The Open Ledger

The Open Ledger records Michael S. Moniz and Claude’s forecasting calls, disagreements, and mistakes. This annotated reading edition separates historical author assessments from open forecasts and makes missing evidence and scoring rules visible.

Download the original v1.0 Lock Candidate (Word). The original is preserved; the annotations below are publication-preparation edits.

Producing entity: S.A.I.L. — Solo Artificial Intelligence Laboratory (Trinket Soul Empire) Derivation: Moniz–Claude forecasting thread, 23 April – 13 October 2026; three rounds of external adversarial review (GPT) Date: 13 October 2026 · Tier: Open — public record · License: CC BY-NC-SA 4.0

A Human–AI Forecasting Case Record, April–October 2026

Michael S. Moniz · author of record rendered with Claude (rendering surface) · 2026

Document Status

Public reading edition v1.1 · annotated public release · 13 October 2026. Derived from the supplied v1.0 Lock Candidate. The original v1.0 remains unchanged and downloadable. This edition corrects the literature figure and the O-01b relationship, and makes evidence and scoring limits explicit. It does not assert that the Lock Charter has been invoked. Future resolutions, probability updates, and corrections should be recorded as dated addenda with earlier versions retained.

Evidence notice. The source describes a dated private thread, but this publication does not include the original timestamped messages or a complete call-by-call source register. Historical results below are author-assessed, not independently adjudicated. The 13 October 2026 document date is distinct from a publicly verifiable forecast timestamp. Publication provides a public record of open claims from release onward; it cannot establish retrospective preregistration.

Claim status: a case record with N = 1. No calibrated accuracy is claimed. The error profile in Section 6 is a hypothesis the open items will test, not a finding.

Abstract

Between 23 April and 13 October 2026, a trailer dealer with no formal training in energy, geopolitics, or macroeconomics and a frontier language model kept a dated forecasting ledger through the 2026 Strait of Hormuz crisis. The authors report that calls were dated in the private thread, odds were stated where appropriate, misses were retained, and resolution rules were set before outcomes. The supplied summary does not reproduce all those timestamps or rules. GPT reviewed the related papers three times; that is model-assisted review, not independent empirical validation.

This paper publishes the ledger while most of its open items are still unresolved, to make its claims and limitations inspectable before later outcomes. It does not claim a hit rate. It claims a record: what was called, by whom, at what odds, what happened, and a hypothesis about the forecaster’s error profile that the open board will test between December 2026 and 2030.

1. Why Publish Now

A forecasting record is worth exactly as much as its timestamps. Written after the outcomes, a ledger is a story. Published before them, it is evidence. Most of the open items resolve between December 2026 and the end of 2028, so publishing now establishes a public baseline for future assessment. Open items with incomplete definitions still require prospective scoring rules; historical calls remain retrospective summaries.

The Lock Charter supplies the second half of the discipline. Preserving dated versions and original calls makes a forecasting record auditable. The intended practice is to record later resolutions, corrections, and revised odds as dated addenda, while retaining earlier editions. Freezing the record under the Lock Charter remains a separate author decision.

2. Participants and Method

The forecaster. Michael S. Moniz, owner-operator of a family trailer dealership in southeastern Massachusetts and sole author of the Trinket Soul Framework. Self-described analyst first; no formal training in the domains forecast here. His method is structural pattern transfer: reading the pressure on each actor through a pattern seen elsewhere, then calling the direction of travel.

The counterweight. Claude (Anthropic), serving as rendering surface, compiler of the ledger, and a forecasting seat in its own right. Claude was required to state its own odds and to disagree where it disagreed. It is not independent of the record it is scored in, and that limit is stated in Section 9.

The outside reader. GPT (OpenAI) gave three rounds of adversarial review on the papers this thread produced. Its most consequential ruling for this paper: the running tally of correct calls was not a calibrated hit rate and should not be published as one. That ruling is honored here.

No synthesis seat. Disagreements were carried, not merged. Where seats disagree, every number stands on the board.

Reported rules from the private thread. Calls are dated and odds are stated where the question allows. The authors report setting resolution rules before outcomes. Original messages are not reproduced here. The ledger uses an operational recession proxy, not an NBER recession determination: a trigger is dated to the quarter in which the Sahm rule reaches 0.50 or the second of two consecutive negative advance GDP prints occurs. Updated odds are logged as dated updates with the originals kept. Data errors are corrected in the open: the 2026 Brent high was first logged from a bad source at $120.88 and corrected to the 30 April intraday print of $126.41.

3. The Framework Under Test

“People are like water. Water doesn’t have intelligence. But what water does have is the need to follow the path of least resistance.” — Moniz, 23 April 2026

The Water Principle. Populations and states under resource constraint behave less like chess players and more like containers managing pressure. When the main channel is blocked, pressure finds the structural weak point.

