Filing operating-cash-flow transition × prior-session short-flow crossing

Date: 2026-08-15
Status: preregistration; no decision-or-later price, return, path, or historical-effect field has been queried for this candidate
Candidate: filing-ocf-transition-prior-session-short-flow-crossing-v1
Phase: bounded discovery only
Capital authority: none

Claim and fixed identity

The claim is long: a filing-time operating-cash-flow transition from nonpositive to positive, while revenue rises year over year, should lead to a more favorable five-session reassessment when the immediately preceding completed FINRA session has crossed into a frozen high short-sale-flow state relative to its own prior 20-observation baseline.

FINRA daily short-sale volume is publicly disseminated off-exchange trade flow. It is not short interest, net short positioning, borrow demand, proof of trapped participants, or consolidated all-venue volume. The candidate tests a repeated information/flow sequence without asserting which of informed selling, intermediation, liquidity provision, or offsetting activity generated the flow.

• Positive seed: current quarterly operating cash flow is positive, the same fiscal period one year earlier was nonpositive, and current revenue exceeds prior-year revenue.
• Anti-seed: current quarterly operating cash flow is nonpositive, the same fiscal period one year earlier was positive, and current revenue exceeds prior-year revenue.
• Crossing state: the last completed research session strictly before filing_date has FINRA short_ratio >= 0.50, while the median of the preceding 20 valid FINRA observations is below 0.50.
• Noncrossing state: the same clocks and 20-observation coverage apply, but the exact crossing rule is false.
• Decision clock: official open of the first research session strictly after filing_date.
• Horizon: five sessions including the decision session; entry is decision-session open and exit is the close at decision_session_seq + 4.
• Direction: long.

Only the first filing for a (symbol, quarterly, fiscal_year, fiscal_period) identity is eligible. Quarterly Q1-Q3 observations must have a prior-year filing for the identical fiscal period, and the two period-end dates must be 300-430 days apart. Amendments and later restatements cannot replace the first-information row.

Frozen epochs and trial accounting

• Discovery filing dates: 2023-07-01 through 2025-12-31.
• Sealed historical screen: 2026-01-01 through 2026-07-31. It remains unopened unless every discovery gate passes.
• Forward evidence: events after 2026-07-31. Unavailable today.
• The complete four-cell discovery evaluation is one semantic outcome trial. Renaming, reordering, changing a threshold, liquidity tier, clock, benchmark, cost, or horizon is not a new trial.
• No discovery result may enter a proposing or reviewing model as rows, symbols, dates, paths, charts, or free-form evaluator prose. Only bounded aggregate failure codes and an input-safe lesson may enter search memory.

Outcome-blind eligibility

At the decision open, every event must have:

• at least 20 prior research sessions;
• no prior-20 source gap and no later-split adjustment hazard;
• prior close of at least $2;
• trailing prior-20 median dollar volume of at least $5 million;
• prior-session realized_vol_20d and momentum_20d available;
• current security metadata classified as common stock or ADR common, treating missing type as common stock, and no ETP flag;
• exactly 20 valid FINRA observations before the participant-flow session;
• source-component reconciliation within 0.001 for the published short_ratio.

The current security map is a disclosed survivorship limitation. It may be used only as an eligibility exclusion, never as a predictor.

The selected $2/$5M tier is frozen from the input-only tier census. It retained 35 positive crossing and 27 anti-crossing events. The looser $1/$1M tier added only one positive-crossing event, no anti-crossing event, one positive-crossing date, and no anti-crossing date, so its extra execution exposure was rejected without inspecting outcomes.

Frozen input gates

The eligible manifest must meet all gates before any outcome task exists:

Every included event must pass deterministic filing, fiscal-pair, revenue, OCF-transition, decision-clock, participant-clock, baseline-order, ratio-reconciliation, security-type, and pre-entry liquidity checks. A deterministic 50-case audit manifest is frozen from the validated roster. Any input gate failure yields UNDERPOWERED_INPUT without reading outcomes.

The ordered event roster, full prior-20 FINRA observations, input-query hash, prevalence-receipt hash, source snapshot, eligibility exclusions, configuration hash, semantic fingerprint, and manifest hash are persisted before outcome unlock.

Primary estimand and outcomes

For each event, compute:

event open-to-session+4-close return
- SPY return over the identical session sequence

The primary interaction is:

[positive crossing - anti crossing]
- [positive noncrossing - anti noncrossing]

where every term is a mean SPY-adjusted event return. The identical interaction is also reported on raw returns as a descriptive audit, but the SPY-adjusted version alone is primary.

The prior-window placebo uses the same frozen event symbols and SPY from the open at decision_seq - 5 through the close at decision_seq - 1.

Primary transaction cost is 50 bps round trip and stress cost is 100 bps. Costs do not alter the four-cell interaction because every cell has the same long convention, but they remain mandatory for the absolute positive-crossing event book.

Frozen statistical and falsification gates

All gates must pass:

1. primary SPY-adjusted interaction is at least +1.00 percentage point;
2. one-sided seed-label permutation within (calendar year, crossing state) has p <= 0.05;
3. one-sided crossing-label permutation within (calendar year, seed class) has p <= 0.05;
4. one-sided 95% lower bounds from decision-date-cluster and issuer-cluster bootstraps are above zero;
5. positive-crossing SPY-adjusted mean and crossing positive-minus-anti contrast are both positive;
6. positive-crossing raw event-book mean remains positive after 100 bps;
7. interaction remains positive after excluding the most favorable decision date and, separately, the most frequent issuer;
8. prior-window interaction has absolute magnitude below 0.50 percentage point and a two-sided permutation p > 0.10;
9. no decision date contributes more than 20% of total absolute interaction contribution;
10. exact target-session exit coverage is at least 98% in every cell for both the event and SPY, with no event silently replaced after outcomes are visible.

Permutations and both cluster bootstraps use 10,000 draws and seed 20260815 with fixed namespaces. Inference is clustered by filing decision date and event issuer. Event-effect rows are hashed and discarded from the model-facing receipt.

Mechanical disposition

• PASS_DISCOVERY: every gate passes. This authorizes only the unchanged sealed 2026 screen.
• REJECTED: powered discovery reaches outcomes but any economic, inference, placebo, robustness, cost, or concentration gate fails. The semantic lineage is retired and a bounded lesson is stored.
• UNDERPOWERED_INPUT: a frozen input gate fails before outcome access.
• INCONCLUSIVE_DATA: the frozen roster is powered but outcome coverage fails. The roster cannot be repaired or replaced after inspection.

Preregistration and manifest freezing do not themselves authorize outcome access. A separate exact evaluator adapter and durable semantic-trial unlock must be reviewed, committed, and deployed before the first price query.

No historical result grants paper orders, live orders, sizing, scheduling, or capital allocation.