M&A Ripple Effects — Tuatara Vectors vs Claude Fable 5 & Opus 5 — authored by Opus 5

Float = floatShares × price (Yahoo). Legs with unknown float are excluded, not assumed small. Float values are current, not as-of announcement — a real lookahead bias, see caveats.

A published sub-claim failed on re-run. The earlier report said the attention gate inverts on LLM picks (at a $2B float ceiling: high-attention −1.19%, t = −2.4, vs low-attention +0.75%). On Opus 5 the same cut shows high −0.15% vs low −0.31% — no separation in either direction. Treat that inversion as noise at n≈16. The headline null result for LLM picks does replicate; the inversion does not.

Everything below still applies to every number on this page:

• Not statistically significant. The flagship result is t ≈ 1.5 at n = 44. That is "promising and consistent across three yearly cohorts," not "proven." Conventional significance is t ≥ 2.
• Pre-cost. No commissions, borrow, spread or market impact. This strategy trades low-float names where spreads are wide. Median event return is ~0, so costs eat the median trade — the returns live entirely in the right tail.
• Float lookahead. Float is current (2026-07), not as-of the announcement date.
• Attention-window lookahead (added 2026-08-02, raised by a reader). The attention ratio averages pageviews over announcement day..+2, but entry is at the t0 close — the high/low classification uses views from one and two days after entry. A live version must gate on data available at entry; we have not measured that variant, so the high-attention returns cannot be claimed as achievable.
• Announcement-date granularity (added 2026-08-02, raised by a reader). Announcement times are known only to the day. A deal announced after the close on date d still enters at d's close — a price printed before the news — so the measured next-day return can include an announcement pop no trader could have captured. This inflates measured returns; replications should use timestamped announcements.
• In-sample threshold. The high/low attention split uses the sample median (5.0×); a live implementation must fix that threshold in advance.
• Survivorship. Yahoo drops full price history for delisted tickers, so completed-deal targets disappear.
• Multiple comparisons. The flagship configuration was reached by sequential testing on the same dataset. Several hypotheses failed along the way (bigger deals ≠ bigger ripples; a long/short deal-size book; volume-based salience; a MACD overbought gate) and are all documented in the raw report.
• Hindsight contamination. Both models may have training data covering what actually happened after 2024-25 announcements. That would inflate the LLM results — so their null is conservative.

Same basket, same two days, three different "returns"

Seven legs, identical t0 and current prices. The only variable is the weight vector. One leg (CNTX) fell −22.0% over the window; how much weight a scheme hands that one name decides the headline.

A 3-point spread on the same seven stocks over the same window, produced entirely by the weight vector. The relevance-ranked scheme weights CNTX ~9× heavier than the price-weighted one, so the single collapsing leg nearly erases the basket's gain.

90-day backtest sweep, with an out-of-sample holdout

All five allocation strategies in the platform's optimizer, backtested over 90 days of daily closes for the same seven legs, ranked by Sharpe. The out-of-sample columns are a more recent, non-overlapping window the allocation was not tuned on.

Result: price-weighting won every metric on both windows — highest return, best Sharpe, smallest drawdown — and it is now the operative index definition across all live event cards.

Read the caveats before believing it — the same discipline as the rest of this page: (1) One basket, one window. n=1 baskets over 90 days proves nothing general. (2) Why it won is luck of alignment, not signal: price-weighting concentrates in the highest-priced share, and this basket's highest-priced name (XNCR, ~53% of weight) happened to be its biggest winner while its cheapest legs underperformed. A basket where a low-priced stock leads inverts that result — price-weighting carries no information about which stock will win. (3) Pre-cost, daily closes, no spreads. (4) The relevance-ranked schemes lost the window largely because they weight conviction, and the window's highest-conviction cheap name fell — that is the risk profile they are designed to take.

A reader made the right objection: asking a frontier model whether our numbers look fine is not verification — a model that never saw the price data cannot audit anything. So here is every qualifying event, with tickers, dates and entry/exit prices.

