This is the latest rulebook. Each change applies from the lock its change log line names and never reaches back into a scored week, so a week may be scored under an earlier version; each Commissioner's Report names the version its week was scored under.

Fantasy Singularity Rulebook

Version 1.9, 2026-10-12. This file is the law of the game. A scoring call it does not cover is a change to this file in a pull request, with a version bump and a change log line, never an ad hoc number. Rule changes take effect at the next weekly lock and never reach back into a scored week, except through the correction log (section 11).

Fantasy Singularity is a game. Scores are not investment advice and say nothing about the value of any company or its stock.

1. The week, the lock and event time

1. A scoring week runs from Monday 00:00:00 to Sunday 23:59:59, US Eastern time (America/New_York, so daylight saving time is followed).
2. Entries for a week lock at Sunday 23:59 Eastern on the day before the week starts. After the lock, an entry cannot change.
3. Event time is the publish time of the event's first primary source, recorded in UTC. An event belongs to the week whose Eastern window contains that UTC time.
4. When a primary source shows only a date, the event time is 12:00 UTC on that date. When no source page shows a date, the earliest timestamped public record (such as a repository commit or upload time; for a repository that is public when read, its creation or first weights commit time is taken as public unless a record shows it was private then) showing the event sets the event time, and the adjudicator's reason names that record. When a dated primary announcement exists, it sets the event time unless an earlier record is shown to have been public at its own time; an upload or commit whose public status at the time is not shown does not move the event earlier.
5. An event scores in the week of its first primary publication, even when it is verified later (section 3.5).
6. For a staged rollout, the first primary publication is the first one showing that the public can use the model, product or capability (section 10.4). A preview limited to selected organizations does not set the week, and a later widening of access is not a new event.

2. The entry

1. The roster is in ROSTER.md: the named organizations plus "the field" (every organization not named). Each is a slider.
2. A player allocates whole percentages across the sliders. The allocations sum to exactly 100. Any slider may be 0.
3. The captain is the player's highest slider. When two or more sliders tie for highest, the player picks one; if they don't, the captain is the tied slider listed first in ROSTER.md.
4. A player may play one chip a week (section 7).
5. A player who does not submit an entry for a week has no score for that week. An entry is never copied forward automatically.
6. Open mode is the public game: one entry per player per week, scored against the global pool of all Open mode entries. A league month is a calendar month; a week belongs to the month that contains its Monday. In a league, members are paired head to head each week by the circle method over user ids; an odd count gives one bye. The higher week score wins, equal scores tie, and a member with no entry loses to one who entered. A member counts in a week's pairs and the league table only if they joined before that week's lock. Chips are per player, not per league.

3. Evidence tiers

1. Tier 1. One of these, and nothing else: The organization's own primary publication: its blog, newsroom, model card, system card, paper, documentation, release notes, status page, or official account. A named board: Artificial Analysis (the Intelligence Index), arena.ai (the text leaderboard, overall), or Epoch AI (the Epoch Capabilities Index and the notable models catalog). For ranks, the Epoch AI board is the Epoch Capabilities Index. A filing or official record: a securities filing, a court record, or a regulator's or government's own publication.
2. Tier 2. Two reputable press sources that report the event independently. Neither source may rely only on the other. Reputable press for v1: Reuters, Associated Press, Bloomberg, Financial Times, The Wall Street Journal, The New York Times, The Washington Post, The Information, CNBC, The Verge, Ars Technica, TechCrunch, Wired, Nature News, Science News. pipeline/SOURCES.md may extend this list only through a rulebook change.
3. Tier 3. Everything else, including a single press report, social posts by people other than the organization's official accounts, leaks, rumors, forum posts, benchmark claims the organization makes without a board, and summaries of any of these. Tier 3 never scores.
4. What each tier can score. The table in section 4 names the lowest tier each event type accepts. A negative event always needs Tier 1 (section 5).
5. Pending. A candidate with some evidence that does not yet meet its type's tier is recorded as pending, with the reason. It scores nothing while pending. It becomes final when the evidence arrives, and then scores in its original week (section 1.5). If it is not final within seven days after its week ends, it is dropped and the drop is logged. For a backfill week, or a row first adjudicated more than seven days after its week ends, the seven days run from the date the row is adjudicated.
6. The quoted sentence. Every scored event keeps its source URL and the exact sentence from the source that it rests on. An event without both does not score.

