Inverse Open-Weight AI Lead — compiled methodology

Over tech-science + companies, questions naming agi, gpt, openai…, read through the weighted mean lens, taking the against side of the thesis "Closed-weight labs lead the frontier AI race.", at horizon 7–365 days, priced across kalshi, gemini, forecastex, combined by two-sided-depth weights capped at 20%.

experimental class · closed-weight-ai-lead · spec v1 · schema worldview-spec/v1

Spec digest	8f267db96ac28e41
Methodology hash	8a45f3768b0d5a00
Compiler	methodology-compiler/v1
Compiled against	no settled history

This document is generated from the spec object. Every parameter it prints is read from that object or from the engine constants the close ran under; nothing in it is hand-written, and it recompiles byte-for-byte from the same inputs forever.

---

1. The worldview

Over tech-science + companies, questions naming agi, gpt, openai…, read through the weighted mean lens, taking the against side of the thesis "Closed-weight labs lead the frontier AI race.", at horizon 7–365 days, priced across kalshi, gemini, forecastex, combined by two-sided-depth weights capped at 20%.

Thesis. Closed-weight labs lead the frontier AI race.

Formal definition. The depth-weighted mean, over the frontier-AI capability and leaderboard board, of the market-priced probability that an open-weight lab holds the leading position. Every mutually exclusive contest ladder collapses to ONE constituent carrying that contest's aligned mass x_F = Sigma_{k in F+} q_k / Sigma_{k in F} q_k, so a 20-leg ladder contributes one voice rather than twenty correlated near-duplicates.

Stance. This reading takes the against side of the thesis. Stance is a whole-index flip and it composes with each member's own declared side by XOR, so a member declared on the same side as the stance is read from its plain venue YES mark and the double flip is arithmetic rather than a special case.

Evaluation rule. Scored at each contest's resolution. A contest resolves FOR when the winning outcome is in the declared for set, AGAINST when it is in the against set, and is NOT SCORED when the winner is unaligned — an outcome the thesis does not read is a gap in the partition, never a silent win for either side.

Scope of the evaluation. Only the contest legs are scoreable. Capability-milestone members (a benchmark threshold, a model release by a date) resolve individually and are scored individually; the index level itself is a continuous reading and is tracked for movement, not scored as a single forecast.

Inverse thesis (wet-logic §6.1, mandatory). Open-weight labs lead the frontier AI race.

Invalidation conditions.

• The leaderboard contests stop listing open-weight labs as distinct outcomes, so the partition has nothing to separate.
• A lab in the for set ships a leading model under a licence that is not open-weight, or a lab in the against set open-weights its frontier model — either makes the declared partition describe the wrong world, and the spec is superseded rather than edited.
• Fewer than four priced constituents survive eligibility, at which point the index prints nothing rather than a number carried by two markets.

What the published number is NOT.

• It is not the probability that the thesis holds. It is a weighted reading of what a board of separate markets is pricing (wet-logic §6.5). Several events do not combine into one probability by averaging them, and this number never claims they do.
• It is not a financial return, and not a performance claim. No position is taken, no order is routed and no custody is held.
• It is not advice. Where accuracy is reported anywhere on this desk it is reported as calibration — how well priced probabilities matched outcomes — never as profit.

---

2. Inverse twin

Twin: open-weight-ai-lead — minted at the same moment, over an identical universe with identical base weights, caps, collapses and rebalance schedule, with every side flipped and the thesis replaced by the declared inverse thesis. The transform is an involution: the twin of the twin is this index.

Gauge space — an exact complement. Under one canonical consolidated qⱼ per market and shared normalized weights,

Ḡ_t = 100 · Σⱼ wⱼ (1 − x_jt) = 100 − G_t

exactly, at every print, on the mid basis. Both documents print the measured daily residual

ρ_t = G_t + Ḡ_t − 100,   |ρ_t| ≤ 0.000001

in their worked-example sections (§14). It is printed so that the identity is verified on the page rather than asserted there: a residual beyond the rounding bound exposes a basis or roster error between the two series — different weights, a member in one and not the other, a side that did not flip.