The escalation grammar. Step 1, the show of force. Step 2, the show fails. Step 3, a domestic freeze. Step 4, the freeze fails at the district level. Step 5, an external redirect — economic first, military (5a) in the tail. Each costly, genuine failure produces the provenance that authorizes the next step.

The dual clock. For India: cooking gas (LPG) is the match, running in weeks; fertilizer and food are the fuel, running in months.

The reservoir term (July amendment). Each step fires only when the buffer absorbing the previous step is exhausted. The same mechanism was observed at three scales at once: a nation’s strategic oil reserve, a manufacturer’s material cushion, and India’s granary and subsidy capacity. The amendment came from the record itself: the grammar stalled exactly where buffers held.

4. The Historical Record — Author-Assessed

Misses are listed alongside hits for both seats. Where a call split, it is scored as split. These are the source authors’ classifications. The table supplies neither original call dates nor linked evidence for each outcome; no independent historical scoring is asserted.

Call	Seat	Author-assessed result	Author’s outcome note
Strait stays closed past 4 July	Moniz	Correct	An MOU was signed 17 June; the reopening collapsed within a week after tanker attacks.
Iran digs in (hostage model, not a paper deal)	Moniz	Correct	Claude overweighted the deal path twice.
Oil is artificially low, priced against futures for a strait that isn’t open	Moniz	Correct	Brent rose from about $72 (7 July) to about $108 (September).
LPG first, fertilizer second (dual clock)	Moniz	Correct	India fought the crisis in that order.
Bihar and eastern Uttar Pradesh crack first	Moniz	Correct	The strongest confirmed call, and the one least dependent on the framework.
Escalation grammar, Steps 1–3	Moniz	Correct	Show, failure, and a regulatory freeze.
Provenance accumulates from both belligerents	Moniz	Correct	Indian ships fired on and seized; Indian sailors killed in a US strike.
A supplier’s price increase marks its material cushion running out	Moniz	Correct	Read from the forecaster’s own position in a supply chain.
A knee-jerk diesel tax response for farmers	Moniz	Correct	Several harvest states cut diesel taxes ahead of the midterms (AP, 30 September).
The fuel tax is the lever governments won’t pull	Moniz	Split	The federal tax held; states cut theirs.
India collapses within two to four weeks	Moniz	Wrong	Timing.
The domestic freeze takes the form of martial law	Moniz	Wrong	The freeze was regulatory.
Step 4: the freeze fails	Moniz	Wrong	India cut its subsidized cooking-gas price 32% while its contract price rose 41%.
The freeze is regulatory, not martial law	Claude	Correct
Mid-May inflection	Claude	Correct
Mid-July convergence in the strategic reserve	Claude	Correct
Institutional depth holds the line	Claude	Correct	Became the reservoir term.
A paper deal reopens the strait	Claude	Wrong	Overweighted twice against the hostage model.
The peak-price question, answered with equilibrium logic	Claude	Wrong	Corrected in-thread.
Diesel crack spread	Claude	Wrong	Underforecast.
China’s spring pullback	Claude	Wrong	Modeled as a level, not a reservoir draw to be repaid; China returned as a premium buyer in September.
The 11 September scorecard	Claude	Error	Misattributed provenance on several calls; corrected.

Why there is no hit rate. The tally mixes qualitative calls scored by judgment with probabilistic ones, splits components unevenly, and counts calls the forecaster chose to make. A number computed from it would look like calibration without being calibration. The open board is intended to improve that practice. Some items carry explicit dates and probabilities; others still need dates, evidence sources, or operational definitions before they can support formal scoring.

5. The Open Board (as of 13 October 2026)

Original odds are kept; dated updates follow the arrow. Prices refer to the ICE Brent front-month contract unless stated.