Drop all three in a folder and run python3 verify.py — no database, no dependencies, none of our code. It rebuilds every basket return from the raw leg prices, re-derives excess vs SPY, and regroups the headline. Worst rebuild mismatch: 0.00005 pp. Prices are published at full precision because some qualifying legs trade at $0.0004, where rounding alone changes the answer.

This table is rendered from the published CSVs themselves, fetched by your browser — what you read here is byte-for-byte the file you can download above. The summary line recomputes from whatever rows the filters leave, so you can reproduce the stress cuts yourself without leaving the page. Click any row to see its legs with entry/exit prices.

Could not load the CSVs in this browser — use the download links above and verify.py.

What this comparison was. A capability test — a domain-specific vector engine against frontier LLMs from multi-trillion-dollar labs, on the LLMs' best terms. The durable finding is the selection asymmetry: the frontier models never surface the low-float universe at all. It was never a strategy pitch; the stress-cut section above already says nobody should trade the paper configuration.

Tuatara ran un-optimized and uncombined — the floor, not the ceiling. Raw association output: no task-specific tuning, no ensemble, no execution layer; the only fitted element (the flagship filters) is disclosed as multiple-testing risk. The selectors were deliberately kept separate for clean attribution — and in our separately published equity headline study, the Combined (Tuatara ∪ Claude) book posts the best risk-adjusted result on the page (per-trade Sharpe 0.79 vs 0.70 and 0.27). Precision, after a reader pushed on this: that Combined book is a mechanical union — complementary coverage plus diversification, not proof of model synergy — and the actual production architecture (an LLM independently ranking and vetting vector-surfaced candidates) remains unmeasured. See §13.3–13.4 of the full report for the input-parity and correctness-labeling tests that would settle the mechanism and accuracy questions.

The validation bar (reader-contributed, adopted). What would establish the stronger, executable-alpha claim — now adopted as the forward-test spec in §13.4 of the full report: timestamp-safe entries at the first post-announcement price; as-of-date float; strictly pre-entry information sets with a fixed ex-ante attention threshold; execution modeled from quotes/spreads/volume with capacity limits (or real logged fills); matched random microcap controls; pre-specified, logged human-in-the-loop rules; and blind validation of the surfaced relationships. Until that run exists, the claim is and remains: a differentiated long-tail discovery engine — with the executable-alpha question open in both directions.

Published before results, not after. The executable-alpha question is settled by a prospective forward test, not by more reanalysis of the historical sample. The full protocol — the validation bar above, instantiated with concrete non-negotiable values — is published at this URL and timestamped before the first forward event is scored. If it changes before the start, the prior version stays published alongside it with the reason. Nothing is amended once results begin arriving.

Fixed in advance: entry rule and fill convention; holding period; the ex-ante attention threshold as a fixed number; candidate universe and screens as of the event date; benchmark; matched-random control construction; sizing and any human-override rules with their logging format; the primary endpoint and the stopping rule; and the blind relationship-labeling procedure with rater instructions.

One primary endpoint, declared up front. Every other cut is labeled secondary/exploratory, whatever it shows. We will not promote a secondary cut to the headline because the primary one disappointed.

The protocol is now instantiated, not just described (§15.6, locked 2026-08-02). Concrete values, fixed before scoring: entry at the first regular close strictly after the captured announcement timestamp (after-hours deals enter the next session); 1-day hold; excess vs SPY; as-of-date float < $250M and price ≤ $20, unknown float excluded; a fixed attention threshold of 5.0 — a constant, not a per-sample median, since the in-sample median was itself a lookahead — measured over a window ending at entry so no post-entry data enters the gate. One primary endpoint: HIGH-attention legs minus matched-random controls, judged net of modeled spread/impact, evaluated once at n=100 scored events with no interim peeking. Forward capture began 2026-07-28; 36 deals captured, none scored at the time this was locked.