4. Event types and points

Each event earns its organization a fixed number of points. An event scores under one type only, the highest type it qualifies for. A separate publication about the same model (for example a system card published the same day as the model) is a separate event. Several models released together in one publication are one event. A model is new when the combination of its family name, version number and variant name (such as Sonnet, Haiku, Flash, Mini or Turbo) is one the organization has not released before. A dated update of a released model (for example an "October" version), or the general release of a model already released as a public preview, is not a new model, and its system card is not an S+ event. A commitment of money or resources the organization makes to a program, or a grant or donation, is not a deal (B1 or B2) unless the primary source names an agreement with a counterparty for that commitment. "Highest type" means the type worth the most points. Rank 1 includes a shared rank 1 as the board displays it.

Negative types are in section 5.

Weekly caps per organization

An event that exceeds a cap is still recorded in the ledger with its reason; it just adds nothing.

Upgrades

A model scored as M or O that later qualifies as F scores the difference (+20 from M, or +25 or +15 from O) in the week the second board shows it at rank 1. A model never scores F twice. When the M or O event an upgrade builds on happened before the ledger's first week, the F row still scores only the difference, and its reason cites the earlier event's primary source.

5. Negative events

1. A negative event needs Tier 1: the organization's own statement or status page, a court record, or a regulator's or government's own publication. Press alone makes a negative event pending, never final.
2. Every negative event is written neutrally in the ledger and every report: what happened, which organization, when, per which source. It never characterizes intent and never uses words the source does not support.

A lawsuit being filed is not a negative event. Neither is an accusation, an investigation that has not reached a finding, or a press report of any of these.

6. The crowd line

1. Crowd share. For each slider, the crowd share s is the mean allocation, as a fraction, across all Open mode entries in the global pool locked for the week. The crowd line is frozen at the lock. During the week before the lock, the site shows the running line.
2. Equal share. e = 1 / N, where N is the number of sliders (the roster plus the field). With the v1 roster, N = 11 and e = 0.0909.
3. Multiplier. M = (e / s)^a, with a = 0.5, capped at 3.0 and floored at 0.5. A slider with s = 0 takes the cap. This normalizes GRAPH-S's (1 / crowd share)^a by the equal share, so a slider the crowd backs at exactly 1/N pays par (M = 1).
4. Small pools. When fewer than 20 entries are locked for a week, every slider uses M = 1. The four-week backfill uses M = 1 (equal weights). Backfill weeks are scored under the rulebook version in force when the backfill is adjudicated, and are never published as live weeks. Only entries from accounts with a verified email count toward the crowd line. The Commissioner may exclude entries that show coordinated manipulation, with a correction log line.
5. Losses are not scaled. M applies only when an organization's net points for the week are positive. Net negative points apply with M = 1.
6. Where the cap and the floor bite (v1 roster). M reaches the cap of 3.0 at a crowd share of 1.01 percent or less, and the floor of 0.5 at 36.4 percent or more.
7. M is rounded to two decimal places before use.

7. Captain and chips

1. Captain. The captain's term counts double (2x), whether it is positive or negative.
2. Chips. Each player gets these chips per league month. A chip not played by the end of the month is lost. At most one chip a week.

1. A chip is chosen before the lock and cannot be changed after it. A Moonshot pick is checked against the frozen crowd line; if the slider's share at the lock is 5 percent or more, the chip is returned unused and the week scores without it.

8. The score

For each slider o in a player's entry:

• x = the player's allocation in percent (0 to 100)
• P = the organization's net points for the week, after caps
• M = the crowd multiplier from section 6 (1 when P is negative)
• C = 2 for the captain, 3 for a Moonshot pick, 1 otherwise; 0 for a negative term under Hedge

Term = x × P × M × C / 10

Week score = the sum of the terms, plus or minus the Quiet Week result, rounded to the nearest whole number (halves round up). A league month score is the sum of its week scores. The season table is the sum of every week score.

Worked examples (v1 roster, N = 11, e = 1/11)

Crowd shares at the lock: Anthropic 30 percent, OpenAI 25 percent, Moonshot AI 3 percent, DeepSeek 1 percent.