Level space — no complement, by construction. Chained twins chain independently: each incepts at 100 with its own divisor struck at its own first governed close, and rebalances rechain each divisor separately. A chained level is a path object — a product of period returns — and returns of x and 1 − x are not complementary, so

L_t + L̄_t ≠ constant

and Ḻ is never derived as 200 − L. Twins are complementary in gauge space, not in level space, and no statement in this document may imply otherwise.

The Friction Gauge. Under side-executable pricing a YES entry strikes at aᵞ and the twin's NO entry at 1 − bᵞ, so per leg the pair sums to 1 + σⱼ where σⱼ = aᵞⱼ − bᵞⱼ is the spread. Hence, exactly:

G^exec_t + Ḡ^exec_t = 100 + Φ_t,   Φ_t = 100 · Σⱼ wⱼ σ_jt

Φ — the capacity-weighted spread — is published as a data product in its own right. The gap between the twins is the market's friction, and it is disclosed rather than hidden as an embarrassment.

---

3. Universe and canonical events

The selector, verbatim as data. This is the canonical admission object — sorted, defaults resolved, entity ids folded through literal/v1 — and it is the same bytes the admission engine reads and the same bytes the spec digest is taken over. It is printed as data rather than described in prose so that a reader can re-run the admission rather than reconstruct it.

{
  "categories": [
    "companies",
    "tech-science"
  ],
  "entities": [],
  "excludeKeywords": [
    "ipo",
    "market cap",
    "share price",
    "stock"
  ],
  "keywords": [
    "agi",
    "ai model",
    "anthropic",
    "benchmark",
    "best ai model",
    "chatgpt",
    "claude",
    "deepseek",
    "frontier",
    "gemini",
    "gpt",
    "grok",
    "leading ai",
    "llama",
    "llm",
    "mistral",
    "model release",
    "open source model",
    "open weight",
    "openai",
    "qwen"
  ],
  "minConstituents": 4,
  "series": [],
  "tags": [],
  "venueWhitelist": []
}

Admission only. Nothing in the object above decides a side. Admission tests universe membership; side rules assign each member's dⱼ ∈ {−1, +1} afterwards (§6), and no side rule may ever eject a member. Narrowing a side rule re-orients; it cannot shrink the roster. A universe object carrying a side-bearing key is refused by name at load time, which makes the August 2026 incident — a side-rule narrowing that dropped "Will US withdraw from NATO by August 31?" out of a live index instead of re-orienting it, and had to be reverted — unrepresentable rather than merely fixed.

Canonical events, not venue listings. Members are canonical events: two venues quoting the same question are one member, marked once and projected to every index that holds it. A duplicate canonical event in one roster is refused outright (duplicate-event) — one question voting twice is the duplicate-exposure defect wet-logic §19 forbids.

Mapping confidence. A cross-venue merge is formed either by an exact structural match on the parsed subject, relation, strike and settlement day, or by title similarity above the published confidence bar. A fuzzy merge below that bar is flagged low-confidence and carries its provenance onto the row; the rules-based approval leg (wet-logic §4.3) admits a structurally matched pair without a human review, and the human confirm gate applies to merges only.

No confirm-gate counts are published for this index yet, and none are invented here. They arrive on the first close row and this paragraph reads them from it.

No venue is admitted beyond the default set forecastex, gemini, kalshi. A whitelist entry is a methodology decision that has to be written down; the absence of one is also a decision and is stated here rather than left to be inferred.

---

4. Input data and pricing

The per-venue mark ladder. A member's venue mark is the consolidated mid when both sides of the book exist; failing that the recent last trade; failing that it is unmarkable and the member carries no price. It is never marked to zero. Marks are clamped to [0.01, 0.99] and taken from the venue of record — never mark-shopped across venues for a better number.