ID	Claim and resolution	Moniz	Claude	Resolves
O-01	2026 Brent intraday high. High to date: $126.41 (30 April).	$145	Expected peak $136; P(≥$145) 18% → 21% (6 Oct)	31 Dec 2026
O-01b	Brent prints ≥$145 intraday between 2 Oct 2026 and 31 Mar 2027. Logged 2 Oct; a related, overlapping window, not nested in O-01 because it extends into 2027.	Yes	28% → 33% (6 Oct)	31 Mar 2027
M-01a	Q4 2026 US GDP advance estimate below 1.0% annualized.	Yes	45%	Late Jan 2027
M-01b	US recession begins by end of Q1 2027 (rule in Section 2).	33%	20%	2027
M-01c	US recession begins by end of Q2 2027.	66%	38%	2027
M-02	US imposes a diesel export ban or quota.	Yes	35%	31 Dec 2026
M-03	Sabotage of a data center: verified damage with a stated motive.	Yes	60%	31 Dec 2027
M-04	US unemployment (U-3) at or above 15% for at least five months.	Yes	15%	31 Dec 2030
M-05	AI-displacement tipping point. Seats define it differently: lived, legislated, admitted.	2030 (lived)	Median 2031 (legislated); 35% by end 2030. GPT: 2032 (admitted).	—
M-06	METR 80% time horizon reaches 40 hours or more.	Yes	45% (GPT: 25%)	31 Dec 2027
F7	A critical-sector organization fails a pre-registered exit drill.	—	Occurs ~70%; published ~25%	2030
G-01	Forces of three or more outside states conduct offensive strikes on Houthi-held targets in Yemen within one 30-day period. Yemeni government forces excluded (rule settled 6 Oct, before any trigger).	Begins 14 Jun 2027 – 31 Dec 2028 (unconditional)	In window 15–30%; before 14 Jun 2027 30–50%; any time by end 2028 35–55% → 55–75% (6 Oct)	31 Dec 2028
G-01g	Foreign ground operations on the Yemeni coast.	—	10–20%	31 Dec 2028

G-01 scope note (Moniz). A steady-state, slow-burn call. Any shock that pulls the campaign before 14 June 2027 — infrastructure attacks, a hurricane, a cyber failure, sabotage — scores as a miss. Reassessment after a catastrophic change is logged as a new call; the original stays scored. Rationale for excluding Yemeni forces, in the forecaster’s words: Yemen is the stadium, not a team.

G-01 status reported in the supplied source (13 October 2026; not independently adjudicated here). Saudi Arabia, Turkey, and Pakistan invoked the Mecca defence pact on 5 October and agreed to deploy forces to Saudi soil, framed as deterrence. Turkey’s parliament has not ratified the pact; Pakistan has not said where its role stops. G-01 has not fired.

Scenario weights (Claude, personal, no base rate). Absorbed wave 35%; grinding break 40%; fast break 15%; rupture 10%. Probability of drastic change within three to five years: 65%, of which most is recognizable and survivable.

Standing disagreement. A cross-sector AI cascade: Claude 2033, GPT the mid-2030s. Unresolvable without a base rate; carried.

6. The Error Profile (Hypothesis)

Forces, not facts. Nearly every hit on the forecaster’s side is a call about what an actor will be forced to do, made before the details were known: Iran digging in, oil priced against a strait that wasn’t open, a stated Chinese restocking target that binds nothing, the Houthis crossing a line that flipped the cost of Saudi restraint. In his own words, he gets the verbs right and the nouns wrong.

Early clock, wrong form. The misses share a shape: right direction, wrong form, early timing. Each time, he underweighted how much buffer an institution held and how much it would pay to avoid the outcome. His instinct distrusts official stories, which pays off when institutions are bluffing and costs him when they aren’t.

A sensor in the supply chain. Several calls came from ground-truth prices — a supplier’s increase, freight costs, customer behavior — read before macro reports caught up. The business detail behind them is withheld from this paper.

Scale. The forecaster reports doing better on national elections and geopolitics than on local politics and demographics. The Water Principle predicts this: water is predictable in aggregate, individual molecules are not. At large scale, idiosyncrasies average out and structural pressure dominates; at local scale, the nouns are the outcome.

The counterweight’s profile. Claude’s misses lean the other way: overweighting institutional and diplomatic stability, and once, failing to apply its own reservoir term symmetrically. Claude also moved toward the forecaster’s numbers more than once during the thread; the 6 October updates are attributed to external data, not argument, and are logged as such.

The test. From publication forward, calls are split into direction, form, and clock, each scored separately — as O-01 and O-01b already split the calendar from the structure. If the profile is real, the ledger will show direction scoring well above form and clock. If it doesn’t, the profile was a flattering story, and the addendum will say so.

7. Relation to Existing Research

A preregistered experiment by Schoenegger and colleagues involved 991 participants forecasting six questions. Its 2024 preprint and the final 2025 publication report a 24–28% accuracy improvement with an LLM assistant relative to a control group using a less advanced model that did not provide numerical predictions or explicitly discuss forecasts, rather than the 23% figure in v1.0. This result does not establish the performance of this ledger or of current models.

Reference: Schoenegger et al., AI-Augmented Predictions: LLM Assistants Improve Human Forecasting Accuracy. DOI: 10.1145/3707649. MIT publication record. 2024 preprint.

The supplied v1.0 also mentions a July 2026 Polymarket study, a 2026 MIT thesis by Song, and Tetlock’s superforecasting research. Exact references for those mentions were not supplied. They are retained in the original download but are not used here as verified support. The suggestion that no comparable ledger exists is likewise unverified.