Multipliers:

• Anthropic: (0.0909 / 0.30)^0.5 = 0.303^0.5 = 0.55
• OpenAI: (0.0909 / 0.25)^0.5 = 0.364^0.5 = 0.60
• Moonshot AI: (0.0909 / 0.03)^0.5 = 3.03^0.5 = 1.74
• DeepSeek: (0.0909 / 0.01)^0.5 = 9.09^0.5 = 3.02, capped to 3.0

The week's points: Anthropic +20 (an M event), OpenAI -10 (an X2 outage), Moonshot AI +15 (an O event), DeepSeek +25 (an O event that tops a board).

Player A follows the crowd. Anthropic 60 (captain), OpenAI 40.

• Anthropic: 60 × 20 × 0.55 × 2 / 10 = 132
• OpenAI: 40 × (-10) × 1 × 1 / 10 = -40
• Week score: 92

Player B goes contrarian. Moonshot AI 50 (captain), DeepSeek 30, Anthropic 20.

• Moonshot AI: 50 × 15 × 1.74 × 2 / 10 = 261
• DeepSeek: 30 × 25 × 3.0 × 1 / 10 = 225
• Anthropic: 20 × 20 × 0.55 × 1 / 10 = 22
• Week score: 508

Player C hedges. Anthropic 50 (captain), OpenAI 50, Hedge chip.

• Anthropic: 50 × 20 × 0.55 × 2 / 10 = 110
• OpenAI: the -50 term counts as zero under Hedge
• Week score: 110

Player D calls a quiet week and is wrong. Anthropic 100 (captain), Quiet Week chip. The Anthropic M event breaks the call.

• Anthropic: 100 × 20 × 0.55 × 2 / 10 = 220
• Quiet Week: -50
• Week score: 170

9. The baseline signal

Some weeks nothing ships. The baseline signal keeps them live with small, steady points measured from snapshots the pipeline stores, with the timestamp and URL of each snapshot. A signal that was not captured scores 0 that week; it is never estimated.

A week's start snapshot of a source is the first stored capture of that source taken after the week begins (Monday 00:00 Eastern) and no later than Monday 12:00 UTC. Its end snapshot is the start snapshot of that source for the following week. A week with no start or no end snapshot scores no baseline signal that needs it.

The baseline is capped at +6 per organization per week and counts toward the +80 positive cap. The field earns no baseline.

10. Attribution, the field and mergers

1. Who earns an event. The organization named as the developer, publisher or party in the primary source. A joint release splits its points equally among the named parties (a research paper is not a joint release: each roster organization among its authors' affiliations earns the full points); a share belonging to an organization not on the roster goes to the field. When the shares are not whole numbers, each party first takes the whole-number part of its share, counted toward zero; the points left over go one at a time to the parties with the largest leftover fractions, roster organizations before the field, and ties in ROSTER.md order. A field share in a type the field cannot earn (section 4's caps table) is recorded and adds nothing. A B1 or B2 deal is the exception: every named roster party earns the full points, and a party not on the roster earns nothing. The field earns a deal only when no roster organization is a party to it.
2. Products built on another's model. An organization that ships another organization's model inside its own product earns only the product event. The model event belongs to the developer.
3. The field. The field slider earns the events of every organization not on the roster, under the same tiers and caps as one organization, limited to the types in the field row of section 4's caps table, with no baseline.
4. Announcements and public use. A model, product or capability scores when the public can use it, not when it is announced or previewed to a waitlist. Public use includes a paid plan that anyone can buy on their own. It excludes waitlists, invitation-only access, and access that needs an eligibility check or a call with sales. A free API limited to non-commercial use counts as public when anyone can sign up for it without an eligibility check. A plan limited to listed countries or markets is public in those markets; a market list is not an eligibility check.
5. Mergers and acquisitions. When a completed merger or acquisition joins two roster organizations, both sliders stay as they are until the end of the league month. Each event goes to the entity named in its primary source. The roster is updated at the next monthly reset. When an organization in the field is acquired by a roster organization, its events from the completion date go to the roster organization. A rename keeps the same slider.

