We publish the predictionbefore the outcome exists.

The record

Counts, not a rate

The record today.

Sealed publicly
3
Voided as unjudgeable
0
Correct
1
Wrong
0
Still open
2

Counts, not a rate: a handful of judged predictions cannot carry one. Voided ones were withdrawn because they could never have been settled, and stay published below with their reasons.

We have audited every prediction on this page, voided the ones we had written badly and corrected the verdicts we had got wrong. Those dispositions are still being applied to the public record, so the statuses below are the ones originally published until they land. The audit sets out what changed and why.

Read the audit

From the public record, 16 November 2026

Judged

1

What we said, and what happened.

What we sealed beforehand, beside what was recorded against it. Misses appear here with the same weight as hits.

  • What we said

    Apple Plans Price Increases on Select Products Due to Memory Chip Shortage

    High risk, sealed 23 June 2026

    It had to meet: Within 45 days, Apple announces specific price increases on at least one major product line (iPhone or MacBook), confirmed via official Apple pricing pages or a press release.

    What happened

    Came out as predicted

    On 25 June 2026, two days after this prediction was sealed, Apple announced price increases across Macs and iPads, blaming a memory chip shortage driven by demand from AI data centres. Apple called the demand spike an unprecedented challenge for the consumer electronics industry. The criterion was a weak test: reporting at the time already said a price rise was planned, so this counts as a hit but a near-certain one.

    Sources: apnews.com, iclarified.com

    Judged 3 August 2026

All predictions

3

Every sealed prediction.

Every call, with its verdict and judgement date. Open a row for the criterion and the hash.

None are past their judgement date without an outcome recorded.

Status

Showing 3 of 3 predictions

  • Resolution criterion

    The latest Deutschlandticket subscriber count published by VDV or the federal government by 15 June 2027, for a month between February and May 2027, is at least 13.92 million, meaning no more than about 4% of the roughly 14.5 million pre-increase holders cancelled after the January 2027 price rise to €66.80.

    Prediction hash
    b0fadd5f8f30a97ffdb2410137004dc455808b5f1088d8b514ff006f9a73a7ca
    Sealed
    25 October 2026
    Verdict
    Low risk
    Decision type
    Pricing
    Outcome
    Not yet judged
  • Resolution criterion

    Disney's fiscal Q1 2027 earnings release reports Entertainment Direct-to-Consumer subscriber or revenue growth at least 17 percentage points below the growth trend implied by the prior year-over-year quarter, consistent with roughly a fifth of affected subscribers actually leaving or downgrading.

    Prediction hash
    621a2c387728fb2452ab6359e78170df0d55897f3da4742aa833f003afb0df6c
    Sealed
    25 October 2026
    Verdict
    Critical risk
    Decision type
    Pricing
    Outcome
    Not yet judged
  • Resolution criterion

    Within 45 days, Apple announces specific price increases on at least one major product line (iPhone or MacBook), confirmed via official Apple pricing pages or a press release.

    Prediction hash
    39938ccdff64486b7ea31483eb421a68882f69269a6161f09b1274b2ce3de7d0
    Sealed
    23 June 2026
    Verdict
    High risk
    Decision type
    Pricing
    Outcome
    Verified, outcome recorded 3 August 2026

Backtests

5

Selected backtests,kept apart from the sealed record.

Backtests are simulations of past decisions run after their outcomes were already public. The language models behind the engine may have seen news of these outcomes in training, so a backtest shows the engine is consistent with what happened — not that it could have foreseen it. Each outcome rule was written down before the simulation ran. Backtests are counted here only, never in the forward-prediction record.

  • Backtest — run after the outcome was public

    Spotify raises US Premium prices for the first time

    Decision date: 24 July 2023 · Simulation run: 4 October 2026 · Engine: v2

    Simulated customer behaviour

    Cleared

    4.8% of simulated customers took an adverse action within 90 days

    What customers actually did

    Matched

    Premium subscribers kept growing at the same pace after the rise (220M to 226M to 236M); Spotify said gross intake in price-rise markets was better than expected and churn showed no real change.