Consolidation as executed, stamped consolidation/prob-v1. Each member's price is a weighted mean of the surviving venue mids, taken in probability space. The close rows stamp the policy version the close ran under (consolidation/v1); the two describe the same executed pipeline from two angles and both are printed so neither has to be inferred.

q_j = Σ_v (w_v · m_v) / Σ_v w_v,        m_v = (bid_v + ask_v) / 2

w_v = (1 + 2·ln(1 + depth_v) + ln(1 + volume_v)) · tightness_v · e^(−Δt_v / λ)
tightness_v = max(0.05, 1 − spread_v / 0.1)
λ = 1,500 s

The depth term is inert, and has been on every published value. The close path supplies no resting-depth figure to the consolidation — depth is passed as null for every venue quote on every basis — so ln(1 + depth_v) evaluates to ln 1 = 0 and the doubled term contributes nothing. The weight actually executed is therefore:

w_v = (1 + ln(1 + volume_v)) · tightness_v · e^(−Δt_v / λ)

The general form is printed above it because it is the expression the engine evaluates and the one a future close would run under if depth were ever supplied — but a third party reproducing a published number must use the reduced form, and a check against the general form with real depths will disagree. The disagreement is not small, and these figures are produced by running the consolidation twice rather than by being typed here. Two venues quoting 0.40/0.42 and 0.50/0.52 on equal lifetime volume and equal staleness consolidate to q = 0.4600 on the executed weight (shares 0.5000 / 0.5000); hand the same two books resting depths of 5,000 and 100 and the general form returns q = 0.4507 (shares 0.5927 / 0.4073) — 0.93 probability points, entirely from a term the engine has never had an input for.

Why the field is empty is a weighting decision rather than an oversight. One mapped venue publishes resting size at the touch and the others publish none, so populating it today would hand a doubled depth term to that venue and zero to the others, tilting every cross-venue consolidation toward it for a reason that is an artefact of which venue documents a size field. Wiring it in is a change to the weighting scheme: it needs the other venues measured first, and it arrives as a versioned change with its own notice and backtested impact, never as a quiet population of an already-printed term.

This says nothing about the weight basis in §8. That is a different depth: b_j = Σ_v min(depth^bid, depth^ask), two-sided resting size near touch, served by the mark recorder's real ladders and reported per close as observed / one-sided / unobserved counts. It is live. Flattening the two would tell a reader the weighting is fabricated, which is false.

Cross-venue dispersion is preserved beside q_j, never folded inside it. The consolidated price answers "what is this worth"; the dispersion answers "how far apart are the venues about it", and one number cannot carry both questions.

The exclusion cascade, in the order it is evaluated. The venue-local guards run per quote, first match wins; the consensus trim runs afterwards, on what survived them.

Refused as	Tier	Threshold in force	What it means
no-two-sided-book	A — no price exists	a bid and an ask are both required	there is no mid, so there is no price to include
crossed-book	A — no price exists	ask < bid	the book is inconsistent with itself; a mid taken across it is arithmetic, not a price
spread-too-wide	B — venue-local quality	ask − bid > 0.1	a mid inside a book this wide is not a price anyone is defending. Excluded, never down-weighted — a bad price scaled down is still a bad price
degenerate-price	A — no price exists	mid not strictly inside (0, 1)	a mid at or beyond the bounds of probability is not a probability
stale-beyond-horizon	B — venue-local quality	age > 21,600 s (6 h)	past this age the quote is dropped outright rather than decayed toward zero
thin-book-anomaly	B — venue-local quality	|last − mid| > 0.15 while venue volume < 100	a venue that printed a large move on trivial size has told us its book is not defended. Checked BEFORE weighting, so deep historical volume cannot launder it
mad-outlier	C — consensus	|mid − median| > 3 × MAD, and only where ≥ 3 venues survived the guards above	decided only by comparison with the other venues. Median absolute deviation, not standard deviation: with three or four venues one bad print moves a standard deviation enough to hide inside it

The tiers are not decoration. A Tier A refusal is forced — no mid exists. A Tier B refusal is a judgement about what counts as a defended price, decided without looking at any other venue, which is what makes excluding it non-circular. Tier C is decided only by comparison with the other venues, which is right for a consolidated price and fatal for a measurement of disagreement — so any dispersion reading is frozen before the consensus trim runs and the trimmed venue is still inside it.