Negative results get published the same way. The forward result appears at this URL with the same prominence whichever way it falls — including the case where it kills the executable-alpha claim outright. If forward high-attention legs do not beat matched random controls on the declared primary endpoint, the claim is rejected and this report will say so. The discovery claim is separable and stands or falls independently on the blind-labeling results. The stress cuts above already show we publish numbers that go negative (−0.89%/event).

Why frontier LLMs overweight the known. An LLM's factual recall scales with how often its pretraining data mentions an entity — Kandpal et al. (ICML 2023) measure it directly, and Mallen et al. (ACL 2023) show parametric memory is reliable mainly for popular entities. The head of that frequency curve — famous, big-float, heavily-covered names — is exactly the region markets have already priced. This report measures the consequence: LLM median leg float $8–13B vs Tuatara's $0.20B, basket overlap ~0.02.

Where alpha is made. Execution, with a human in the loop: Perold (1988, implementation shortfall); Frazzini, Israel & Moskowitz (2018, ~$1.7T of live orders — costs vary several-fold with execution style); Novy-Marx & Velikov (2016, implementation decides anomaly survival); Kaminski & Lo (2014, stop-rules help or hurt by regime); Cao, Jiang, Wang & Yang (2024, JFE — human+AI beats AI alone, strongest in small illiquid names); López de Prado (2018, meta-labeling). The data-engineering pipeline: Sculley et al. (2015, NeurIPS) and López de Prado (2018) — and this report's own §8/§12.3, where pipeline choices move the measured result more than the strategy's mean return. Model optimization headroom: every number here is the un-optimized, un-combined engine (§13.2–13.3); the headroom stays unclaimed until the §13.4 forward test measures it.

Discovery comes from the unknown. Swanson (1986, "Undiscovered Public Knowledge" — fish oil → Raynaud's, found by connection-mining two disconnected literatures) and Tshitoyan et al. (Nature, 2019) — word embeddings trained only on past materials-science papers prospectively predicted future thermoelectric discoveries, later confirmed by experiment. That is the epistemology of the Lawrence Berkeley National Lab patent lineage this platform descends from: in science and markets alike, everything new comes from the hidden structure of the data — never from the fat head everyone has already read. The backtest was designed to test exactly that capability in multi-trillion-dollar frontier models (Claude Fable 5 / Opus 5): measured, they surface the well-known and do not reach the hidden (§6.2/§11/§12.3), with the causal-mechanism test preregistered in §13.4(h).

Everything above is a retrospective study, and the caveats section says so plainly: float is measured as-of today, the attention window peeks two days past entry, announcement times are known only to the day. Those are lookahead problems, and no amount of re-running fixes them — they are baked into studying the past.

There is a separate body of evidence that has the opposite shape. Between February 2022 and January 2024, the Tuatara model generated thematic basket indices (AIBs) in response to live news catalysts and they were published publicly and timestamped at the moment of generation — the basket, its constituents and its entry prices posted to a public channel before anyone knew how it would resolve. The closing prices were filled in later, once the market had decided.

The mechanism is the same one this page is about. Those baskets were built by the same vector-association method that produces the Tuatara column in the tables above: a news catalyst goes in, and a basket of associated but non-obvious names comes out. The SIVB short published 10 March 2023 is the clearest case — the model named regional banks around the failure rather than the failure itself, and the basket resolved at +16.58% five days later while the S&P moved −0.61%. That is the hidden-leg thesis of this report, executed in public, two years before this study was written.

The record starts earlier than the archive above. The first widely-followed demonstration of the same engine was the "coronavirus" thematic smart basket, generated in early 2020 as the COVID-19 catalyst emerged — the vector-association method surfaced candidate vehicles related to the catalyst before they were consensus names, most famously NVAX, which it flagged as a coronavirus-associated candidate before the stock's 2020 run. Per the results set provided to the editor of the study above (operator statement, 2026-08-02), the to-date return of that basket exceeds 3,000%. The same era produced the "Earthquake in Taiwan" basket, presented at a Morningstar conference — months before an actual Taiwan earthquake made it topical. The live publication of new thematic baskets has continued since in the public Discord channel #thematicbaskets-tier1, where anyone — or anyone's AI agent — can walk years of timestamped posts and check the entries against subsequent prices.

