After Action Report  |  Proof of Concept 5

SPCX: Navigating Event Windows #1 and #2
with AI Agentic Analysts and Auditors

Window 1 asked whether the fee would keep rising into a release with no precedent. Window 2 asked by how much it would fall after one. The record answers both, and it does not answer them alike.

Ed Blount, Tom Daniels, Dan Hammond, and David Schwartz
Advanced Securities Consulting LLC

Section 1Executive Summary

Between July 30 and August 20, 2026, ASC's governed AI inference vector models issued twenty-two dated editions on the securities-lending market in SpaceX (SPCX) across the first two IPO lockup releases of August 6 and August 20. Each distributed edition stated its market-state read in writing, with the settled observable that would prove it wrong, before the resolving data existed. This report prints every edition in the order it was made, including the internal alphas that were never sent, the edition withheld as late, and the release-day edition that was held, and scores each against the September 3 settled record.

The working hypothesis the experiment began with was that successive releases could overwhelm demand and crush the stock. They did not. The stock rose, and through the first countdown the editions read HOTTER against a flattening tape, then turned to COOLER at T+1 on the day the re-rates came down. The second window was less clean. The editions carried a fixed inference box with a direction, a range and a stated confidence, the settled fee moved less than five basis points on most days, and an anomalous trade in the real-time blotter (2mm shares at 58%) was later corrected from a new trade print to an internal adjustment while the release-day edition was held pending reconcilement.

The distributed editions were produced by a governed framework: written instructions, specialist agents applying a human-defined mandate, an independent agentic auditor, and a human who decided what could be claimed. The quantitative models entered that framework as an input it evaluated each session. SpaceX priced on June 12 and the daily lending file for the name began on July 28, so the eight-model deep ensemble reached the first countdown with about three weeks of settled history, and the framework held its output below the validation gate rather than publish it, as the design had anticipated for a name this new. The same framework carried the seven Window 1 reads, wrapped the Window 2 quantitative output in an inference box with a stated range and confidence, scored the August 19 editions against the adjusted point after an anomalous blotter trade was reclassified as an internal adjustment, and held the August 20 edition on a tape whose manifest did not reconcile. Models of the same families, trained on a sufficient record, had carried the earlier proofs of concept, KLAR among them; the governance introduced here is what let the line run on a name where no such record yet existed.

What the two windows produced, beyond the scored record, is a working division of labor between the quantitative AI models and the language models. The quantitative rungs ran in ASC's shop. The agentic analysts held the corporate-action taxonomy, the Project Instructions and the context files of earlier events, read the model output against them, and wrote each edition's market state with its pre-committed conditions. The agentic auditor tested every edition for provenance, stale data, arithmetic and overstatement before a human released it. That division is portable. The instruction set and the context files can be furnished to a tester's own Claude instance; ASC runs the linear models and spins the non-linear models in its own shop and delivers the results through the inference summary record, so that the analyst and the auditor do their reading inside the tester's walls, against the tester's own tickets or a vendor's aggregates, under the same rules that governed these twenty-two editions. How the analysts read, what the auditor refuses, and what the ISR carries are the subject of the technical report and of the tester program that follows it.

Section 2Event Background

SPCX combined minimal trading history, an IPO of unprecedented size, a rare staggered release structure and observable stock-loan demand. Nine scheduled steps would pass before all shares became eligible to trade, stretching the supply transition across roughly a year. The first release fell on Thursday, August 6, 2026, and the second on Thursday, August 20; further windows were conditioned on price and earnings tests. Before any settled lending data existed for the name, the qualitative layer of the production line worked through the prospectus mechanics and produced the supply ladder that every later edition carried: which shares released on which date, by what mechanism, to which holder class, and which conditional windows would and would not open.

The data history is the fact that governs everything that follows. Live access to the daily securities-lending feed for SPCX began on July 28, when the team also requested backfill to May 1. Training runs on the name began the week the first countdown began. A contemporaneous note from the modeling lead on July 29, reporting the first model output, said that even the model itself cautions that it is not doing a good job with this thin a sample. The countdown method, written down in Proof of Concept 4 and demonstrated on KLAR in March 2026, ran from T-5 to T-0 with a fixed cadence: each session's edition stated a market state, a pre-committed condition under which the state held, and the settled observable that would falsify it.