A member whose every venue quote is refused publishes no price and is counted, by reason, on the close row. Members priced from a single venue are flagged on every row.

Future methodology — declared, not in force

A log-odds aggregation is the declared migration target consolidation/logit-v2 and no published value has ever been produced by it. It is printed here, inside a subsection whose heading says it is not in force, because a thousand documents need one place that states what is coming; it appears nowhere else in this document and a fixture enforces that.

p̃ = min(1−ε, max(ε, p)),   z = ln( p̃ / (1 − p̃) )
ᾱ_jv = L · F · R · C,      z̄_j = Σ_v α_jv z_jv / Σ_v α_jv,      q_j = σ(z̄_j)
D_j = sqrt( Σ_v α_jv (z_jv − z̄_j)² / Σ_v α_jv )     (divergence, preserved beside q_j)

Until it ships, every document on this desk prints consolidation/prob-v1 and the arithmetic above this subsection is the arithmetic that ran. A document that printed math the engine does not run is the defect class this compiler exists to kill, so the two are separated by a heading rather than by a footnote.

---

5. Eligibility

Every threshold in force, with the published default beside it. Version eligibility/v1.

Parameter	Published default	In force here	Source
Minimum trailing volume	250	250	published default
Maximum spread	0.100	0.100	published default
Price band — floor	0.020	0.020	published default
Price band — ceiling	0.980	0.980	published default
Shortest horizon (days)	1	1	published default
Longest horizon (days)	730	730	published default
Confirm gate enforced	no	no	published default
Admit reference-price settlement	no	no	published default
Constituent floor	8	4	per-spec
Resolution window (days)	—	7 – 365	per-spec horizon axis

No eligibility parameter is overridden. Every threshold above is the published default, which is a statement about this index and not an absence of one.

Resolution-source tiers. A member is admitted only where the settlement authority is one this desk can name. The tiers, in rank order:

Rank	Tier	Label	Admitted
1	venue-discretion	Venue discretion	no
2	major-outlet	Major-outlet consensus	yes
2	reference-price	Published reference price	yes
3	court	Court or regulator ruling	yes
4	government	Official data or an official act	yes

A member whose settlement rests on the listing venue's own discretion is measuring the venue rather than the world, and is refused for that reason. Reference-price settlement is refused by default and admitted only where a spec declares it, because absent rules text means the settlement authority cannot be verified and failing closed is the only safe default when a price is about to be averaged into a published number.

Venue admission. At least one contributing venue must be CFTC-regulated, or named on this spec's venue whitelist (§3). Venues outside both may inform the consolidated price of an already-admitted member but can never be the sole basis for one.

What happens when the thresholds bite is not an abort. This desk publishes a graded ladder with a hard floor rather than a cliff, with this index's own κ_min = 95.00%; the full table, every threshold and the resulting published state are in §11.

---

6. Side selection and framing

Elected framing axis: alignment — the THESIS declares a partition of a contest's outcomes, and the side follows the partition.

Orientation (wet-logic §6.3). Every formula in this document is stated and computed over the oriented value xⱼ, never the raw venue price qⱼ:

x_j = q_j        when d_j = +1 (the member is held YES)
x_j = 1 − q_j    when d_j = −1 (the member is held NO)

with the invariant that an increase in xⱼ always means stronger market support for the thesis. A mixed YES/NO index whose published formula read qⱼ would be publishing a formula the engine does not run.

Side is a declaration, never an admission filter. A member matching the universe is admitted with d = +1 by default; a side rule flips it to −1; no side rule ever ejects a member.

