> ## Documentation Index
> Fetch the complete documentation index at: https://docs.beliefsystems.xyz/llms.txt
> Use this file to discover all available pages before exploring further.

# Market Grades Methodology

> How Belief Systems assigns a letter grade to an individual prediction market for the quality of its liquidity and the integrity of its order flow.

**Family:** Market Grades by Belief Systems
**Methodology Version:** `mg-1.2` · **Calibration Version:** `mg-cal-1.0`
**Last Updated:** October 2026

> [!NOTE]
>
> Market Grades are published as a beta. The validation work and its published schedule are set out
> under [Beta status and validation roadmap](#beta-status-and-validation-roadmap).

> [!NOTE]
>
> This page is the methodology of record for the Market Grades family. Every grade is a function of
> three things: the archived market data at the time of computation, the methodology version, and
> the calibration version. Each published grade carries both version stamps, so a grade is always
> read against the rules that produced it.

## What a Market Grade is

A Market Grade is a letter grade, from A+ to F, assigned to one prediction market on one venue. It measures **institutional usability**: how far a professional can rely on the market's price as a signal and transact against it without hidden microstructure risk.

The grade measures the conditions under which a probability is observed, never the probability itself. It carries no information about whether an outcome is likely, cheap, or mispriced. It grades markets, never venues, and it is not a forecast, a solvency opinion, an audit, or investment advice.

Three lines accompany every surface that shows a grade:

* "Market Grades measure market quality — liquidity, integrity — not the likelihood of any outcome."
* "Market Grades are unsolicited: no graded party requests, pays for, or previews a grade. Belief Systems holds no positions in graded markets."
* "A grade is not a recommendation to buy, sell, or hold any position, and is not a substitute for a user's own assessment."

### The grading unit

The unit is a single binary market: one question, two complementary outcomes, on one venue. Constituents of a multi-outcome event are graded individually; coherence across the whole outcome set is outside the scope of this version.

A grade belongs to a **grade episode**. If a market's question is edited in a way that changes its meaning, or the venue reuses or migrates the market, the episode closes with its record frozen and a new episode opens and seasons from the beginning. A different question is a different market. Actions, corrections, and citations all key on the episode.

### Letters of record and component grades

Only the grade of record carries a letter with rating-action discipline. There are no provisional letters and no letter-shaped indicators.

The two pillar results, Liquidity and Integrity, are also published as letters on the same scale. These **component grades** are recomputed with every daily computation and carry no hysteresis, no rating actions, no Watch, no caps, and no notching. They are withheld whenever the grade of record is withheld. Only the grade of record moves by rating action, and a component grade is never cited as a Market Grade. Numeric pillar scores are not published.

### Independence

Grading is user-pays, never venue-pays. No graded party requests, pays for, or previews a grade, and there is no pre-publication response window. The graded universe is defined by published objective criteria; membership of any Belief index plays no part in it. No commercial consideration enters any grade determination, administrator act, or methodology proposal.

## Inputs and coverage

### What is observed

Every grade is computed from the market-data archive, never from the venue live. Three inputs are archived per market:

| Input | What it contributes |
| - | - |
| Order books | The full price ladder, top 20 levels per side, with tick size and minimum order size. The basis of every depth, cost, and book-state measure. |
| Trades | Every executed print with notional, price, size, taker side, and taker wallet. The evidence that displayed liquidity is real. |
| Lifecycle metadata | End date, trading status, resolution status, question text history, and tick regime. Drives the lifecycle states below. |

Books for graded and seasoning markets are sampled on a fixed schedule whose interval is set by the market's capture tier. **Sample times are randomized within each interval**, and the schedule is never published. A market that displays liquidity only when it expects to be observed cannot time the observer: presence at a given share of sample times requires presence at about that share of wall-clock time. The capture cadence is a governed quantity: changing it for graded markets is a noticed methodology change.

On the venue graded today, the YES and NO books are one order pool presented twice: a NO bid at `q` is the same resting order as a YES ask at `1 − q`. Every book measure is therefore computed on a single ladder and valued in both directions, so that **renaming the two outcomes can never change a grade**. Every metric below declares its unit and satisfies this property.

### Missing data is never zero

"We could not look" and "there was nothing there" are different facts, and the methodology keeps them apart. A failed fetch is excluded from the computation and counts against coverage. A book the venue reports as empty or one-sided while the market is open enters each metric at its worst honest value. Missing observations are never filled in.

### Observed days and seasoning

An **observed day** is a day with usable coverage on which the market either showed a two-sided book for at least 20% of the day's observed time or printed at least one trade. Days with a dead book season nothing.