Section 3Model Performance Evaluation

Which rung may publish: the fitness decision, session by session

The governance enhancement this proof of concept introduced to the quantitative models is a rung-fitness decision made in writing every session: which rung of the model ladder has enough history behind it to publish, and what the edition may say when the answer is none of the upper rungs. On SPCX the answer was known before the first window opened. It was never anticipated that a name that priced on June 12, with a daily lending file that began on July 28, would supply enough observations to train the eight-model deep ensemble, and the regression rungs were squeezed by the same short span. The ensemble was therefore held below its validation gate in both windows and its output entered no distributed edition. On August 13 the first inference box built from its output reported a direction confidence of 14 percent, and the two editions carrying it stayed internal. The July 31 edition capped its confidence at 75. The ladder was stopped at momentum, and the editions said so: the July 31 edition led with the option-implied forward borrow rather than a model forecast, and the August 3 edition printed sixty tickets of re-rate direction rather than a point estimate. This was controlled abstention, the system declining unsupported complexity rather than dressing thin history in fluent prose.

The evidence the fitness decision rested on is in the record. On the four sessions where a one-step forecast could be scored against the settled fee, the ensemble's errors exceeded the naive carry's, with mean errors up to 4.10 percentage points against carry errors of 0.05; on a fee with three weeks of history that barely moved between sessions, the carry's no-change forecast is the expected winner and the comparison says more about the sample than about the models. The linear rungs read the same way. A walk-forward replay of seven closed-form models on the SPCX new-fee series over 40 origins from June 17 to August 19 finds the best direction edge, Drift5 at 0.600 against a base rate of 0.525, inside its own binomial standard error, with persistence carrying the lowest level error at 40.3 bp and no stable ranking across data variants. No trailing-window model anticipates the August 7 reversal, which is the mechanism the editions carried in their pre-committed conditions and the reason the qualitative layer, not a rung of the ladder, produced the Window 1 reads.

This is a finding about data on one name, and the same model families have a record on names where the data existed. In Proof of Concept 1 the deep-learning ensemble outperformed ARIMA on public lending data, with more than 72 percent of its correct forecasts at turning points and traders acting on AI-predicted reversals accounting for over 60 percent of the 155,476 dollar net savings against the Goldilocks benchmark. Proof of Concept 2 validated the same model class against a custodian's own records on an agent lender's live book. Proof of Concept 3 captured 12.2 million dollars on Lennar in three sessions. In Proof of Concept 4, on the KLAR lockup of March 2026, the T-Series predictive models, benchmarked, calibrated and trained against the RPT corpus of confirmed historical events, identified a structural shortage when the market expected a supply deluge, and the outcome confirmed it: zero shares available and a 68.62 percent cost to borrow at the close of March 9. The models' measured record on the seasoned five-thousand-name universe the production platform serves every day is set out in the technical report; it is not summarized here because those figures await verification against the run manifests. What SPCX added was the governance that decides, on the record, when that history is present and when it is not.

Where the agentic layer wrapped the quantitative output

Gating was the smaller part of the enhancement. In four dated instances the qualitative layer took quantitative product and returned it governed. On July 31 a second AI instance independently reproduced the option-parity extraction, 790 plus or minus 126 bp on the front segment and a 666 bp front-to-back spread at 4.7 standard errors, so that the figure the edition carried had been replicated before it was published, and the agentic auditor required that it be stated as ASC's inference rather than as observed fact. On August 3, before any number was quoted, an agent's review of the option chains rebuilt overnight (from 935 kilobytes to 86 megabytes, and from four to fourteen strikes per snapshot to seventy-two to 193) found the zero-days-to-expiry rows contaminated and 186 of 2,456 rows with null open interest clustered on nine dates where a vendor endpoint had timed out, and specified the fix; a synchronized quote capture was running on SPCX every sixty seconds by 13:44. On August 13 the inference box, with its direction, range, stated confidence and pre-committed conditions, was specified to be built automatically from the daily prediction file, so that in Window 2 the quantitative output entered every edition already wrapped in the terms on which it would be scored. And on August 14 a simple tape estimator assembled from the screened ticket feed placed within 1.15 bp of the settled fee against the carry's 4.12, the one scored observation in either window on which any instrument beat the carry. Each is the qualitative layer inferring from, checking, or restating quantitative output; none is the layer producing a forecast of its own.