The alignment maps, verbatim. A contest carries no risk instrument, no severity noun and no numeric threshold, so the text-reading axes are correctly silent on it — which is why an unpartitioned contest corpus defaults almost entirely to YES. Direction here comes from the thesis, through a declared and auditable partition:

Contest: Which lab holds the leading frontier model

Source: content/indices/specs/open-weight-ai-lead.json — desk-declared partition (Designer §3.3). Declared, never derived: the licence status of each lab's frontier model at authoring time.

Outcome entity	Partition	Match patterns
Anthropic	against	anthropic, claude
Google DeepMind	against	deepmind, gemini, google
OpenAI	against	chatgpt, gpt-, openai
xAI	against	grok, xai
Alibaba (Qwen)	for	alibaba, qwen
Allen Institute (OLMo)	for	ai2, allen institute, olmo
DeepSeek	for	deepseek
Meta (Llama)	for	llama, meta
Mistral	for	mistral

An outcome the partition does not name is UNALIGNED. It stays in the contest's denominator, so the listed mass a reader sees is the real one, and it is never silently counted as support. A contest family the thesis takes no side on is excluded from the oriented gauge with the counted reason no-thesis-side — never defaulted to YES — and remains admitted to side-symmetric gauges, where no side is needed at all.

No leg is ever flipped to improve a YES/NO ratio. A mixed book is a consequence of a thesis taking a real side; it is never a target.

The dead-zone rule. A member priced at or beyond 0.1 / 0.9 on the held side carries almost no information, whichever way it is priced. Dead-zone membership is invariant under the side choice — it is a property of the question, not of the holding — and a composition more than 50% dead on one side is refused at cultivation as one-sided-and-dead rather than published as a reading of a settled board.

No per-member side table is published yet, because this index has no settled history: no close row has been published for it yet, so every worked number below is stated as unavailable rather than invented. The section fills itself from the first published close and from nothing else.

Axis coverage is not yet measured for this index. It is published per close as the share of admitted members the elected axis could read; an unmeasured share is reported as unmeasured rather than as full.

---

7. Family collapse

Time-space collapse: OFF. Each dated leg is its own constituent. That is a declaration rather than a default: with the collapse off, two dated legs of one question are two members, they carry two weights, and a reader comparing this index with one that collapses is comparing different rosters over the same board.

Rank-space collapse: ON. The mutually exclusive outcome legs of one contest are grouped into a family through the event graph's exclusivity structure, and the family — not its legs — becomes the constituent:

x_F = Σ_{k ∈ F⁺} q_k / Σ_{k ∈ F} q_k          (the aligned mass; F⁺ is the thesis's "for" set)

• Vig-strip by normalization. The overround cancels in the ratio: no clipping, no bias. The unnormalized clipped-sum form is excluded — it is vig-biased upward, incomplete-listing-biased downward, and the clip masks the overround that would have revealed both.
• Fréchet-legal by construction. Outcomes inside a family are mutually exclusive, so summing them is additivity. There is no independence assumption anywhere in this construction and never a product of the individual probabilities.
• Listed-mass disclosure. Σ_{k ∈ F} q_k is published per family. An incompletely listed ladder shifts x_F, and a reader has to be able to see the denominator. A family whose listed mass falls outside [0.9, 1.1] does not form at all.
• Nested ladders are refused by name, read directly from the comparator and strike rather than from the mass gate. "Above $110,000" at 0.62 and "above $120,000" at 0.31 sum to 0.93 — inside the gate — and are not a partition: the first strictly contains the second, and the true mass is 0.62.
• One voice per contest. A twenty-leg ladder contributes one x_F rather than twenty correlated near-duplicates. Per-leg YES/NO holdings inside a mutually exclusive family are excluded for directional gauges: six discordant legs at q ≈ 0.13 would each read x ≈ 0.87 and drag the gauge toward 54 while the thesis is priced at 20.
• Family weight is b_F = Σ_{k ∈ F} b_k, the sum of the legs' two-sided depth bases, exactly as time families weight.
• Side-symmetric routing. For a selfInverse gauge the family contributes normalized categorical entropy, H_F = −Σ_k q̂_k log₂ q̂_k / log₂ |F| with q̂_k = q_k / Σ_F q, needing no side at all — so unaligned outcomes still carry product value instead of being discarded.

