Research note 014: memory strengthens with stress, sessions stay uniform
Date: 2026-10-16 · Status: one extinction week, hourly buckets, stated pairs. Second crisis datapoint for note 013's open question: Luna/UST death spiral (2022-05-08..14, BTC ~$35k→$26k), identical lag-1 method over hourly taker-buy shares — overall, per session, and split by first-hour magnitude.
Question
Note 013 found r=+0.49 in the SVB week against +0.36 calm — suggestive, not established. A second, harsher crisis decides whether the ordering (stress strengthens memory) replicates or was noise.
Dataset and method
- Seven daily spot aggTrades slices, 13,960,241 prints into 168 UTC
hours, overall taker-buy 50.2% (balanced tape through a crash).
Identical pairs to notes 011/013 (consecutive hours, same-session
only, median magnitude splits). Pins are extracted-CSV SHA-256 in
research/scripts/stress_014.py(stdlib only, one command). - Daily flow ramps mid-week (3.0M/3.3M prints May 11–12 vs ~1M on the quiet days) — the stress is in the data, not assumed.
Results
| Scope | Calm (011) | SVB stress (013) | Luna stress (014) |
|---|---|---|---|
| All consecutive hours | +0.36 ± 0.08 | +0.49 ± 0.08 | +0.62 ± 0.08 |
| Asia 00–08 | +0.34 ± 0.14 | +0.44 ± 0.14 | +0.67 ± 0.14 |
| EU 08–16 | +0.34 ± 0.14 | +0.43 ± 0.14 | +0.60 ± 0.14 |
| US 16–24 | +0.34 ± 0.14 | +0.60 ± 0.14 | +0.49 ± 0.14 |
| High-magnitude first | +0.32 ± 0.11 | +0.60 ± 0.11 | +0.75 ± 0.11 |
| Low-magnitude first | +0.03 ± 0.11 | +0.20 ± 0.11 | +0.29 ± 0.11 |
Reading
Three facts, one upgrade and one refusal. First, the ordering replicates: calm +0.36, SVB +0.49, Luna +0.62 — memory strengthens monotonically with stress across three independent weeks, and the Luna-vs-calm gap (+0.26 at ~2.3 sigma) now clears the bar the SVB gap only approached. Second, the magnitude mechanism replicates with it: +0.75/+0.29 high/low, the same conviction-hours story at higher amplitude. Third, sessions refuse to rank: Asia reads highest (+0.67) and US lowest (+0.49) — the reverse of the SVB week — but no pairwise gap clears combined error bars, so the uniform reading from note 011 stands extended, not overturned: memory is session-flat within error in calm, SVB, and Luna weeks alike. The epicenter moved (US hours for SVB, Asia hours for Terra) while the uniformity did not — consistent with one tape-wide carryover process whose level, not whose memory, follows events.
What would strengthen or overturn this: a third crisis (each additional week is another independent draw — the honest unit is still small); the perp cut in stress (does balanced flow remember any better when it matters?); maker-side persistence; and the live-captured version, where hours stop being archive buckets.
Reproduce
python3 research/scripts/stress_014.py --data-dir /tmp/stress014
Seven CSV content hashes pinned in the script, refused on any mismatch — verified by a from-scratch rerun with identical figures.