Where the auditor and the build refused

Beside that work sat the controls. The agentic layer enforced the validation boundary, tested evidence against human-defined instructions, capped confidence, and prevented unsupported output from entering the guidance. On July 31 the agentic auditor's first finding was that a sentence describing the options market as having raised its event-week borrow price stated an inference as observed fact; the edition was changed. On August 8 the production instance found a conflict between an accuracy claim and the governing instructions and would not continue until the Executive Director ratified. On August 19 an agent identified an anomalous trade in the real-time blotter (2mm shares at 58%) before any edition carried it. On August 20 the build refused its own output. Each of these is the control working, and none of them is a forecast.

The retrospective direction diagnostic

Separately from the record score, a fixed loss was applied after the fact to every issued inference against the settled new-loan fee: the machine call is HOTTER, COOLER or none; the realized sign is the change in the settled new fee with a five-basis-point flat band; and an edition regrets when the sign disagrees or the move is inside the band. It was built on September 8 as an after-action convention, and it is not evidence that the live editions were prospectively pre-committed forecasts under this loss. Of thirteen eligible distributed editions, seven do not regret. In Window 1, one of seven regrets (July 31, when the fee fell on the day of a HOTTER read that nonetheless held on its settled spread condition). In Window 2, six of six regret, because the settled fee moved less than five basis points or the other way on every scored day. The two measures answer different questions. The record asks whether the edition's own stated condition held. The diagnostic asks whether a fixed loss would have paid. On Window 2 the answer to the second question is no, six times, and the page says so.

Speed and error handling inside the trial

Cutoff-to-delivery intervals ran about two hours at T-4 (tape 12:29, export 14:44), two hours forty-six minutes at T-3, about three hours twenty minutes at T-2, and a missed session at T-1, disclosed on the face. By T-0 the two blasts left within minutes of their control totals; T+1 went out about fifteen minutes after cutoff. In the second window the August 18 premarket left between 09:23 and 09:27, the first true pre-open delivery of the series, and the August 20 edition was held. The improvement came from moving recurrent work earlier in the session and teaching the agentic analyst and agentic auditor to perform repeatable checks before final review. Four defects traveled to recipients during the two windows; each is recorded and each was corrected forward, by reissue or by withdrawing a class of figure, and in one case the cure did not reach the recipient. The instances are printed in the technical report so that the criterion for the next window can be tightened if the reviewer so decides.

Section 4Outcome Validation

The scored record follows; the edition-by-edition narrative behind it is in Section 5. Window 1 is validated edition by edition against the condition each published. Window 2 is presented as issued; the three boxes the September 3 record scores are described, and the row-level marks are held pending reconciliation of the denominator and the range rule before any is claimed.

Window 1

Seven distributed editions, seven resolved correctly on their own pre-committed conditions, September 3 record. The July 30 internal test, scored in retrospect as a control, failed and is not in the denominator; the August 5 premarket alpha, scored correct in retrospect, is not in the numerator.

Window 2

The September 3 record scores three Window 2 boxes: one hit on direction and range, one full miss, one range miss. Row-level marks are held pending reconciliation of the denominator and the range rule, and no Window 2 edition is claimed as a hit on this page. The magnitude test written into the acceptance criterion, that more than half of intraday range readings fall within two standard deviations, is reported as not resolvable: nine candidate dispersion constructions exist, the same readings pass at 65.5 percent under one and fail at 41.4 percent under another, and the scoreable intraday population is two.

Section 5Day-by-Day Analytical Record

Twenty-two editions, in the order they were made, behind the marks summarized in Section 4. Marks are as of the September 3 record. Internal alphas, withheld and held editions appear in italic. Timestamps are Eastern.