    Mean net change per period: before +7.5, after +8 million subscribers/quarter

    Sources: Spotify / SEC, Spotify / SEC, CNBC

    • bt-rule-v1 (fixed before any backtest was run; the same rule for every case). One primary customer-behaviour metric per case, chosen before the run: the customer count in the affected market, or the reported churn rate where the decision hits existing customers and counts are not the cleaner measure. COUNT metric: take the metric at the end of five consecutive reporting periods L-2..L2, where L0 is the last period ending before the decision took effect. Before = (L0 - L-2)/2, after = (L2 - L0)/2 (mean net change per period). Observed 'adverse' if after < 0, or before > 0 and after < 0.40 x before; 'tolerated' if after >= 0 and (before <= 0 or after >= 0.75 x before); otherwise 'unclear'. RATE metric: mean of the two periods ending before the effective date vs the first two periods ending on or after it; after/before <= 1.10 is 'tolerated', >= 1.25 is 'adverse', otherwise 'unclear'. REVERSAL: if the company withdrew or materially reversed the decision within 90 days citing customer reaction, the observed group is 'adverse' regardless of the metric. Missing figures give 'unclear'. SCORING: the engine's behaviour verdict folds into 'tolerated' (CLEARED, CLEARED_WITH_MODIFICATIONS) or 'adverse' (HIGH_RISK, DO_NOT_LAUNCH). Hit = same group as observed; miss = different group; unclear = either side has no group (e.g. INSUFFICIENT_EVIDENCE, or an 'unclear' observation). Every scored backtest counts in the denominator, misses and unclear included. This case: metric — Global Premium subscribers, millions, quarter end; decision announced 2023-07-24, effective 2023-07-24.

    • Spotify Premium — price change (announced 24 July 2023). Spotify is raising Premium prices in the US and other markets, the first increase of the US Individual plan since launch: Individual $9.99 to $10.99 a month, Duo $12.99 to $14.99, Family $15.99 to $16.99, Student $4.99 to $5.99. New subscribers pay the new price immediately; existing subscribers are notified and get a one-month grace period. Spotify had 210 million Premium subscribers at the end of March 2023. Apple Music and Amazon Music Unlimited (non-Prime) already cost $10.99, and YouTube Music Premium rose to $10.99 in July 2023.

  • Backtest — run after the outcome was public

    Peacock raises prices for the first time

    Decision date: 17 July 2023 · Simulation run: 4 October 2026 · Engine: v2

    Simulated customer behaviour

    Cleared

    8.3% of simulated customers took an adverse action within 90 days

    What customers actually did

    Matched

    Paid subscribers grew faster after the rise: 24M at the end of June 2023, 28M in September and 31M in December.

    Mean net change per period: before +2, after +3.5 million subscribers/quarter

    Sources: Comcast / SEC, Comcast / SEC, Comcast / SEC

    • bt-rule-v1 (fixed before any backtest was run; the same rule for every case). One primary customer-behaviour metric per case, chosen before the run: the customer count in the affected market, or the reported churn rate where the decision hits existing customers and counts are not the cleaner measure. COUNT metric: take the metric at the end of five consecutive reporting periods L-2..L2, where L0 is the last period ending before the decision took effect. Before = (L0 - L-2)/2, after = (L2 - L0)/2 (mean net change per period). Observed 'adverse' if after < 0, or before > 0 and after < 0.40 x before; 'tolerated' if after >= 0 and (before <= 0 or after >= 0.75 x before); otherwise 'unclear'. RATE metric: mean of the two periods ending before the effective date vs the first two periods ending on or after it; after/before <= 1.10 is 'tolerated', >= 1.25 is 'adverse', otherwise 'unclear'. REVERSAL: if the company withdrew or materially reversed the decision within 90 days citing customer reaction, the observed group is 'adverse' regardless of the metric. Missing figures give 'unclear'. SCORING: the engine's behaviour verdict folds into 'tolerated' (CLEARED, CLEARED_WITH_MODIFICATIONS) or 'adverse' (HIGH_RISK, DO_NOT_LAUNCH). Hit = same group as observed; miss = different group; unclear = either side has no group (e.g. INSUFFICIENT_EVIDENCE, or an 'unclear' observation). Every scored backtest counts in the denominator, misses and unclear included. This case: metric — Peacock paid subscribers, millions, quarter end (rounded by Comcast); decision announced 2023-07-17, effective 2023-08-17.

    • Peacock — price change (announced 17 July 2023, effective 17 August 2023). Peacock Premium (with ads) rises from $4.99 to $5.99 a month and Premium Plus (ad-free) from $9.99 to $11.99 — the first increase since the service launched in 2020. New subscribers pay the new prices from 17 August; existing subscribers move on their next billing date after that. Earlier in 2023 Peacock closed its free tier to new users and ended free Premium for Xfinity customers. Peacock had 22 million paid subscribers at the end of March 2023.

