The research log

A note before the entries. Everything here was checked against the record — the gauge tables, the frozen specs, the transcripts — before it went on this page. Dates are exact. What you will not find: position sizes, dollar amounts, or anything that describes the live book. Direction and reasoning only. And where an entry doesn't prove something, it says so.

August 5, 2026 — The refunding fork: pre-registered, gauged, resolved

The question all July: would the Treasury's August-5 refunding statement break the long end loudly, or pass quietly? Instead of holding an opinion, we froze a fork on July 4 — three branches with probabilities written down before the event: silent 45%, violent 35%, reactive-hybrid 20% — and trigger rules for what would move us between branches. About 65% of the probability mass said nothing breaks before the print. Then we instrumented it.

A standing gauge table ran daily from July 8 — term-premium anchor, stablecoin float, auction cover, bank-credit reads, a variance-ratio conditioner with its threshold written into the row. Over the first two weeks of July, the ten-year term-premium anchor climbed about nine basis points — from 0.594 on July 6 to 0.685 on July 15 — while the risk-neutral leg was flat, moving about a basis point the other way. The entire move was term premium, not rate expectations. That decomposition, not the level, was the finding.

Before the print, the adjudication test was frozen to two limbs: coupon auction sizes raised, or the "maintain for at least the next several quarters" guidance language dropped. Bill-share shifts would not count, no matter how the tape felt about them. On August 5 the statement held sizes and kept the language — neither limb fired. Silent branch, adjudicated on the frozen test, not on the mood.

The honest wrinkle, recorded the same session: the metals tape bid hard into the quiet print (gold up ~3.9%, silver up ~4.7%). The system's read — the statement adjudicates supply guidance; the metals bid reads policy trajectory; the two can disagree without contradiction, and the unresolved half points at the August 11–13 auctions. And where the quiet print authorized an action under a standing rule, the system surfaced the authorization explicitly rather than assuming it exercised — the decision stayed human.

On the record: probabilities and triggers pre-registered July 4; the two-limb test frozen before the statement; gauge rows dated and receipted. This doesn't prove the fork was right — it proves the call was made before the outcome and scored against a test that couldn't be bent afterward.

July 20–21, 2026 — The instrument that refused its own thesis

Every correlation instrument we'd built detects decay — a crowded trade coming apart. None detects a trade being born. So we designed one: a two-leg gate where cluster coherence and actual capital adoption both have to clear, specifically so a synchronized selloff can't masquerade as formation. The spec was frozen to the repository before any data contact — thresholds, roster, benchmarks, and the pass/fail bar, all written first.

Then the data. Across 16.5 years the frozen gate almost never opened, and the one event that fired failed the leave-one-out test. An outcome-blind amendment re-gated it in residual space and found eight events across three decades — of which one survived, and the forward claim missed its frozen bar at every filtering level. The design did do one thing exactly right: all three synchronized-selloff candidacies were rejected by the adoption leg — the precise failure the two-leg design existed to refuse.

The verdict, written blunt: not supported; the phenomenon is genuinely rare; the bars stay frozen; the instrument is retired. The pre-commitment — written before any outcome was seen — said it best: reaching the end without enough events "means the phenomenon is genuinely rare, and that is itself the finding — not a reason to loosen."

On the record: spec frozen before data contact; point-in-time data; an independent spot-check pass; kills recorded as wins. This entry exists because the instrument failed — a research process you can trust is one that shows you its dead ends at full price.

May 8 → ongoing — One framework, three months of instruments

Articulated May 7–8: a two-regime bifurcation thesis — a medium-of-exchange leg (AI capital spending defended, a managed dollar) and a store-of-value leg (metals, with the fiscal arithmetic binding) — both able to win at once inside a narrow band. Ten days later the band's binding condition was corrected and the correction was kept, dated, next to the original: productivity below two percent — strong enough to defend the narrative, not strong enough to force the policy response.

What makes this a log entry isn't the thesis. It's what grew on it. By August the framework had produced the term-premium decomposition above, the standing gauge table, two market-dynamics needles, and the refunding fork — which is this same framework's August chapter, adjudicating one of its named break conditions. Each instrument was built on the last one's findings, and the corrections stayed in the record with their dates. That's the compounding claim on the front page, shown rather than asserted.

On the record: articulation and correction dates exact, held in the system's memory with turn-level sources. The framework is still open — it's a lens under calibration, not a victory lap.

May 2026 → ongoing — Three passes at a market I'd never traded

Before May I had never traded a bond instrument. The system taught me the market first — why the long end matters, what an auction tests — and the first expression, into the May refunding, was small and profitable. The second, into the June 11 auction, lost: we exited the same session, whole, on the rule we'd set before the print — no tail, no thesis. That loss bought the most. The post-mortem became the sectoral-balance framework the entry above is built on. The third pass was pre-registered on July 4 with a frozen two-limb test, adjudicated on August 5 exactly as written, and its remainder is live into the August auctions. Neither of us makes these trades alone — and the losing one is why the process exists.

June 5, 2026 — The cold open

The oldest entry here, kept for time-depth. Every morning I open a blank conversation and type one line: "market open update, please." No context, no reminder of what I hold or what I decided yesterday. Here is part of what came back at 8:45 AM PT on Friday, June 5 — as-is:

"This is a hard risk-off open — let me give you the read, not a price list. S&P −1.4%, Nasdaq −2.4%. This is not a rotation — it's everything down together, with semis and silver leading the fall. When both clusters crack on the same tape, the driver is usually liquidity, not a thesis-specific break in either. [...] Silver moved ~2x gold, the miners ~2x silver — the bidirectional leverage firing as designed. Thesis intact, P&L worse. This is the add-on-weakness tape your framework names — but your standing posture since May 19 is no fresh metals, and the June 11 auction is the deliberate gate. Holding that is the discipline; I'm flagging the flush, not calling the add."

Notice what didn't happen. It didn't ask what I own. It didn't see silver down 7% and cheerfully suggest buying the dip. It read the print through a prior it has worn long enough to reason from — and it held my own rule against the tape's temptation, while explicitly leaving the decision with me.

On the record: response verbatim from June 5, lightly trimmed; the positions visible in it were mine that morning, shared by choice. It triggered on my ask — the flagging is the system's, the discipline is ours, the decision is mine.