Grouping version contest-family/v1.

---

8. Weighting

The basis.

b_j = Σ_v min(depth^bid_jv, depth^ask_jv)

Two-sided resting depth near touch — the lesser of the bid-side and ask-side visible resting size within the published band of the best quote on that side, summed over venues. Side-symmetric by construction, so an index holding a mix of YES and NO members weights them on one comparable basis. Unit: claim units of resting depth near touch — the lesser of the bid-side and ask-side visible resting size within the published band of the best quote on that side, summed over venues. One unit pays $1 on YES on every venue indexed. Never a dollar figure, and never an execution claim: W.E.T. handles no order fulfilment. Basis version two-sided-resting-depth/v1.

Cap and redistribute, iterated to a fixpoint. No single member may exceed c = 20.00%; the excess above the cap is redistributed across the uncapped members and the procedure repeats until no member breaches it. Weights are then normalized, wⱼ = w̃ⱼ / Σ_k w̃_k.

Where the cap is mathematically unreachable — the engine's own test is c·n < 1 + 1e-12, tolerance included, which on this cap means a roster of fewer than 6 constituents — no valid capped vector exists and the weights fall back to equal, the least-concentrated valid answer, with the fallback recorded on the row.

Coverage. κ_t = Σ_{j priced} w_jt — the share of the published weight vector that actually carried a price today. This index's floor is κ_min = 95.00% and below it the index prints nothing at all (§11). κ is measured in the space the gauge publishes: on the weight vector wⱼ, because a coverage figure taken over a vector nobody consumes describes a number that was not published.

Concentration is published, not summarised away. HHI is Σ wⱼ² over the published vector and effective N is 1 / HHI: the count of equally-weighted members this weighting is worth. A roster of ninety names with an effective N of thirty is a thirty-name index that lists ninety, and the gap between those two counts is the thing a cap exists to bound.

No weighting report is published for this index yet. HHI, effective N and the concentration shares are read from a close row, never written beside one, and this index has no settled history to read.

---

9. Gauge

Declared family: mean. The published formula, stated over the oriented xⱼ:

G = 100 · Σ wⱼ xⱼ

A weighted mean of the oriented values. It reports how much of the worldview the board is pricing — not how dramatic the watchlist happens to be.

Self-inverse: no — a property of the formula, not a choice, and not authorable in the spec. Flipping every side maps G to 100 − G, so a twin is minted by default (§2).

---

10. Level and continuity

Continuity: chained. Event contracts resolve and vanish, and an index cannot jump when they do. Composition changes adjust the divisor, never the printed number:

V_t = Σ_j u_j m_jt,     L_t = 100 · V_t / D_t,     D_new = D_old · V_new / V_old

At a rebalance or a resolution the prior composition is valued at today's prices, so the divisor absorbs only the composition change and never a market move. The discontinuity at a rechain is 0.000000 by construction — that is what the divisor is for, and it is the property that makes a level comparable across a reconstitution when a spot reading is not.

Inception. The divisor is struck at the first governed close after mint, never from composition-time mids. A composition-time strike prices the basket at whatever the mids happened to be during the minting job — a moment nothing governs and nobody can reproduce — and every level in the series would inherit that arbitrary instant. There is no other inception rule and the grammar cannot express one.

Two numbers, and which one to quote. The gauge is the spot reading, recomputable from one close row's own members. The level is that gauge chained through the divisor, and it needs the divisor's whole history. They answer different questions and neither substitutes for the other.

No level is published yet. The divisor is struck at the first governed close after mint, so a newly minted index has a gauge and no level until that close runs.

---

11. Missing data and quality states

Three measured quantities, all printed on every close row.

κ_t = Σ_{j priced} w_jt                     (weight coverage)
π_t = |{ j : no price }| / |J_t|            (missing-price COUNT share)
χ_t = |{ j : no two-sided basis }| / |J_t|  (missing-basis count share)

κ and π are both required because they fail independently. κ catches the heavy members vanished; π catches the long tail vanished — a breadth failure that weight coverage hides completely while the top members still price.