A grade of record requires at least **30 observed days**. To enter the graded universe at all, a market's end date must sit at least **65 days** after it was first observed: 30 days of seasoning, 21 days of minimum meaningful graded life, and the 14-day terminal grace described below. This is an entry test only. Once a market holds a grade of record, end-date changes are handled by the lifecycle rules, never by re-testing eligibility. A market with no stated end date passes the test and carries the U flag (see [Resolution-risk notching](#resolution-risk-notching)).

### Coverage

Book coverage is the usable observed time divided by the time the capture schedule expected. Trade coverage requires a continuous trade record, with any gap closed by backfill.

| Book coverage over the window | Treatment |
| - | - |
| 0.80 or above | Full computation. |
| 0.50 to 0.80 | The last full grade carries forward, marked data-limited. No rating action confirms. After 14 consecutive days the market moves to NR-Data. |
| Below 0.50 for 7 days | NR-Data: the last grade of record stays displayed beside the suspension marker. |

Recovery requires coverage of 0.80 or above for 7 days. After an outage of 30 days or less, the grade resumes with a fresh confirmation clock; after a longer one, the market seasons again.

If a single input is unavailable, its pillar renormalizes over the remaining metrics and the absence is disclosed. If more than 40% of a pillar's weight is unavailable, the grade is data-limited. Losing the trade record also forces evidence tier T0.

Every grade carries a **coverage class**: full (coverage of 0.95 or above), reduced (0.80 to 0.95), or limited (carry-forward).

### Evidence tiers

Displayed quotes are claims; executed trades are evidence. The highest letter a market can reach is capped by how much trading corroborates it. The tier input is **TN\_p25**, the 25th percentile of daily traded notional over the 30-day window, in dollars per day. A 25th percentile at a level requires roughly 23 of 30 days at or above it, so the tier rewards sustained activity, not bursts.

| Tier | Requires | Letter cap |
| - | - | - |
| T2 | TN\_p25 of \$500 or more | None |
| T1 | TN\_p25 of \$50 or more | B+ |
| T0 | Below T1, or insufficient trade coverage | C+ |

A market keeps its current tier while TN\_p25 stays at or above 0.8 times that tier's entry threshold; otherwise it takes the highest tier whose threshold it meets. Tier caps only ever lower a letter, and every binding cap is disclosed on the grade.

Trade evidence can be fabricated at low cost on any venue. The tier does not treat prints as proof. What it buys is duration and visibility: to hold a higher tier, an actor must sustain wallet-attributed, publicly traceable activity on most days of every rolling window, while simultaneously displaying real near-mid depth under randomized sampling. An actor who does that for months is supplying the liquidity the grade measures.

## Pillars and metrics

A grade is built from two scored pillars: **Liquidity** (weight 0.60) and **Integrity** (weight 0.40). Each metric is mapped to a 0–100 score on a ramp between two absolute anchors. Anchors are fixed values, never percentiles of the graded population, so a letter means the same capability in every period and on every venue.

Window statistics are time-weighted over usable observations in the trailing window `[T − 30 days, T)` in UTC. Metrics where more is better use the 25th percentile; metrics where more is worse use the 75th percentile. Holding a property for 75% of observed time therefore requires holding it for about 75% of wall-clock time, given randomized sampling.

### Liquidity

**L1 – Corroborated deployable depth** (weight 0.60 within the pillar). At each observation, the cash a taker can deploy within 5¢ of the mid in the **worse-served direction**:

```
D = min( Σ asks within 5¢ of mid: p · s ,  Σ bids within 5¢ of mid: (1 − p) · s )
```

The first term is the cost of buying YES; the second is the cost of buying NO, since a YES bid at `p` is NO-entry capacity at `1 − p`. An empty or one-sided book gives `D = 0`. The statistic is the time-weighted 25th percentile of `D`, then bounded by trading evidence:

```
L1 input = min( p25(D),  κ × TN_p25 ),   κ = 25 days
```

Displayed depth counts only up to 25 days of sustained turnover. **Unit:** dollars of deployable taker cash. **Ramp:** 0 at $250 or less, 100 at $50,000 or more, log-linear between.

**L2 – Reference-clip execution cost** (weight 0.40 within the pillar). At each observation, a $1,000 reference order is walked through the ladder in both directions: buying YES through the asks and buying NO through the bids. Each direction's cost is its average execution price over the mid, divided by `max(0.05, min(p, 1 − p))` so that cost is measured relative to the price being paid. The venue's taker fee is charged per fill, from the market's own fee regime at the time of the observation, inside each direction. The observation's cost is the larger of the two directions. If the book runs out before $1,000 fills in either direction, the observation scores the worst value, 0.20. A fee regime that cannot be established is a computation error, never a zero fee; observations from before per-market fee archiving began are priced from a pinned venue fee snapshot and labeled imputed.

