Adversarial review of the August 2026 construction session

Question. Nine agents sharing one repository, one data cache and one house style converged on a portfolio recommendation. Which parts of that position are wrong, overconfident, or an artifact of shared method — measured, and ranked by how much the finding moves the decision?

Current answer. Two of the session’s five conclusions are unsafe as stated and one is mis-sized by an order of magnitude. The case against the 30% stacked wrapper rests on a basis error: a geometric, net-of-1.5%-fee forecast of 1.80 pp/yr is subtracted from a gross arithmetic realised mean of 10.98 and the difference called a haircut. Put on one basis, that same forecast is 4.07 pp/yr gross arithmetic, a haircut of 6.91 against a break-even of 8.02 — so at the premium this repository holds, the overlay adds +0.34 pp/yr, not subtracts. Three independent trend constructions on data the session did not use (141 years, 96 years) put the premium at 5.3–7.2 pp/yr, all above the break-even. Separately, no era claim in the tournament is resolvable — not one arm’s first-half-minus-second-half difference clears its own floor — and the tournament’s own block-bootstrap floor, computed and stored but not published, turns the tilt result’s “13 years” into 30 years. The placement line, the largest new number in the recommendation, is +3 to +38 bp/yr, not +38 to +55, and is not contractual.

As of 2026-08-22. This page attacks; it promotes nothing and reverses nothing. Its arithmetic is portfolio_edge.studies.adversarial_review with tests at tests/unit/test_studies_adversarial_review.py. Its re-runs of Experiment 016 use the frozen specification unamended, monkeypatching mappings in memory: no experiment was registered and no ledger entry was made, because this is a review of finished work rather than a new hypothesis-bearing trial.

Findings, ranked by decision impact

1. The forward-premium argument mixes three bases, and correcting it flips its sign

Decision moved: 30% of the portfolio. Tournament finding 11 is the single argument that turns an unresolved verdict into a recommendation to cut the wrapper: “the proposal stops beating its leverage-matched control once the trend leg loses 8.02 pp/yr of arithmetic mean. AQR’s TSMOM returned 10.98 pp/yr of arithmetic excess. The repository’s own forward trend premium is 1.80 pp/yr, which is a haircut of 9.18 — past the 8.02 break-even.”

Where 1.80 comes from. Decision 0004 cites “a post-publication trend excess return this repository measures at roughly 1.8 pp/yr”, with no experiment reference. Computing directly from the pinned AQR workbook:

3.35 − 1.50 (Experiment 004’s stated management fee) = 1.85. The 1.80 is the same AQR TSMOM file’s last 168 months, taken geometrically and net of a fee. It is a 39%-subsample of the 427 months being judged, not an independent prior.

Three basis errors, each measurable. Comparing it with 10.98 subtracts a geometric net figure from an arithmetic gross one and calls the residue a haircut:

• variance drag 0.5 × 12.38² / 100 = 0.77 pp/yr;
• a 1.50 pp/yr management fee that the tournament’s arms already charge separately — RSST’s 99 bp expense ratio is inside the rule, so the fee is counted twice;
• and the point estimate is used with no error bar, when the 168-month subsample’s own 95% interval is [−2.67, +11.00] and therefore contains the 10.98 it is being used to overturn.

Restated on the basis the break-even was computed on, 1.80 + 1.50 + 0.77 = 4.07 pp/yr. The like-for-like haircut is 6.91, not 9.18, against a break-even of 8.02.

Re-running the frozen tournament at each candidate premium (trend_haircut_pp_yr against the same specification, nothing else changed):

Only the published figure, on its uncorrected basis, puts the proposal below zero. Every estimate on the break-even’s own basis leaves it positive. The correct statement is not “at the premium this repository believes the overlay subtracts” but “the overlay’s sign is decided in the fourth decimal of a premium nobody can estimate, and the repository’s own estimate leaves it positive by about a third of a point.”

That does not make 30% right. It removes the only argument the session had for a particular direction of error.

The error has already propagated. The trend-weight regret study built this session carries gross_premium = 0.0180 as one scenario at a 15% prior weight, labelled “a convention this page traces and does not inherit” — correctly sceptical of the number’s provenance, and still entering a geometric net-of-fee figure on a gross axis. On that page’s own basis the same scenario belongs at 4.07, and every regret surface computed with it is shifted toward zero trend.

2. Not one era claim in the tournament clears its own resolution

Decision moved: the second pillar of the case against the wrapper, and the standing of the tilt result. Tournament finding 2 says the wrapper’s whole contribution is pre-2008 and “a full-window figure for these arms describes 1990–2008 more than it describes the future.” AGENTS.md requires that “before interpreting a null, compare the effect of interest with the design’s resolution.” The era table was never so compared.