The ladder, evaluated top-down, first match wins.

#	Condition	State	Prints?
1	administrator action	paused	no
2	κ_t < 95.00% (this index's floor) or π_t > 25.00% or fewer than 4 priced members	insufficient_data	no — never a carried value
3	newest member mark older than 48 h	stale	yes, badged — and no directional statement
4	π_t > 15.00% or χ_t > 50.00%	degraded	yes, badged
5	κ_t < 98.00% or π_t > 5.00% or χ_t > 25.00%	limited	yes, badged
6	otherwise	current	yes

Why a ladder and not a cliff. A single abort threshold makes a 5% day and a 19.9% day identical on the page, and then makes a 20.1% day identical to a total outage. That is the wrong shape for a number published every day, so the floor stays hard and everything above it is graded.

Rules that hold at every rung.

• Per-member absence sets availability A_jt = 0 and renormalizes the weight over the priced members, always disclosed (wet-logic §6.4). Renormalization is never silent: every member's full weight, used weight and availability are published, so a reader can re-derive the renormalization from the row.
• A missing depth basis degrades but never blocks. A member with a price and no two-sided book falls back to the observed mean basis within its cohort, disclosed. Losing the weight basis is a precision loss; losing the price is a truth loss; the ladder treats them differently on purpose, which is why χ never appears in the no-print rung.
• Q^idx_t = Σ_j w_jt Q_jt is published BESIDE the value, never inside it. A quality score folded into a level produces one number that answers neither question.
• The state is a published field on the close row, not a rendering decision. A reader consuming the JSON sees exactly what a reader of the page sees.
• π and χ thresholds are fleet-wide constants and change only under the methodology-change process; κ_min is per-index and is 95.00% here.

No close row exists yet, so no measured κ, π or χ is shown. The thresholds above are in force from the first close.

---

12. Sentinel and quarantine

The screen runs before the close, and the close refuses to run without it. Version sentinel/v1. The rules:

• volume-vs-move
• book-depth-collapse
• divergence-spike
• stale-mark-as-move

Threshold	In force
large move	≥ 0.15 in probability
undefended book	spread ≥ 0.1 AND ≥ 3× the prior spread
prior-spread floor	0.005 — so a one-tick book does not make every widening infinite
depth collapse	depth_now / depth_prior ≤ 0.25
divergence	≥ 0.2 from the other venues' median on the same canonical event
divergence jump	≥ 0.15 increase in that detachment since the prior snapshot
stale mark	3 consecutive identical observations

Quarantine is exclude-and-disclose. A quarantined input is dropped from the number and counted on the close row by rule. It is never silently dropped, and it is never left in with a caveat attached.

The screen observes anomalies and never asserts intent. No arrangement of a price tape can evidence a state of mind, so no output of this screen names a cause, and the vocabulary that would assert one is refused by the publishing layer rather than avoided by convention.

A rule that could not be evaluated is counted as unassessed, not as clear. An input nobody could check is not an input that passed, and a screen reporting only its hits would be claiming full coverage while running at a fraction of it.

No sentinel counts are published for this index yet. They arrive on the first close row and are read from it.

---

13. Governance and lifecycle

Class: experimental. Lifecycle state: experimental.

The standing class sits between the reserved top class — created only by human ratification under a formal consultation, and carrying its own reserved trust vocabulary that no document on this page may borrow — and the weekly slates. Its levels and its documents are public; its rows carry a spec digest and a methodology hash as flat stamps.

Published specs are immutable. "Never change a rule to produce a number" is structural here rather than aspirational: a published standing rule cannot change at all. Evolution is by succession — mint a successor spec, link it, retire the predecessor with notice. During the 28-day probation window an experimental spec may be re-minted in place with a version bump and a logged diff of every canonical field that moved; graduation freezes it.