11. Reports, disputes and corrections

1. The Monday report. The Commissioner's Report for a week is published after the lock on the following Monday. It shows every event with its source and quoted sentence, each organization's points, and provisional scores.
2. The dispute window. For 48 hours after the report is published, any player may dispute an event or a missing event by giving a source. The Commissioner rules within 72 hours of the dispute. D.J. hears appeals filed within 48 hours of a ruling.
3. Final scores. A week's scores become final at 23:59 Eastern on the second Monday after the week ends, or when the last open appeal for that week is decided, whichever is later. After that, a change happens only through the correction log.
4. The correction log. Every change to a scored event, its points or a week's scores is written to the event's corrections field in the ledger and to the correction log in the next report: what changed, why, the source, and the date. Nothing is changed silently.

12. The not-advice line

Every surface that shows a score, a point value, an organization's standing or the crowd line carries this line:

"Fantasy Singularity is a game. Scores are not investment advice."

Change log

• 1.0 (2026-10-08): First version, with the opus read's findings applied before merge.
• 1.1 (2026-10-08): Gaps met in backfill week 2026-09-07. M now names the main general-purpose model line, with variants and public previews, and sends specialty lines to P2 or P1 (section 4). Public use now covers self-serve paid plans and excludes eligibility checks and sales calls (section 10.4). Non-roster deal parties earn nothing (section 10.1). Backfill weeks score under the version in force at adjudication, and their pending rows get seven days from adjudication (sections 6.4 and 3.5).
• 1.2 (2026-10-09): Gaps met in closing week 2026-09-07. A staged rollout's week is set by the first public availability (section 1.6). A specialty-line release to the API and a consumer product at once scores once, as P1 (section 4). R1 needs an independent party's proof check; the organization's own formalization does not count (section 4). A free non-commercial API that anyone can sign up for is public (section 10.4).
• 1.3 (2026-10-09): Gaps met in week 2026-09-14. A timestamped public record may set the event time when no page is dated (section 1.4). Open weights include research and non-commercial licenses (O). Only the listed safety publications score (S+). A multi-year expected total is a disclosed value (B1). Several named partners are several deals (B2).
• 1.4 (2026-10-09): An F upgrade whose M or O event predates the ledger scores only the difference and cites the earlier source (section 4, Upgrades).
• 1.5 (2026-10-09): Gaps met in week 2026-09-21. Several models in one publication are one event, and omni-modal main-line models are M (section 4). The single-P1 rule for a specialty release now covers API and consumer releases in the same scoring week, not only at once (section 4). A program provided to a named government under an agreement is a partnership (B1). A new AI device generation on sale is P1; a pre-order is not (section 4). The seven-day pending clock runs from adjudication for a backfill week or a row first adjudicated more than seven days after its week (section 3.5).
• 1.6 (2026-10-09): Gaps met in week 2026-09-28. A dated announcement sets the event time unless an earlier record is shown to have been public then (section 1.4). Speed and service tiers are not P2 (section 4). A market-limited plan is public in its markets (section 10.4). A public repository's creation or first weights commit counts as public unless shown private (section 1.4). A research paper earns full points for each roster organization among its authors (section 10.1). ROSTER.md 1.1: the Alibaba Qwen slider covers all of Alibaba's AI models, including DAMO Academy.
• 1.7 (2026-10-10): League head-to-head pairing, results and late joiners (section 2.6). Scoring of points and player scores is unchanged.
• 1.8 (2026-10-12): Gaps met in live week 2026-10-05, in force from the lock of Sunday, October 18, 2026 (the week of October 19); per the rule at the top of this file they do not reach back into week 2026-10-05, whose four rows that wait on these questions stay pending and drop on October 18. A new model is a family, version and variant combination the organization has not released before; a dated update or the general release of a public preview is not, and its card is not S+ (section 4). A commitment, grant or donation to a program is not a deal without a named agreement with a counterparty for it (section 4, B1 and B2). An opt-in setting that adds a new API function is a P2 tool; a level, length, speed or price setting is not (section 4, P2). Split shares are whole points by largest remainder, roster organizations first (section 10.1).
• 1.9 (2026-10-12): Baseline snapshot times the pipeline can meet, in force from the lock of Sunday, October 18, 2026 (the week of October 19): a week's start snapshot is the first taken after Monday 00:00 Eastern and no later than Monday 12:00 UTC, and its end snapshot is the next week's start snapshot; board movement, token share and store rank read those snapshots; board movement now also scores a rise into the top 10 from below it (section 9).