An era gap is a mean on a sub-sample, so its floor scales as 1/sqrt(n); the difference between two halves has twice the full sample’s standard error, so its floor is exactly 2 × the full floor. Using the tournament’s own stored block-bootstrap floors:

Nineteen arms, nineteen unresolved era differences. Every one of the twelve trend-bearing arms’ post-GFC gaps sits between 3.5 and 7.5 floors inside noise. The era table is a table of nineteen numbers none of which this design can distinguish from its own full-window estimate, and it was read as if it distinguished all of them.

The asymmetry is the finding. proposal_no_trend’s post-GFC gap is +0.41 against a floor of 1.04 — the tilt arm is unresolved on the last seventeen years by exactly the standard applied to trend, and its first-to-second-half decay (+1.24 → +0.32, a fall of 74%) is proportionally larger than the difference between the trend arm’s halves as a share of their own floors. The session read one era table as evidence of decay and the other as evidence of persistence. Neither is either.

Assumption stated: the sub-period floors above scale the full-window floor by sqrt(n_full / n_sub), which assumes the gap’s volatility is constant across eras. Where it is not, the floor is larger in the volatile era and smaller in the quiet one; the 1990s–2000s were the volatile half for every trend-bearing arm, which cuts against the session’s reading rather than for it.

3. The published detection floors use a different inference basis from the intervals beside them

Decision moved: whether the tilt result is resolvable in a human holding period. The tournament’s intervals are stationary block bootstrap, mean block 12 months — the right choice for autocorrelated monthly gaps. Its MDE is the i.i.d. sd/sqrt(T). Both are computed and both are stored in the run artifact (mde_80pc_power_pp_yr, mde_80pc_power_block_bootstrap_pp_yr). Only the i.i.d. one is published, including in the headline.

Two of the three arms the tournament reports as separating from their benchmark do so only under the narrower inference. The min-variance arm loses the property outright. The tilt arm — the recommendation’s load-bearing result — clears by 0.07 pp/yr rather than by 0.32, and needs thirty years of holding, not thirteen. Its implied t falls from 4.73 to 3.05, which is exactly the hurdle Harvey, Liu and Zhu propose for anything drawn from the published factor family. A result that arrives precisely at the deflation threshold is not a result that supports “the tilts are the only resolvable edge” without qualification.

4. The placement line is +3 to +38 bp/yr, not +38 to +55, and is not contractual

Decision moved: the largest single line in the recommendation. The arithmetic in structural and tax edges §8 reproduces exactly — every priority, the §901 credit cap, the two-bucket yield split, all twelve cells. I could not break the calculation. What does not survive is what is booked from it.

• The range is truncated. §8.6 states plainly that one input decides whether placement is worth 6.5 bp/yr or 38, and the ledger’s own range column prints +2.8 to +38.2. The headline, the conclusion list and the +38 to +55 bucket all use +21 to +38 — bracket uncertainty only, with the entire distributed reading dropped. −18 bp on the floor.
• 94.7% of the +38.21 is one holding. Decomposed by fund: RSST +36.18, IDMO +2.47, DFIV +2.12, VEA +1.87, AVES +1.40, IEMG +1.07, AVLV −3.51, VTI −3.39. The 36.18 is 40.8% charged on 8.3% of NAV of undistributed controlled-foreign-corporation income from a single fiscal year with a 19.94% total return. A RIC shareholder is not taxed on undistributed investment-company taxable income; the page’s own independent check (prospectus 17.17% pre-tax against 16.85% after tax on distributions, a 32 bp gap) is evidence for the distributed reading. Scaling that yield by 0.75 / 0.5 / 0.25 gives +29.75 / +21.28 / +12.82 bp.
• The +14 for rebalancing inside the shelter is not incremental. Under the pro-rata control every fund is two-thirds sheltered, so any rotation smaller than two-thirds of a position is also executable with no realisation. It is a property of having 66.7% shelter, not of optimising location. §8.6 concedes the plan makes rebalancing harder in one direction. −14 bp. It is also booked against §4’s explicit instruction that the figure is “a hurdle, not a saving. Double count: not additive.”
• +14 and ≤+2.8 are mutually exclusive. Lot selection pays only on realised gains; the +14 is the value of realising none. −2.8 bp.
• Three benchmarks are summed in one line, two paragraphs after the same section warns that lines measured against different benchmarks are not added — the error aggregate() raises on.
• The optimum may not be reachable. It requires 30% of total net worth in RSST inside the traditional account. Traditional balances are typically employer plans with closed fund menus, and RSST/MATE/JPFP appear on none. That single line is 36.18 of the 38.21 bp. The page checks the Roth contribution phase-out and never checks whether the traditional account can hold the wrapper at all.
• What is right, and worth recording: the nominal-versus-after-tax objection is not present. Rebuilding the whole thing in after-tax dollars (traditional discounted at 24%) gives 40.88 bp of after-tax wealth against the published 38.21 — the nominal treatment is mildly conservative, because the saving accrues in the taxable account.