**Unit:** dimensionless relative cost. **Statistic:** time-weighted 75th percentile. **Ramp:** 100 at 0.01 or less, 0 at 0.20 or more, log-linear between. The clip walk replaces the quoted top-of-book spread, which a minimum-size quote can tighten while real size pays much more. Quoted spread remains published as an indicator.

**L3 – Turnover persistence** (indicator, unweighted). The median of daily traded notional over the window, in dollars per day, scored 0 at $25 or less and 100 at $10,000 or more. L3 is computed and reported but does not enter the grade: it ranks markets almost identically to L1, and it is the more easily fabricated of the two. Turnover still reaches the grade through L1's corroboration bound and the evidence tiers.

### Integrity

**I1 – Participant diversity** (weight 0.45 within the pillar). The effective number of takers over the window:

```
N_eff = min( distinct taker wallets,  1 / HHI of taker notional )
```

**Unit:** effective participants. **Ramp:** 0 at 2 or fewer, 100 at 30 or more, log-linear between. Only takers are observable, so I1 measures the taker side.

**I2 – Book degeneracy** (weight 0.55 within the pillar). The time-weighted share of observations, while the market is open, in which the book is crossed, locked, one-sided, or empty. **Unit:** share of observed time. **Ramp:** 100 at 0.01 or less, 0 at 0.20 or more, linear between.

**Integrity is a screen on flow concentration and book state, not proof of an unmanipulated market.** Many wallets can belong to one actor, and the methodology does not claim to know beneficial ownership. I1 raises the operational cost and forensic footprint of manufacturing diversity; it does not rule it out.

### Diagnostics

The methodology defines an unscored diagnostic battery for administrator review: price moves without trades, self-trade and round-trip signatures, conformance between trades and the book, sweep signatures, and **depth flashing** – displayed depth concentrated around observation times. During the beta, most of this battery is not yet computed; diagnostics never enter a grade, and a diagnostic can become a scored metric only through a noticed methodology change.

### Live indicators

The methodology provides for named numeric indicators between daily computations, with the window in the name, such as 24-hour deployable depth, 24-hour clip cost, 24-hour turnover, 24-hour taker count, and 24-hour quoted spread. They are not published during the beta. When they are, each will carry its as-of time, observation age, count, and coverage, will be withheld once it is more than twice the market's capture interval old, and will never carry a letter.

## Scoring and the letter scale

### Aggregation

Each pillar score is the weighted mean of its available metric scores. The pillars combine as a weighted geometric mean, with a floor so that a zero pillar does not collapse the result:

```
composite = exp( 0.60 · ln max(L, φ)  +  0.40 · ln max(I, φ) ),   φ = 5
```

A geometric mean means strength in one pillar offsets weakness in the other less than an arithmetic mean would.

### Letter cuts

| Letter | Composite at or above |
| - | - |
| A+ | 95 |
| A | 90 |
| A− | 85 |
| B+ | 78 |
| B | 71 |
| B− | 64 |
| C+ | 56 |
| C | 48 |
| C− | 40 |
| D | 30 |
| F | below 30 |

### Caps and the weakest-link rule

The composite letter is then capped. The **weakest-link rule** applies to the lower of the two pillar scores, with the most restrictive match taking effect:

| Lower pillar score | Letter cap |
| - | - |
| Below 20 | D |
| 20 to 39 | C+ |
| 40 or above | None |

The order of operations is fixed: composite → letter → weakest-link cap → evidence-tier cap → resolution-risk notching → published letter. Caps and notches only lower a letter, and every binding cap is published with the grade.

### What the bands mean

| Band | What a professional can do in a market graded in it |
| - | - |
| A-range | Institutional size executes with negligible microstructure risk on demonstrated, corroborated evidence: a $50,000 deployable clip within 5¢ of mid, the $1,000 reference clip fills at 1% relative cost or less, turnover corroborates the depth, 30 or more effective participants, and the book is essentially never degenerate. |
| B-range | Professional use with caveats: a $1,000–$10,000 clip executes at a cost a desk would accept (about 5% relative cost or less), depth is corroborated most days, and participation is real but not deep. |
| C-range | Citable with material caveats, not size: the price is a usable signal, but the \$1,000 clip fits only at a cost that forecloses size, or the evidence is uncorroborated. |
| D | Deficiencies dominate: the reference clip does not reliably fit, or the book is degenerate a material share of the time. |
| F | Unusable on the evidence: no deployable cash within 5¢ of mid, or no corroborating prints. |

These descriptions are the intent of each band. How the composite relates to them is set out in [Known limitation: band descriptions are not individually enforced](#known-limitation-band-descriptions-are-not-individually-enforced).