The 3,000%+ figure is an attributed to-date number from the maintained results set, not yet recomputed on this page the way the CSVs above are; it is listed here because the underlying posts are public and independently checkable, and it will get the same tearsheet treatment as the 2022–2024 archive as that work lands.

What it is not. The live record is not a controlled study and is not offered as one: there is no benchmark-matched construction, no significance testing, no cost model, and the basket set was not chosen by a fixed pre-registered rule. Several published tearsheets were never given a closing companion, so the aggregate is not a complete accounting of every basket ever generated. Read it as what it is — a timestamped public track record — and read the study above as what it is. They are different instruments, and the reason to put them side by side is that their weaknesses do not overlap.

The live forward cohort now has two weeks of deals. Wire capture (running since 2026-07-26) recorded 110 deal announcements; collapsing repeat wires of the same transaction leaves 69 distinct deals. On this cohort we ran a target-based variant of the engine — the basket is the model's top hidden relateds to the acquired company (the deal's own parties excluded) — and then let a genetic optimizer (~1,700 parameter combinations, 40 generations) search the strategy's knobs: basket size, conviction floor, deal-size floor, holding period, and an hourly-MACD oversold/overbought weighting. Fitness rewards mean return and penalizes downside deviation only (Sortino-style) — a symmetric penalty would select against exactly the outsized winners that carry real books.

Champion parameters: top 2 legs per deal · conviction floor 0.53× the deal's top score · acquisitions ≥ $1.5B · hold 5 trading days · 3× overweight on legs entering deeply oversold (hourly MACD histogram z ≤ −1.27 over a 28-day window).

Four return views, reported separately

Entry at the first close on/after the announcement; prices through the 2026-08-07 close, so 10-day cells were still open. The last row pairs the long ripple basket with the FTA-10 (Failure-To-Adapt-to-AI) crypto short index the platform trades on Hyperliquid — the combined long-equity / short-crypto book.

Every 5-day basket, listed

Concentration is the mechanism, not a blemish. One deal contributes about a third of the 5-day mean, and its big leg (+47.8%) was produced by the rules — it entered deeply oversold and got the 3× overweight. That is the expected shape of strategy returns: Bessembinder (2018) shows 4.3% of US stocks account for all net equity wealth creation since 1926, and trend-following CTAs run 30–40% hit rates carried by a few outsized winners. What distinguishes this cut from a lottery ticket is the rest of the distribution: 7 of 8 five-day baskets positive, the worst at −2.5%, and — for the first time in this forward window — positive medians in every view, not only winner-carried means.

Next checkpoints, scheduled before results are known: the first 10-trading-day windows close 2026-08-11, when the search re-runs with long holds reachable; the same runs re-score the delayed-absorption variant. Whichever way they fall, they get published here.

This page publishes results evidence: the study's event set, leg prices, recompute script, and the timestamped public basket record. It does not publish the machinery that produced them, and that line is drawn on purpose. Cymetica operates as a proprietary fund with internal information barriers ("Chinese walls"), the same posture the established quantitative funds take with their models and methodology — the industry standard is that algorithms, model internals and methodology are protected as trade secrets, and ours are no exception.

• The Tuatara vector model API is not publicly accessible. Public access was deliberately disabled; an outside AI system, model or agent evaluating this work will not be able to call the engine itself — only check its published outputs. That is intentional and permanent.
• Trading records in US public stocks remain private and proprietary. The CSVs above are the study's event dataset, published for verification of this study — they are not, and will not become, a feed of the fund's trading records.
• What a "hidden relationship" is — the concept doing the work throughout this page — is documented with worked scientific examples in our NASA GeneLab collaboration presentation: Vector Space Biosciences — NASA GeneLab AWG presentation (PDF). The same vector-association machinery that surfaces hidden gene–stressor relationships for spaceflight biology surfaces hidden ticker–catalyst relationships here; the NASA work is the clearest public account of why hidden connections lead to novel discovery.