Defensible range: +3 to +38 bp/yr, sensitive to a tax-character reading and to whether the traditional account can hold the fund. The claim that placement is worth more than the overlay survives only at the top of that range.

5. Eighty-six per cent of the tilt edge is measured on the repository’s shortest series

Decision moved: how much weight the one “resolvable” conclusion can carry. Stripping each fund’s non-market loadings from proposal_no_trend one at a time and re-running the frozen tournament:

US tilts contribute +0.125 pp/yr; non-US tilts contribute +0.683. The 1990-start problem the session named for trend applies with far more force to the tilts: the whole non-US factor evidence in this repository is 427 months with zero out-of-sample, and it carries 86% of the only result the session calls resolvable. The US value premium — the one with a century of history — was +2.16 pp/yr over this window with t = 0.95, and contributes about 10 bp.

Trend, by contrast, can be tested outside 1990–2026 and was not. Two independent constructions using this repository’s own frozen time_series_momentum:

The tournament’s window is not the problem for trend. The Sharpe ratio of an independently built book is 0.58 before 1990, 0.58 after 1990 and 0.58 over the whole 96 years — the 427-month window is representative, and the trend premium is resolvable once the instrument is long enough (t = 5.5 and t = 6.6). What the window cannot resolve is the portfolio gap, which is a different estimand. Restated at AQR’s 12.38% volatility these give 7.18 [4.61, 9.75] and 5.32 pp/yr — both above the 2.96 break-even, and the monthly book’s entire 95% interval is above it.

Charged and unflattering: the own book runs 262% annual one-way turnover; at 20 bp one-way its mean falls from 3.64 to 3.12 and its Sharpe from 0.58 to 0.50. Four legs against a vendor’s fifty-eight understates by the breadth identity; annual data with a one-year lookback is a coarser signal than monthly. Read the two as a bracket.

6. The alpha-charging rule is dominated by the estimator it should have used

Decision moved: whether the tilt result survives a hostile arm. Two agents charged a fitted alpha when its estimate exceeded its own detection floor. Exactly one fund qualifies. That rule charges whatever was measured precisely and forgives whatever was not — it conditions on the estimator’s variance, not on the truth, and it takes a single extreme estimate at face value at the exact moment the winner’s curse is largest.

The standard alternative shrinks the whole cross-section by its own precision: tau² / (tau² + s_i²) with tau² from moments. On the five measured alpha − pedestal figures and their HAC standard errors:

The threshold rule is harsher than the statistically defensible treatment — and it is harsher for the wrong reason, because it takes −3.80 whole while zeroing −1.66. Under shrinkage the tilt result survives the published floor and fails the bootstrap floor; the honest summary is that the tilt finding is fragile to the alpha treatment under every inference basis except the pair the session happened to publish.

Two asymmetries make even that generous. DFIV’s −3.80 pp/yr is not credible as a fund property for a 27 bp index fund; it is far more likely absorbing a loading error, which means it is evidence about the mapping rather than about the fund. And the wrappers’ alphas were never measured, so charging measured alphas charges only the tilt side and gives RSST, MATE and JPFP a zero implementation shortfall by absence of measurement.

7. The market betas were measured, then discarded, and the specification says otherwise

Decision moved: levels, not the ranking. Experiment 016’s specification states that “a market beta of 1.000 is an assumption everywhere: no market beta was measured for any of these funds.” That is false. Every tilt fund’s market beta was fitted, with a HAC standard error, in Experiments 009 and 013, on the same French files with the same sha256: AVLV 1.0137 (0.0274), DFIV 1.0547 (0.0261), IDMO 0.9540 (0.0385), AVES 1.1766 (0.0346), VEA 1.0498, VTI 0.9966. The betas were not unmeasured; they were unrepresentable — src/content/shelf.ts’s FactorLoading union has no "MKT" member — and the perturbation grid then protects them by name (_MARKET_LEGS).

