ViskaStrat · Study · 2026-07-27 · internal — for the decision-layer pitch

Was the warning in our own briefs?

The conviction engine scores every theme a broker touches on a signed bullish–bearish scale; the Mímir morning brief renders that record to the fund. This study joins the engine's conviction record against the fund's own AI-sleeve tape for June 1 to July 24, 2026 — and separately checks, day by day, which of those readings the delivered brief actually rendered (full-archive cross-check, ViskaRes 2026-07-27).

Verdict, in three layers: the engine generated the warning in time; the delivery layer dropped part of it; and no rule forced action on the part that did arrive.

1 — Delivered. The decisive warnings reached the inbox: the morning of the peak (Jun 30) the brief rendered the month's heaviest read — 0.12, 10 brokers, CONTESTED — and the Jul 1 brief rendered a second heavy contested read (0.17, 9 brokers). The fund still added $517k on Jul 1. An add-gate on delivered data alone blocks that add (≈ $31k saved).

2 — Generated but not rendered. The engine had scored the deterioration by Jun 25 (0.17, 6 brokers, contested) — but the briefs of Jun 22–26 carried no conviction block, so the fund entered its $3.15M Jun 29 re-entry having been shown only a wall of thin-breadth bullish reads. On the engine's own record, the evidence-weighted gate blocks both re-entry days, avoiding −$244k of realized losses.

3 — The fund's own exits needed neither. Jun 17–18 (+$533k banked) and the fast Jul 6 cut both led the signal. The manager's exit instinct is not the problem; the entry discipline is. And the gap between $31k and $244k is the cost of consuming this signal as an email instead of as a decision input — which is precisely the decision-layer proposal.

The three layers, one clock

Signal shown on the day the underlying broker material is dated; when rendered, it reaches the fund in the next morning's brief. Dots marked × with dashed outline existed only in the engine — the next morning's brief carried no conviction block (Jun 16, 22, 23, 25). Gaps are absent coverage, not neutrality. Hollow dots = contested. The index is flow-neutral: it moves only on price, never on the fund's own buying or selling.

Event by event

When	Tape	What the fund did	What the signal said (available that morning)	Who led
Jun 9–11	BUY +$2.89M	MU, SNDK, STX, WDC, TSLA — the fourth build of the year	No AI read since May 31 in the engine; briefs carried no conviction block at all before Jun 18. Rules abstain.	—
Jun 17–18	SELL −$3.15M	Full exit of MU/SNDK/STX/WDC + TSLA, realized +$533k	Engine record bullish (0.50 → 0.61, corroborated) — the Jun 17 read was never rendered; the first delivered block (Jun 18 brief) read 0.40 bullish. Either way the fund sold against it.	Fund led. Its own instinct beat the signal here.
Jun 25–28	—	Sleeve nearly flat ($0.4M remnant)	Engine scores the deterioration: 0.17, 6 brokers, contested, 11 snippets — the heaviest evidence weight of the month to that date (Jun 25, four days before the $3.67M re-entry) — but briefs Jun 22–26 render no conviction block. The fund is shown only thin bullish reads (0.60 on 1 broker, twice).	Warning generated — not delivered.
Jun 29	BUY +$3.15M	Re-entry in 5 names at the exact index peak window, 4–6% above own exits	Delivered picture that morning: 0.31 on 4 brokers, uncontested, after a wall of thin bullish prints — delivered 5-brief mean 0.48. Delivered gates do NOT block. Engine record (with the dropped Jun 25 print): mean 0.35 falling from 0.54 — engine gate blocks.	The costly decision — and the rendering gap's price.
Jun 30	—	Index peaks at 123.7. Year-high 35.3% AI concentration	The brief renders the month's heaviest read: 0.12, 10 brokers, CONTESTED — behind it, JPM's scored "SHORTS — AI Vulnerable".	Warning delivered, position already on.
Jul 1	BUY +$0.52M	Adds 500 MU on the first break day	The brief renders a second heavy contested read: 0.17, 9 brokers, CONTESTED; delivered 5-brief mean 0.21, down from 0.48. Both gates block, on delivered data alone.	The indefensible add.
Jul 6	SELL −$2.72M	Unwind, realized −$246k on the day	Delivered read 0.28/3 brokers contested — the full corroboration collapse (0.06) printed in the engine the next day.	Fund led again. Exit was fast and right.
Jul 13	trim −$0.75M	GOOGL/TSLA remnants trimmed, +$8k	Signal recovered (0.46–0.54) on thin breadth.	—

What was behind the delivered readings at the top

These are verbatim evidence spans from the scored broker calls underlying the Jun 30 and Jul 1 delivered readings. The brief renders the score, broker names, a CONTESTED flag and a one-line rationale — not these spans; they are what the engine scored to produce the numbers the fund saw:

"SHORTS — Span AI Vulnerable (ex-Software)" JPM, scored −0.55 bear on ai-infrastructure, material dated Jun 29 — in the Jun 30 brief, the peak day

"Mid-year 2026: take profits in S&P/secular growth, buy Large Value/cyclicals" BofA, scored −0.55 bear, dated Jun 30 — in the Jul 1 brief, the day the fund added $517k

"The railroads laid thousands of miles of duplicate tracks simply to block rival lines… a map of unprofitable ghost lines." UBS, scored −0.60 bear, dated Jun 30 — in the Jul 1 brief

Two statable rules, tested honestly

Both rules gate adds only — neither forces an exit, so no hindsight exit-timing is claimed. Both use only data available the morning of the trade (T−1 material). The cash from a blocked buy is assumed to sit idle — no alternative deployment is credited.

Gate on DELIVERED data only

Using strictly what the briefs rendered (score, broker count, CONTESTED flag): no adds when the latest delivered read has ≥ 6 brokers and is CONTESTED, or the delivered 5-brief broker-weighted mean is below 0.40 and falling.

−$31k avoided

Blocks the Jul 1 MU add ($517k lot: bought $1,035, sold Jul 6 at $973.35). Does not block Jun 29 — the deterioration evidence had not been rendered.

Gate on the ENGINE record

Same deterioration gate, computed on the engine's full scored record (snippet-weighted, including the reads the rendering layer dropped on Jun 22–26).

−$244k avoided

Blocks Jun 29 ($3.15M, mean 0.35 falling from 0.54) AND Jul 1. All Jul 6 realized losses on those lots never happen; the kept GOOGL lot's further ~$35k mark-down is also avoided (unrealized).

Against the sleeve's full-year realized of +$1.85M, the failed final round trip gave back about 13%. The engine-record gate keeps essentially all of it — without touching either of the fund's good exits. Two conclusions follow. First, the decision layer is complementary to the manager: his exits were better than the signal, the signal's entry discipline was better than his. Second, the $31k-vs-$244k spread is the measured cost of consuming this signal as a rendered email rather than as a machine-readable decision input — the delivery layer dropped exactly the five days of deterioration that would have armed the gate before the re-entry.

Caveats — read before quoting numbers

This is a one-episode backtest — and no longer scored series exists to validate on. The 0.40 threshold was chosen looking at this episode. The full-archive parse (ViskaRes, 2026-07-27) shows the conviction block first appears 2026-06-18: 29 scored days across 204 briefs. There is no Dec–May scored series to backtest against; the honest validation path is forward (paper-gate the rules daily from here) plus the archive's narrative layer, which spans all 7.5 months but carries explicit direction on only ~13% of bullets.
The delivery layer has holes, and they were load-bearing. Eight briefs inside the scored era carried no conviction block (Jun 22–26, Jul 4–5, Jul 18); no brief at all was issued Jul 12, 19, 26; no AI-complex engine read exists Jun 1–15 (rules correctly abstain on the Jun 9–11 build). A missing day is absent coverage, never neutrality.
Breadth thins late in the window. Broker counts fall to 1–2 by late July; a ±0.45 on one broker is not the same observation as on nine. Dot size in the chart carries this.
Attribution is conservative but not audited to the cent. FIFO lot math from the statement's own realized figures; the R1 lot loss is computed from lot prices (−$30.8k).
The signal was wrong where the fund was right. It read bullish through the Jun 17–18 exit that banked +$533k. The case here is for a signal-gated add rule, not for trading the signal outright.

Method & provenance. Tape: IBKR Flex statement as-of 2026-07-24 (sha-verified, ops task 040); trades deduped by tradeID, daily values from SUMMARY position snapshots, net traded = −proceeds × fxRateToBase; sleeve = MU, SNDK, STX, WDC, GOOGL, TSLA; index is flow-neutral, chained on names held across consecutive days. Signal: theme_corroboration view exported from the production conviction pipeline 2026-07-27 (task 047, 279 rows, sha f17e1418…), AI complex = ai-infrastructure + ai-memory + ai-power + datacenter-cooling, snippet-weighted — this is the engine record. Delivered-brief coverage cross-checked against the full-archive rendered-brief parse (ViskaRes research/signals/full/, 204 briefs, 2026-07-27): × dashed dots were never rendered; delivered-gate arithmetic uses only rendered fields (score, broker count, CONTESTED). Evidence spans verbatim from broker_theme_calls (702 rows). Tier: [live-verified] on tape, engine record, and rendered-brief coverage; counterfactual rows are stated-rule arithmetic, tagged inferred. Author: ViskaStrat. Not client-facing in this form. Rev 2 — corrects rev 1's "as delivered each morning" framing after the ViskaRes cross-check.