Comuvia ForeGlass™
ComuviaComuvia ForeGlass™

Forecast results & evidence

Read the first recorded results, compare the published forecasts, and inspect the evidence behind them.

FIRST RECORDED REVIEW · 15 OCTOBER 2026

What we predicted.
What the data showed.

METRIC: U.S. BOTTOM-50% SHARE OF HOUSEHOLD NET WORTH · 2026 Q2

The bottom half of U.S. households held 2.304% of household net worth in the retained Federal Reserve estimate for 2026 Q2. Both ForeGlass central forecasts were 2.452%, or 0.148 percentage points above that estimate.

In everyday terms: of every $100 of total U.S. household net wealth, the least-wealthy half of households collectively held about $2.30. This is the group’s combined share, not an amount held by each household.

How to read this metric

Who is the bottom half? Imagine 100 households ordered from lowest to highest net worth. The first 50 make up the bottom half. The grouping is by wealth, not income, age or earnings.

What is net worth? The value of assets, such as homes, bank balances and investments, minus liabilities (debts), such as mortgages and loans. A household can have negative net worth when its debts exceed its assets.

What is the calculation? Add up net worth for the least-wealthy 50% of households, divide by net worth for all households, and multiply by 100. Group net worth ÷ total household net worth × 100 = wealth share (%). It is not the average household’s wealth, an income share, or a statement that every household in the group owns the same amount.

What are we comparing? Forecasts of this same U.S. wealth share for 2026 Q2 against the Federal Reserve estimate retained on October 1 and recorded in the October 15 review. “Actual” in the tables means that reported estimate; it is not an exact census and later source revisions may differ.

What does the error mean? The absolute difference is |2.452% − 2.304%| ≈ 0.148 percentage points, calculated from unrounded values. Percentage points measure the gap between two percentages. They are different from relative percent error and from an accuracy or success rate.

What can it tell us? It describes how household wealth is distributed. On its own, it does not measure cost of living, safety, financial hardship or overall economic fragility, and a change in the share does not by itself show how the group’s wealth changed in dollars.

Read the Federal Reserve’s definition and estimation method ↗

Reported estimate2.304%

Federal Reserve · 2026 Q2

ForeGlass central forecast2.452%

Same central value in two records

Absolute forecast error0.148 pp

pp = percentage points

THE FOREGLASS RESULT

Same central forecast. Different ranges.

Compare all 19 public records ↓

The July range contained the reported estimate; the August range did not. Range containment and central-value error answer different questions. The interval score prices both at once: the range's width plus 2/(1 − coverage) times the distance outside it, at the coverage the record declared, lower is better; the July record declared no coverage, so it carries no score. A wider range is not, by itself, evidence of a better forecast.

Registered 2026-07-30

1.978% – 2.925%

Recorded range · central forecast 2.452%

No coverage declared · range not scored

Registered 2026-08-11

2.339% – 2.557%

Recorded range · central forecast 2.452%

Interval score 0.565 pp at 80% coverage
ForeGlass forecast ranges and the reported estimateHorizontal bars show the recorded ranges, dots show central forecasts, and the dashed vertical line shows the reported estimate. Exact values are in the table below. The percentage axis does not start at zero.1.90%2.12%2.34%2.56%2.78%3.00%Reported 2.304%Registered 2026-07-30Registered 2026-08-11
Bars: recorded ranges. Dots: central forecasts. Dashed line: reported estimate. The axis shows wealth shares and does not start at zero.
How to read the errors and the comparison anchor

Absolute error is the distance between the central forecast and the reported estimate, in percentage points. The ForeGlass relative error is 6.41% of the reported estimate; it is not a success rate. The July record stated no probability level for its range; the August record declared 80% coverage. Those are different declarations, so the ranges are not presented as equivalent confidence intervals. These coverage annotations were checked against the original records; this public projection does not expose a coverage field.

The recorded comparison anchor is 2.471%, from 2025 Q1. Its absolute error is 0.167 percentage points, compared with 0.148 for the two ForeGlass central forecasts. The anchor is preserved as recorded; it is not the same quarter one year before the 2026 Q2 target. This single comparison is not evidence of general predictive skill.