Re-running the frozen tournament with the fitted values substituted:

The trend contribution moves by 0.016 pp/yr across all of it — the ranking is invariant. The tilt level moves over +0.787 to +0.984, a range of 0.197 pp/yr, which is 42% of the published floor and 27% of the bootstrap floor. The direction matters: the tilted funds have higher betas than the control, so about 11 bp of the +0.79 that an investor would actually realise is market beta rather than tilt — the β = 1 assumption manufactures a beta neutrality the real funds do not have, and is conservative for the level while being exactly wrong for the attribution.

Two coefficient sets were also silently truncated. AVLV’s fitted RMW +0.074 and CMA +0.118 were dropped though the panel carries both legs. AVES’s fitted SMB +0.367, RMW +0.267, CMA +0.303 and UMD −0.176 were dropped because the emerging panel has no such series — the emerging leg is structurally unable to carry its own fitted vector, and AVES is 12% of the tilt result.

8. Ninety-eight per cent of the portfolio’s active risk sits in one boutique fund

Decision moved: a concentration nobody counted. proposal_rsst runs 7.087% tracking error against the cheap control; proposal_no_trend, the same portfolio with the wrapper replaced by plain US beta, runs 0.999%. The wrapper therefore carries 1 − (0.999/7.087)² = 98.0% of the portfolio’s active variance at 30% of its capital.

Under the placement optimum that 30% of total net worth sits in a single ~$313m fund from a boutique issuer, inside the traditional account, whose N-PORT structure JPFP has never filed and whose own sister page records that 13 of 25 managed-futures funds (52%) filing at the window’s open stopped filing within 6.5 years. Issuer shares of the proposal: Return Stacked 30%, Vanguard 30%, Avantis 20%, Dimensional 10%, Pacer 5%, iShares 5%. The five-year closure and merger risk on 98% of the active risk is a cost that appears in no arm, no floor and no fee table.

9. The whole resolvable edge is worth about three points of savings rate

Decision moved: what the investor should spend attention on. Nobody asked what any of this is worth against the contribution rate, which is the one lever with no tracking error and no detection floor. Solving for the constant monthly contribution that buys the same terminal wealth as extra growth, at 10.8%/yr:

The only conclusion in this recommendation that the data resolves is worth, over a full career, about the same as raising the savings rate by three points of the starting balance — one budgeting decision, available immediately, certain, and with no fund risk. That is not an argument against the tilts. It is an argument that the marginal hour of this project has been spent on the smaller lever, and it is the comparison the charter’s reference investor should see beside every sleeve number.

Read with its own caveat: the equivalence depends entirely on the ratio of contributions to balance. Early in accumulation the flow dominates by an order of magnitude; near the end it does not.

10. The timing-rule verdict rests on the underpowered half of its own evidence

Decision moved: conclusion 5’s scope, not its direction. Timing rules reaches “no timing rule at any weight, in any account” from a US test that is unresolved (+0.74 pp/yr, MDE₈₀ 3.03) — while its JST pooled test, the only design on the page with resolution, finds the effect (+0.97 pp/yr, HAC t = 2.74, MDE₈₀ 0.99, 148 years, 16 countries). Two further window mismatches: the −2.96 pp/yr after-tax cost is measured on 1990–2026 and juxtaposed with a +0.74 pre-tax gap measured on 1926–2026, where the matched post-1990 pre-tax gap is +0.35 (t = 0.24); and roughly 0.73 of the “1.92 pp/yr tax cost of the rule” is the cost of holding 27% in ordinary-rate bills and rebalancing monthly, which a static blend at the rule’s own average weight incurs with no timing signal at all.

The recommendation is probably still right — the overlap with the wrapper (ρ = 0.566) and the deflation result carry it in the sheltered accounts, and the tax arithmetic carries it in the taxable one. But the verdict’s scope is set by the instrument that could not see, not by the one that could.