Errata are not changes. A transcription error — a typo'd threshold that never matched the ratified proposal — is corrected under the restatement policy's existing input-error language: appended, disclosed, never silently edited. A methodology change is never an erratum.

Restatement and cessation. A published value is final unless it meets one of the narrow conditions in the restatement policy; corrections are appended, never edited in place. Cessation follows the cessation policy: 20 uncomputable days trigger a notice and a 30-day period, after which the series retires, its twin retires with it, and its history stays published permanently. See also methodology change and consultation, conflicts, data errors and complaints.

Hourly output, where it exists, is indicative and is labelled so on every surface. It carries no sentinel screen and it is never the governed print. Any surface showing an indicative value beside a governed one badges which is which.

Succession. None declared. This spec supersedes nothing and is superseded by nothing today.

The version block, stamped on this document and on every close row.

Axis	Version
compiler	methodology-compiler/v1
spec schema	worldview-spec/v1
spec version	v1
methodology hash recipe	methodology-hash/v1
engine	wet-engine/v1
gauge interpreter	spec-compute/v1
weight basis	two-sided-resting-depth/v1
consolidation policy	consolidation/v1
eligibility	eligibility/v1
quality	quality/v1
sentinel	sentinel/v1
entity resolution	literal/v1
severity rubric	— (this family reads no severity)

---

14. Worked example and limitations

This index has no settled history: no close row has been published for it yet, so every worked number below is stated as unavailable rather than invented. The section fills itself from the first published close and from nothing else.

Limitations, generated from this spec rather than curated.

• Single-venue exposure is not yet measured. The share of members priced by one venue only is published per close, and this index has none to read.

• The alignment partition is a declared judgement. It is auditable — the entities, the patterns and the source are printed in §6 — and it is still a judgement about which outcomes support the thesis. A reader who disagrees with the partition disagrees with the index, and that is the honest place for the disagreement to sit.

• Indicative output is not governed output. Any hourly value carries no sentinel screen and is labelled indicative wherever it appears.

• L + L̄ is not a constant. The twins are complementary in gauge space only; their levels chain independently and no statement here implies otherwise.

---

15. Disclaimer and reproducibility

Informational only. This document and the values it describes are data about what prediction markets are pricing. They are not advice, not a recommendation, and not a forecast issued by this desk.

No order routing, no custody, no execution. W.E.T. operates no venue and handles no order fulfilment. Any percentage quoted anywhere as an illustration is a hypothetical equal-notional percentage over the members described here, with no fees, no slippage and no fill assumption; it is arithmetic on published prices and it is not a return anyone received.

Not the probability that the thesis holds. Neither the gauge nor any level derived from it is the probability that the thesis is true or false. It is a weighted reading of separate markets pricing separate questions.

Multi-event statements use Fréchet–Hoeffding bounds, and nothing else. For any claim about several of these events occurring together, the only claim this desk makes is:

max( 0, Σ_{i=1..n} p_i − (n − 1) )  ≤  P(E_1 ∩ ... ∩ E_n)  ≤  min_i p_i

Never a product of the individual probabilities, and never a directional joint claim. The events on one board are correlated by construction — they are drawn from one world — so a product understates a conjunction whenever they move together, which is most of the time and always in the direction that flatters the thesis. The compiler's prose guard refuses a document that states one.

Reproducibility.

Stamp	Value
Spec digest	8f267db96ac28e41
Methodology hash	8a45f3768b0d5a00
Hash recipe	methodology-hash/v1
Compiler	methodology-compiler/v1
Consolidation stamp	consolidation/prob-v1

Both stamps are FNV-1a over a canonical serialization — two 32-bit lanes over the same bytes, concatenated to 16 hexadecimal characters, dependency-free so a third party can reproduce them without this code. The spec digest is taken over the axes that decide a number; the methodology hash is taken over those axes plus the thesis prose, the class and the twin link, because the thesis is what the number means. Both are carried on every close row, and the close job recompiles this document from the spec it executed and refuses to print on any mismatch — binding the document to the ledger rather than only to CI.