  • Backtest — run after the outcome was public

    Netflix ends free password sharing in the US

    Decision date: 23 May 2023 · Simulation run: 4 October 2026 · Engine: v2

    Simulated customer behaviour

    Cleared with modifications

    9.7% of simulated customers took an adverse action within 90 days

    What customers actually did

    Matched

    US and Canada paid memberships kept growing after the change: 74.40M at the end of Q1 2023, 75.57M in Q2 and 77.32M in Q3 — net additions accelerated rather than reversed.

    Mean net change per period: before +0.505, after +1.46 million members/quarter

    Sources: Netflix

    • bt-rule-v1 (fixed before any backtest was run; the same rule for every case). One primary customer-behaviour metric per case, chosen before the run: the customer count in the affected market, or the reported churn rate where the decision hits existing customers and counts are not the cleaner measure. COUNT metric: take the metric at the end of five consecutive reporting periods L-2..L2, where L0 is the last period ending before the decision took effect. Before = (L0 - L-2)/2, after = (L2 - L0)/2 (mean net change per period). Observed 'adverse' if after < 0, or before > 0 and after < 0.40 x before; 'tolerated' if after >= 0 and (before <= 0 or after >= 0.75 x before); otherwise 'unclear'. RATE metric: mean of the two periods ending before the effective date vs the first two periods ending on or after it; after/before <= 1.10 is 'tolerated', >= 1.25 is 'adverse', otherwise 'unclear'. REVERSAL: if the company withdrew or materially reversed the decision within 90 days citing customer reaction, the observed group is 'adverse' regardless of the metric. Missing figures give 'unclear'. SCORING: the engine's behaviour verdict folds into 'tolerated' (CLEARED, CLEARED_WITH_MODIFICATIONS) or 'adverse' (HIGH_RISK, DO_NOT_LAUNCH). Hit = same group as observed; miss = different group; unclear = either side has no group (e.g. INSUFFICIENT_EVIDENCE, or an 'unclear' observation). Every scored backtest counts in the denominator, misses and unclear included. This case: metric — UCAN (US + Canada) paid memberships, millions, quarter end; decision announced 2023-05-23, effective 2023-05-23.

    • Netflix US — change to account sharing (announced 23 May 2023, effective immediately). A Netflix account is for use by one household. From today Netflix will email US members whose account is being used by people outside their household. Those people can transfer their profile — keeping viewing history and recommendations — to a new account they pay for themselves, or the account owner can add an 'extra member' for $7.99 a month (one extra member on the Standard plan, two on Premium). Members can still watch while travelling. US prices are unchanged: Basic with ads $6.99, Basic $9.99, Standard $15.49, Premium $19.99 a month. Netflix estimates more than 100 million households worldwide share accounts. Paid sharing launched in Canada, New Zealand, Portugal and Spain earlier in 2023.

  • Backtest — run after the outcome was public

    Adobe stops selling Creative Suite licences: Creative Cloud only

    Decision date: 6 May 2013 · Simulation run: 4 October 2026 · Engine: v2

    Simulated customer behaviour

    Cleared

    7.1% of simulated customers took an adverse action within 90 days

    What customers actually did

    Matched

    Paid Creative Cloud subscriptions accelerated after the switch, from 700,000 at the end of May 2013 to 1.03 million in August and 1.44 million by the end of November, despite a widely publicised customer petition.

    Mean net change per period: before +187000, after +369500 subscriptions/quarter

    Sources: Adobe / SEC, Adobe / SEC, Adobe / SEC

    • bt-rule-v1 (fixed before any backtest was run; the same rule for every case). One primary customer-behaviour metric per case, chosen before the run: the customer count in the affected market, or the reported churn rate where the decision hits existing customers and counts are not the cleaner measure. COUNT metric: take the metric at the end of five consecutive reporting periods L-2..L2, where L0 is the last period ending before the decision took effect. Before = (L0 - L-2)/2, after = (L2 - L0)/2 (mean net change per period). Observed 'adverse' if after < 0, or before > 0 and after < 0.40 x before; 'tolerated' if after >= 0 and (before <= 0 or after >= 0.75 x before); otherwise 'unclear'. RATE metric: mean of the two periods ending before the effective date vs the first two periods ending on or after it; after/before <= 1.10 is 'tolerated', >= 1.25 is 'adverse', otherwise 'unclear'. REVERSAL: if the company withdrew or materially reversed the decision within 90 days citing customer reaction, the observed group is 'adverse' regardless of the metric. Missing figures give 'unclear'. SCORING: the engine's behaviour verdict folds into 'tolerated' (CLEARED, CLEARED_WITH_MODIFICATIONS) or 'adverse' (HIGH_RISK, DO_NOT_LAUNCH). Hit = same group as observed; miss = different group; unclear = either side has no group (e.g. INSUFFICIENT_EVIDENCE, or an 'unclear' observation). Every scored backtest counts in the denominator, misses and unclear included. This case: metric — Paid Creative Cloud subscriptions, fiscal quarter end; decision announced 2013-05-06, effective 2013-06-17.

