vintage (growth score 4 of a possible 4: real GDP 2.09% annualised for 2026 Q1, payrolls averaging 188k over three months, a Sahm gap of 0.10pp, industrial production 1.67% year on year), and inflation is above target and firming (core PCE 3.41% year on year with a 3.52% three-month annualised run rate). The position that follows is not a growth call. It is a call on the price of two things. First, policy: the 3-month bill at 3.87% against core PCE at 3.41% is a real policy rate of 0.46%, below any plausible neutral, so the stance is neutral rather

What was verified against what was recalled

The Capital Markets desk fetched every house forecast in session and each carries a URL. Two inputs are marked recalled and stated as such at the point of use: the long-run real earnings growth assumption in the bottom-up equity cross-check, and parts of the historical context around prior CAPE peaks. The Systematic desk fetched the primary papers rather than summaries of them where the paper was reachable, and marks eight of forty claims as recalled. Where a desk could not reach a primary source it says which secondary source it used instead, which happened once, for Research Affiliates, whose own tool requires a login and whose figures were taken from Morningstar.

No paywalled source was used anywhere in this study and no credential was required. The one place a subscription would have improved the work is a clean forward earnings yield for the equity risk premium proxy, which the Macro desk records as out of scope by design and substitutes a dividend-yield construction for.

Where to find the rest

The five-year decision record, twenty quarterly entries with reasons, binding constraints and compliance results, is a separate document: decision_record.html. The interactive record is dashboard.html, which opens from disk with no network. The methods notebook, tying every method to the paper it comes from, is methods.ipynb. Every algorithm is a runnable Python file in taa/, the desk papers are in desks/, and the verification artifacts are in tests/.

Two supplementary notes sit in outputs/. Three things the office did that were worth watching records how six concurrent desks behaved as a system, with the clock time of each episode. The three decisions that earned the most is an outcome-selected analysis of the best three of the twenty, written with hindsight deliberately and labelled as such, and it does not change the recommendation.

Before any of this is published or quoted outside the Committee, read AUDIT.md. It records what in this study can and cannot be trusted: that the five-year record is a simulation of a rule rather than a track record, that the performance numbers cannot be separated from zero and change sign with the measurement window, and that the committee minutes are a construct whose reasoning is genuine and whose meeting is not.