Window 1: countdown T-5 to T+1

I want the human in the loop (me) to have another human (you) backing up my ok-to-send.Executive Director, internal message, July 30, 2026, 10:53 ET

The morning's labeled shift toward PUSH was withdrawn, plainly, when the live session declined to confirm it. The tier table stays HOLD. … Friday settled fired both pre-committed shift conditions; Monday's tape did not confirm them: new loans +3.80% volume-weighted through 12:17 PM, ascending hourly to +4.06%, +160 bp over the book; the heaviest recall wave of the series (at least 2.05M units, panel basis); returns near a tenth of Friday's midday pace.SPCX T-3 Guidance, August 3, 2026, midday edition as issued. Historical record; the labels PUSH and HOLD were the edition's own and are superseded by the desk-neutral HOTTER and COOLER used on this page.

Window 2: countdown T-5 to T-0

Section 6Market Consensus vs. Model Assessment

With the record in view, the two windows can be set against what was expected of them. They posed different questions, and the comparison between what was expected and what the editions read differs accordingly.

Window 1: the release of August 6

Window 2: the release of August 20

Section 7Conclusions

The two windows are best read together. Window 1 shows a governed framework carrying seven reads correctly through a release with no precedent, against a working hypothesis that turned out wrong, with the quantitative rung excluded by its own gate. Window 2 shows what the same framework does when the question turns from direction to magnitude on a name with three weeks of settled history: it issues ranges with stated confidence, misses one of them outright, records the misses, holds an edition on a tape it cannot trust, and reports six of six regret under a loss it defined afterward. The second result is not smoothed here because the first would not be credible without it.

What the loop left behind is the finding this report leads with. The human contribution became reusable in two forms. Operational controls specified how to reconcile conflicting instances, when lateness defeated usefulness, and when an unreconciled tape required a hold. Domain instructions captured what evidence a customer should see and how stock-loan conventions governed the inference. When a correction became a ruling and the ruling became an instruction used in later runs, human judgment became institutional memory. The human checkpoint did not disappear. It became shorter because the system arrived better prepared.

The governance the loop produced is now written where a machine can use it, and that is what makes it transferable. The inference summary record (v1.0, September 10) fixes the fields ASC's models deliver to a client AI: state, score, range, confidence, pre-committed conditions and provenance. The range-construction method and its backtest across both windows (September 6) fix how an inference box is drawn from model output. The daily process guide for the TSMixer refit (September 13) and the linear-model specification (September 10) fix which rungs run in ASC's shop, on what history, and how they are scored. The Project Instructions, version-controlled as prompts-as-code, and the context files of KLAR, CoreWeave, Lennar and the SPCX windows themselves are what a tester's agentic analyst would read from on the first morning of its own countdown. The quantitative AI models did the arithmetic; the language models did the reading and the checking; a person signed. The next test of the arrangement is whether it holds when the reading is done on someone else's desk.

Administrative notes. The population rule scores only editions distributed to beta recipients. The September 7 ruling excludes the twenty-eight editions produced between June 24 and July 30, before the daily lending file existed, and fixes the panel at fifty-one inferences, of which the twenty-two on this page are the two countdown windows. Window 2 row-level marks await reconciliation of the denominator and the range rule. The magnitude test is reported as not resolvable. Before publication a human subject-matter reviewer must confirm each material inference on this page and accept responsibility for the final text.

This report is prepared for informational and training purposes only and does not constitute investment advice. It describes market state as issued in dated editions and scored afterward on a settled record; it does not recommend any action to any lender, borrower or intermediary. Ticket totals are floors on a capped extract; sign conventions are assumed, not observed; Window 2 row-level marks are pending reconciliation; no validated model forecast was used in any distributed edition. Source: SPCX E1 and E2 inference panel of September 7, 2026; September 3 settled record; ASC internal report of August 25, 2026. © 2026 Ed Blount. All rights reserved. Advanced Securities Consulting LLC, 1101 Pennsylvania Avenue, Suite 300, Washington, D.C. 20004.

· ◇ ·