    • Adobe — Creative Suite to Creative Cloud (announced 6 May 2013, new apps from 17 June 2013). Adobe will no longer develop new versions of Creative Suite. Future versions of Photoshop, Illustrator, InDesign, Premiere Pro and the other applications will be available only through a Creative Cloud subscription: the full set for $49.99 a month on an annual plan, or a single app for $19.99 a month. Owners of CS3 or later get the first year of the full set for $29.99 a month. CS6 stays on sale with bug and security fixes but no new features. Creative Cloud launched in May 2012 and had more than 500,000 paying members at the announcement.

  • Backtest — run after the outcome was public

    Netflix splits DVD and streaming into separate plans

    Decision date: 12 July 2011 · Simulation run: 4 October 2026 · Engine: v2

    Simulated customer behaviour

    Do not launch

    24% of simulated customers took an adverse action within 90 days

    What customers actually did

    Matched

    Netflix lost about 800,000 US subscribers in the quarter the new prices reached existing members (24.59M to 23.79M) and was still below its June level at year end (24.40M); DVD subscriptions fell sharply.

    Mean net change per period after: -0.095 million subscribers/quarter

    Sources: Netflix / SEC, Netflix / SEC

    • bt-rule-v1 (fixed before any backtest was run; the same rule for every case). One primary customer-behaviour metric per case, chosen before the run: the customer count in the affected market, or the reported churn rate where the decision hits existing customers and counts are not the cleaner measure. COUNT metric: take the metric at the end of five consecutive reporting periods L-2..L2, where L0 is the last period ending before the decision took effect. Before = (L0 - L-2)/2, after = (L2 - L0)/2 (mean net change per period). Observed 'adverse' if after < 0, or before > 0 and after < 0.40 x before; 'tolerated' if after >= 0 and (before <= 0 or after >= 0.75 x before); otherwise 'unclear'. RATE metric: mean of the two periods ending before the effective date vs the first two periods ending on or after it; after/before <= 1.10 is 'tolerated', >= 1.25 is 'adverse', otherwise 'unclear'. REVERSAL: if the company withdrew or materially reversed the decision within 90 days citing customer reaction, the observed group is 'adverse' regardless of the metric. Missing figures give 'unclear'. SCORING: the engine's behaviour verdict folds into 'tolerated' (CLEARED, CLEARED_WITH_MODIFICATIONS) or 'adverse' (HIGH_RISK, DO_NOT_LAUNCH). Hit = same group as observed; miss = different group; unclear = either side has no group (e.g. INSUFFICIENT_EVIDENCE, or an 'unclear' observation). Every scored backtest counts in the denominator, misses and unclear included. This case: metric — Total unique US subscribers, millions, quarter end; decision announced 2011-07-12, effective 2011-07-12.

    • Netflix US — plan change (announced 12 July 2011). Netflix is separating its DVD-by-mail and streaming plans. The $9.99 plan with unlimited streaming plus one DVD at a time is replaced by two plans at $7.99 each — $15.98 a month for members who keep both, up to a 60% increase. Streaming-only stays at $7.99 and DVD-only (one at a time) is $7.99. New members pay the new prices now; existing members move at their September billing date. Netflix had 22.8 million US subscribers at the end of March 2011.

How it works

Three steps

How a prediction is sealed.

  1. We write a criterion that states exactly what would count as a hit, and by when.
  2. The simulation id, title, verdict and judgement date are hashed with SHA-256 at the moment of registration, so the call cannot be revised afterwards. The criterion is published in the open beside it.
  3. On the judgement date the outcome is recorded as a hit or a miss, with the evidence cited, and published either way.

Check the record yourself.

The track record and every certificate are public and need no account.