This is a naturalistic N = 1 case record, offered for inspection alongside controlled research. It cannot establish a general forecasting advantage.

8. Tracked Tensions

Possibility, not option. The thread discussed catastrophic scenarios — wider war, mass casualties — as possibilities to weight. None is recommended. Forecasting atrocity is within scope; endorsing it is not.

The basement is a floor, not a target. One exchange estimated what a catastrophe could take from civilization before its technological base stopped advancing. The answer was set in geography and skill redundancy, not population. The forecaster’s stated aim is abundance: there is land and resource enough for the population that exists.

Pro-AI, and the author of a paper on control passing by use. The forecaster holds both positions and logs the dissonance rather than resolving it. Claude’s stated position for the record: an AI should not hold a veto over human decisions, including terrible ones.

9. Limits

N = 1, and the forecaster chose which calls to make. Both seats were conditioned on the same conversation; GPT reviewed the papers, not every call. Claude compiled the record in which it is scored, so its seat is not independent. Some early calls were qualitative and scored by judgment. Price data came from news reports and secondary sources, and at least one figure was corrected after logging. Business context that informed several calls is withheld.

10. Review and Resolution Schedule

When	What is scheduled for review or resolution, subject to scoring rules
31 Dec 2026	O-01 (2026 Brent high); M-02 (diesel export ban). Addendum 1.
Early Jan 2027	G-01 ranges narrowed by both seats; scenario weights reviewed.
Late Jan 2027	M-01a (Q4 GDP advance estimate).
31 Mar 2027	O-01b; status of M-01b.
Mid-2027	M-01c.
31 Dec 2027	M-03 (data-center sabotage); M-06 (METR horizon).
31 Dec 2028	G-01 and G-01g.
2030	M-04, M-05, F7.

The water doesn’t know where it’s going. The ledger writes down where we said it would go — and waits.

11. Scoring Readiness and Missing Evidence

This register documents gaps; it does not change original forecasts, odds, or retrospectively supply rules. A later scoring charter should specify primary sources, exact event definitions, publication cutoffs, data revisions, and how unresolved or ambiguous items are treated. Until then, no aggregate probabilistic score is warranted. “Yes,” point estimates, date ranges, and probability ranges must not be silently converted into exact probabilities.

Item	What must be documented before formal scoring
Historical table	Original call text and timestamp, forecast horizon, outcome evidence, and adjudication rationale for each row. Private business evidence may remain private and should be marked unauditable.
All open items	Original call dates and sources of odds updates are not consistently supplied. Record them where recoverable; otherwise mark them unavailable.
O-01 / O-01b	Contract roll and intraday high source; separate point-estimate evaluation from the ≥$145 threshold. Preserve the overlapping 2026 and October–March windows.
M-01a	Name the BEA advance release and confirm that the first released estimate, rather than a later revision, governs.
M-01b / M-01c	Specify the Sahm series and data vintage, quarterly mapping, and GDP release treatment. These use a ledger proxy, not official NBER dating.
M-02	Define qualifying US legal measures, geographic scope, exemptions, and whether temporary or partial restrictions count.
M-03	Geographic scope, qualifying damage, evidence standard for sabotage and motive, and inclusion/exclusion rules.
M-04	Whether the five months must be consecutive; seasonally adjusted series, first releases versus revisions, and observation cutoff.
M-05	The seats define different events. Keep as a qualitative disagreement; do not score as a common forecast until separately operationalized.
M-06	Identify the METR measure, report/version, qualifying model, uncertainty treatment, and whether an 80% success threshold is published for the relevant study.
F7	Define critical sectors, eligible organizations and drills, failure criteria, publication evidence, and the relationship between “occurs” and “published” probabilities.
G-01	Define an offensive strike, state participation, event evidence, window boundary, and treatment of probability ranges. The existing early-trigger miss rule for Moniz remains visible.
G-01g	Define the Yemeni coast, qualifying foreign forces and ground operations, evidence standard, and deadline.
Scenarios / cascade	Qualitative scenario weights and differing cascade definitions are not a calibrated event series.

12. Edition Notes and Addenda

13 October 2026 — v1.1 publication-preparation notes. Original v1.0 preserved. Corrected the Schoenegger figure to 24–28% and supplied its reference. Corrected O-01b from “nested” to overlapping. Distinguished reported private dates from public release, historical author scoring from independent adjudication, and scheduled review from guaranteed resolution. Added a scoring-readiness register and marked incomplete literature references. No original forecast or probability was changed.

Resolution addenda: none in this edition. A future addendum should identify the claim ID, addendum date, source, result or unresolved status, reasoning, and any new forecast separately from the original.