THE PUBLISHED COMPARISON

All 19 public results for this question

Download the public ledger ↗

Smallest absolute error first. Ordering uses unrounded errors for this one question. Equal errors are ordered by registration date, then record ID; displayed errors are rounded. This is a comparison with one reported estimate, not an overall ranking of forecasting quality.

“With engine context” means the model received simulation information. The first-round “cold” prompt design also reused the simulation’s sealed claim, which could contain forecast numbers. The original verbatim prompts were not retained. “Cold” therefore does not establish independence from the engine for this cohort. Ensembles combine model forecasts. These input conditions matter when interpreting agreement.

All 19 published central forecasts were above the reported estimate. This describes these records on this one question, not performance on other outcomes.

Shared reported estimate: 2.304344% of household net worth. All errors are absolute. The interval score is the range's width plus 2/(1 − coverage) times the distance outside it, at the declared coverage; lower is better; it describes this one estimate.
Forecaster / inputRegisteredCentralRecorded rangeError ↑Interval score
Ensemble · medianWith engine context · retired
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-ens-median-grd
2026-08-172.450%2.340% – 2.560%0.146 pp0.577 pp
Ensemble · medianCold arm · first round
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-ens-median-raw
2026-08-172.450%2.340% – 2.560%0.146 pp0.577 pp
Ensemble · trimmed meanCold arm · first round
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-ens-trimmed_mean-raw
2026-08-172.450%2.342% – 2.560%0.146 pp0.599 pp
DeepSeek-V4-ProCold arm · first round
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-foundry-deepseek-v4-pro-raw
2026-08-172.450%2.340% – 2.560%0.146 pp0.577 pp
gpt-4oWith engine context · retired
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-foundry-gpt-4o-grd
2026-08-172.450%2.340% – 2.560%0.146 pp0.577 pp
gpt-4oCold arm · first round
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-foundry-gpt-4o-raw
2026-08-172.450%2.350% – 2.550%0.146 pp0.657 pp
Grok-4.3With engine context · retired
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-foundry-grok-4-3-grd
2026-08-172.450%2.340% – 2.560%0.146 pp0.577 pp
Grok-4.3Cold arm · first round
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-foundry-grok-4-3-raw
2026-08-172.450%2.340% – 2.560%0.146 pp0.577 pp
Kimi-K2.5With engine context · retired
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-foundry-kimi-k2-5-grd
2026-08-172.450%2.380% – 2.520%0.146 pp0.897 pp
gpt-5-miniCold arm · first round
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-gpt-5-mini-raw
2026-08-172.450%2.340% – 2.560%0.146 pp0.577 pp
kimi-k3Cold arm · first round
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-kimi-kimi-k3-raw
2026-08-172.450%2.340% – 2.560%0.146 pp0.577 pp
Ensemble · trimmed meanWith engine context · retired
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-ens-trimmed_mean-grd
2026-08-172.451%2.340% – 2.559%0.146 pp0.573 pp
ForeGlassForeGlass Engine
Record IDpred-2026-07-30-025
2026-07-302.452%1.978% – 2.925%0.148 ppNot scored · no coverage declared
ForeGlassForeGlass Engine
Record IDpred-2026-08-11-069
2026-08-112.452%2.339% – 2.557%0.148 pp0.565 pp
DeepSeek-V4-ProWith engine context · retired
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-foundry-deepseek-v4-pro-grd
2026-08-172.452%2.339% – 2.557%0.148 pp0.565 pp
Ensemble · inverse varianceWith engine context · retired
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-ens-inverse_variance-grd
2026-08-172.453%2.408% – 2.497%0.148 pp1.130 pp
Ensemble · inverse varianceCold arm · first round
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-ens-inverse_variance-raw
2026-08-172.458%2.417% – 2.499%0.154 pp1.206 pp
gpt-5-miniWith engine context · retired
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-gpt-5-mini-grd
2026-08-172.470%2.320% – 2.580%0.166 pp0.417 pp
Kimi-K2.5Cold arm · first round
Record IDpred-2026-08-17-tourney-us-dfa-bottom50-networth-share-interval-foundry-kimi-k2-5-raw
2026-08-172.480%2.400% – 2.560%0.176 pp1.117 pp
Which records are visible? The October 15 review recorded 39 outcomes for this question. This edition publishes 19 result records from the earlier registration cohort; the other 20 are represented as hash-only commitments, without their forecast values or results. The cohorts differ. This selection must not be presented as the system’s overall accuracy.