What I tried hardest to break and could not

• That trend decayed after 2008. It did, and it is not a vendor artifact. My own 4-asset book runs Sharpe 0.87 over 1990–2008 and 0.17 over 2009–2025; the JST book runs 0.55 pre-1990 and 0.19 over 1990–2020. Three constructions, three sources, same shape. What the century of data adds is context, not a rescue: the 1960s Sharpe was 0.07 on the same book, and the decade that followed was 0.80. A fifteen-year dry spell has happened before and did not end the effect — which is why the decay is not resolvable, not why it is not real.
• The vendor-cost objection to AQR’s TSMOM. I expected the gross-of-trading-costs index to flatter badly. On the only window where it can be checked, live managed- futures funds net of their own fees returned +2.84%/yr against the index’s +1.95% — the funds beat the index. The 7.7 pp/yr CTA-bias scenario is well outside the measured interval. Live managed futures already says this and is right.
• The basis mapping. I expected β = 1.000 to be load-bearing. It moves levels by up to 0.20 pp/yr and the ranking by 0.016. Attack 7 above is what survived; the rest failed.
• The symmetric-haircut argument. I expected that haircutting the value premium the way the session haircut trend would sink the tilts. It does not: realised US HML over 1990–2026 was +2.16 pp/yr (t = 0.95), below the repository’s own pooled post-publication estimate of +4.74, so the “haircut” is an increase and the tilt gap rises from +0.79 to +1.14. On the US-only post-publication figure of +1.57 it falls only to +0.71. The tilt result does not depend on the US value premium — which is finding 5, arrived at from the other side.
• Sequence-of-returns risk for an accumulating investor. Recomputing every arm’s terminal wealth with constant monthly contributions rather than a lump sum, the money-weighted advantage over the cheap control moves from +0.876 to +0.825 pp/yr for the tilt arm and from +4.440 to +4.170 for the proposal at contributions up to 2% of the starting balance per month. The ranking is invariant to the contribution rate. The accumulator objection is real for the size of the prize (finding 9) and empty for the choice.
• The placement arithmetic itself. Every priority, the §901 credit cap, the two-bucket yield split and all twelve cells reproduce exactly by hand. The foreign-tax credit is not double-counted. The pro-rata control is the right control. Finding 4 is entirely about what is booked from a correct calculation.
• 65/35. Global cap weight is ~64% US on a single 2026-06-30 read, and the same page’s own arithmetic puts a 10 pp shift at 14 bp/yr against 80–144 bp of tracking error — 55 to 178 years to resolve. 65/35 is unfalsifiable here, and so is 60/40 and 70/30. This is the safest of the five conclusions precisely because nothing is being claimed.

Which conclusions are unsafe

Scope and limitations

• Instrument: all tournament re-runs use Experiment 016’s frozen specification, its pinned source bytes and its own code paths, with mappings replaced in memory. The published central case reproduces to four decimals in every re-run, which is the check that the substitutions are the only difference.
• Not registered. No experiment file, no ledger entry, no promotion. This reviews finished work; it does not open a hypothesis. run_kind would be exploratory if it had one, and the look-ahead is total — every number here was chosen after reading the session’s results.
• The long-window trend books are lower bounds on breadth and upper bounds on cleanliness. Four and thirty-six legs against a vendor’s fifty-eight; no financing, no market impact, no capacity; annual data for JST with a one-year lookback. Costs are charged only in the sensitivity row.
• The sub-period floors are scalings, not measurements, and assume constant gap volatility across eras. Where that assumption fails it fails against the session’s reading, not for it.
• Empirical-Bayes shrinkage uses the raw HAC standard errors from Experiments 009 and 013 as proxies for the standard errors of the pedestal-adjusted alphas the tournament charges. The two differ; the pedestal is itself estimated, so the true errors are larger and the shrinkage stronger than shown.
• The placement findings are an independent re-derivation of §8’s arithmetic and of what is booked from it. They are not a new tax measurement, and the character question (recognised versus distributed) is unresolved here as it is there.
• Not examined: human capital and job correlation; the operational cost of holding nine lines across three brokerages with partial shares and settlement; what happens if the investor stops contributing; estate and §1014 interactions beyond noting that the placement optimiser scores annual distribution drag only and never asks which asset should receive the step-up — a term §4 puts at 78 bp/yr at thirty years, 40× the boundary §8.5 does price.

What would change these findings

1. A stated prior on the trend premium with an interval, arrived at without touching the 1990–2026 AQR file. Finding 1 exists only because the repository’s forward premium is that file’s own tail, on a different basis. The two independent constructions here are a start and are not a forecast.
2. Publishing the block-bootstrap MDE beside the block-bootstrap interval, or stating in one place why the i.i.d. floor is the right companion to a bootstrap interval. Finding 3 is a reporting choice, not a modelling error, and it is one line to fix.
3. A market-beta field on the shelf. Finding 7 is downstream of a TypeScript union with no "MKT" member; the numbers already exist in two committed artifacts.
4. A pre-1990 international value or momentum series. Finding 5 is currently untestable in this repository: nothing here can put DFIV’s and IDMO’s factor legs outside the single window they were measured on.
5. Resolving RSST’s tax character, and checking whether a traditional account can hold the wrapper at all. Those two facts move finding 4’s range from a factor of thirteen to something a decision can use.