SOURCE AND TIMING

From reference period to review

  1. 2026 Q2The period measured. It had already ended when these forecasts were registered.
  2. 30 July – 17 AugustRegistration dates of the records shown here.
  3. 1 OctoberWe retrieved the Federal Reserve source file. The publication date of these exact source bytes is not established in the retained evidence.
  4. 15 OctoberOutcomes were recorded against that retained copy. This is the review date, not the date the source was published.

Federal Reserve Distributional Financial Accounts ↗

The figures are reported estimates, including modeled allocations between wealth groups. The current source may contain subsequent revisions.

SCOPE OF THIS REVIEW

A partial, explicitly bounded evaluation

Of 179 forecasts due for this review, 39 outcomes were recorded and 140 remained on hold. All 39 recorded outcomes concern the same wealth-share question.

  • 39 top-1% wealth forecasts: no admitted resolution policy.
  • 38 Treasury-holdings forecasts: sector-definition and source-admission issues.
  • 63 employment-revision forecasts: restricted to historical diagnostics; no admitted resolution policy.

These are separate methodological holds, not a claim that every missing result is waiting for new data.

Why this is not blind validation

The source-admission policy was adopted on October 14, after the outcome had been inspected. Registration, evidence selection and later scoring are different steps. Neither an unchanged hash nor a newly timestamped release proves that an outcome was unknowable when a forecast was made.

What comes next

Resolve each outstanding hold under an explicit evidence policy, collect more independent outcomes, and use the errors to guide model improvements. Future evaluations should specify source admission before outcome exposure. The immediate milestone is a checkable record of this result, including its limitations.

Inspect verification details →Explore consensus agreement →

sealed ledger ● 0 shown
Lanei Claimi Predictedi Forecasteri Resolvesi Againsti Resulti

How to read a row: each row represents one registered forecast. Claim describes the committed statement (or a labelled summary where the original wording is withheld); Against names its source or grading rule. Click a row, or focus it and press Enter, to inspect the available record, hash and source-series chart. Forecast contents, outcome entries and publication visibility are distinct.

One question, many forecasters

Two different comparisons: for a question with a recorded outcome, the observed value and absolute error show how each forecast compared with that estimate. The LLM median and consensus board measure agreement between forecasts. Input conditions differ: retired engine-grounded entries received the simulation forecast and are not independent alternatives to it. Other prompts or inputs may produce different results.

ForeGlass LLM Ensemble LLM median

Parts that must add up to the whole official published statistics

In plain terms: a forecast about a part should be consistent with the whole it belongs to. Each chain takes an officially published total, splits it into the pieces the statistics agency itself publishes, adds them back up, and shows how far off the round trip lands. A rung marked ◆ carries one of our sealed forecasts. Each chain is a group: open the one you want. The per capita switch redraws the ladders per person or per household where the chain has a stated denominator - that view is a ratio, so the shares, the running total and the residual stay on the totals.

About this tab

These are measured, published values reconciling with each other - the frame our sealed forecasts sit inside.

Where the parent total is independently compiled (the World Bank's EU aggregate), landing on it checks our member set, joins and arithmetic against an outside compiler. Where the parent is only the sum of its own published bands, exact agreement is the expected result - it verifies arithmetic, not two independent measurements.

What this tab does not show
  • Closure is not accuracy. Almost none of these rungs carry a forecast yet; coverage is stated per chain as a number.
  • Exact agreement where both sides rest on the same underlying series is expected, not corroboration.
  • A chain that cannot be built is shown as a refusal naming the blocker.
  • Per-capita values are ratios on a stated denominator - they do not add up, and where the denominator comes from a different statistical universe than the numerator the chain says so.

What this page is, and what it is not

This page brings together registered forecasts, hash commitments and admitted outcomes. Registration, source admission, scoring and public release are separate steps. The first recorded review and its limitations are explained in First results; the ledger lets you inspect the individual public records.

  1. Separate registration from independent timing evidence. Each record carries a registration date and a hash. The hash identifies committed content; it does not by itself prove when that content existed or when an outcome became knowable. Archived releases and independent timestamp evidence can support chronology, with the limitations stated under Verify.
  2. Central error and range containment. Absolute error is |actual − forecast|, in the metric's units. A range contains the result when the reported estimate lies between its recorded endpoints; the ledger calls this hit. The headline result of a range is its interval score, a proper score: the range's width plus 2 / (1 − coverage) times the distance by which the reported estimate fell outside the range, at the coverage the record declared (a record that declared none has no score). Lower is better. A narrow range that misses is penalised, a wide range that contains the estimate pays for its width, and the score is in the metric's units, so it compares records on one question and never across questions. Relative absolute error divides central error by the magnitude of the reported estimate and multiplies by 100. It is not a success rate. Where a published record has naive_value, its comparison-anchor error can also be calculated. This edition publishes no skill figure or aggregate winner. For the first review the anchor is 2025 Q1, with its original historical label preserved in the record.
  3. Source vintage and timing. A result is graded against the retained copy of the source in our evidence pack. A later source revision is not silently substituted into that result. Grading after the scheduled date is recorded separately from the source's publication date. release_observed contains a documented publication date or the explicit value unknown for a late row; if neither is present, the page reports the missing metadata. Retrieval and grading dates must not stand in for an unknown publication date.
  4. Distinguish the committed rule from source admission. The sealed statement and rule are kept as recorded; some public fields are deliberately withheld or restated with labels. Implementation and evidence-admission decisions require separate review. For the first review, the source-admission policy was adopted after outcome exposure, so the evaluation is not described as fully blind or preregistered.
  5. Separate recording from publication. Public release involves content review and visibility rules. This edition includes both contained and missed ranges, but it does not publish every recorded result. The first review's selection is set out beside the first-results table.
  6. Compare registration dates and inputs. Forecasts answering one question can have different registration dates, context and methods. The first-results table shows those differences. Multiple model variants and ensembles are not independent economic outcomes.
  7. Held forecasts are still committed. Additional forecasts appear in the public ledger as hash-only commitments. Comparing a later disclosure with the earlier commitment can detect changed content; independent timing evidence is a separate check.

Disclosure boundary: this public edition does not include model internals, calibrations or full run lineage. You can inspect the published forecasts and reported results; you cannot reproduce the complete simulation from this page alone.

Verify it yourself

Alongside this page sits ledger.json — the machine-readable ledger behind everything shown here.

The ledger identifies the report source and its SHA-256 digest. The source digest covers the report before website navigation and related videos were added; it does not cover this composed web page.

  1. Seal algorithm:
  2. What you can check here:
  3. Independent timestamp evidence. The release workflow pairs this ledger with an OpenTimestamps receipt. A new receipt may still be pending; only an upgraded, verified Bitcoin attestation establishes the corresponding block-time bound. Verify its status with an OpenTimestamps client. A timestamp obtained for this results edition does not prove that its forecasts existed before outcome exposure. For earlier evidence, inspect the archived releases.
  4. Check a past release. Every stamped release keeps a dated copy of the ledger and its proof under ledger-archive/, listed in index.json. Each entry is addressed by its own date (ledger-archive/<date>.json), so a claim made months ago can be checked against the bytes that carried it rather than against today’s ledger. What the index is: a record of what the last site build saw in that directory — not a guarantee that every release is listed, and not a promise that any particular entry is still served. Neither the index nor this page republishes the underlying source data; those are linked from their own landing pages, byte-for-byte copies are not redistributed here, and a link that resolves is not a reconstruction.
  5. Commitments. ledger.json carries a commitments array: one {id, hash, sealed date} row per held forecast. When a held forecast is released, its hash on this page must equal the hash committed earlier — a forecast rewritten in between cannot match.
  6. Time-witness it. Snapshot ledger.json at an independent web archive. That timestamp is outside our control, which is what makes the ordering claim checkable by someone who does not trust us.
Grading status:

Honest limit: two of the fields inside each hash are internal run identifiers, so this ledger does not let you recompute a hash from scratch — it lets you check that a published hash never changed. Independent archive snapshots are what turn that into a time-anchored claim.