### Resolution-risk notching

After caps, a defined set of flags lowers the letter for resolution risk:

| Flag | Notch | Determination |
| - | - | - |
| D – active resolution dispute | −2 | From the archived resolution status, confirmed by the administrator. Can reach F. |
| C – material rules clarification after listing | −1 | Administrator, on a recorded question change. Floors at D. A change of meaning closes the episode instead. |
| U – no defined resolution date | −1 | Automatic, when the market has no end date. Floors at D. |

Applying or removing a flag is a disclosed administrator act. The administrator's discretion is closed: only the enumerated acts are permitted, each with a written evidence file that a third party can reproduce from archived observations, a 24-hour cooling period except in cases of active harm, and same-day disclosure.

## Lifecycle and Not Rated states

Every grade carries one lifecycle state. Not Rated is a refusal to opine with the reason stated. It is not a grade and does not rank below F. Suspension states never erase an established grade: the last grade of record stays displayed beside the marker.

| State | Meaning |
| - | - |
| NR-New | Fewer than 7 observed days. No grade. |
| NR-Horizon | Failed the entry test: the end date is too near for the market to season and hold a meaningful grade. Applied only before a first grade of record. |
| Provisional | 7 to 29 observed days. Publishes seasoning progress (for example, "Provisional – day 17/30"). No letter. |
| Active | 30 or more observed days. Grade of record, subject to rating actions. |
| Terminal | The market is near its end. The letter freezes and is shown as terminal. |
| NR-Terminal | Entered the terminal window before earning a grade of record. |
| Resolved | The market has resolved. The final record is permanent and immutable. |
| Retired | Archived by the venue, or closed without resolution data for 30 days. The record closes at its last letter, marked retired-unresolved. |
| NR-Data | Our coverage of the market fell below threshold. The last grade stays displayed with the marker. |
| NR-Inactive | No two-sided book for more than 80% of the window and no trades at all. The last grade stays displayed with the marker. |
| NR-Withdrawn | Our own inputs for the market were corrupted. Restored by administrator disposition. |

**Terminal.** A market enters Terminal when its end date is 14 days away or less; when the venue stops accepting orders before resolution; or when the price has sat within two ticks of 0 or 1 for 5 consecutive days with the end date 30 days away or less. The frozen letter is the lower of the letter at onset and the current letter, so Terminal can never revive a higher grade. For the end-date trigger, onset is dated 3 days before entry. A market leaves Terminal only when every entry condition is false: end date more than 21 days away or removed, orders accepted, and, where the price trigger applied, the price at least 0.05 from 0 and 1 for 5 consecutive days. It returns to Active at the frozen letter with a fresh confirmation clock.

The freeze exists so that a downgrade keeps its meaning. Without it, nearly every market would be downgraded at the end of its life as the book thins ahead of resolution. Deterioration that begins well before the terminal window is real and is reported by ordinary rating actions.

**NR-Inactive** requires true absence. A small book that is present is evidence, and the market is graded on it, down to F if warranted. Withdrawing liquidity cannot buy an escape into Not Rated. Restoration requires a full window of observed days.

## Rating actions and Watch

The grade is computed daily at 00:00 UTC. The pipeline above produces a **candidate letter**; rating actions govern when the published letter changes to it. Every action names its driver, and a move of more than one letter is a single action naming both letters.

**Score-driven actions.** A downgrade requires the composite to sit at least **2 points** beyond the letter boundary; an upgrade requires **4 points**. Either must hold for **2 consecutive computations**. Warnings arrive faster than praise.

**Cap-driven actions.** A cap that begins or stops binding follows its own hysteresis (the tier retention rule above, and a 2-point band on the pillar score for the weakest-link rule), with the same 2-computation confirmation. While a cap binds at or below the current letter, no upgrade of any kind can occur.

**Flag-driven actions** take effect on the disclosed administrator act, after its cooling period, and bypass score hysteresis.

**Accelerated downgrade.** A downgrade can be fast-tracked from the trailing 72 hours of data, but only when coverage over those 72 hours is 0.80 or above, so an outage on our side can never trigger one. The triggers are a qualifying two-sided book for less than 20% of the 72 hours, a degenerate book for more than 50% of it, or a composite 15 points or more below the current letter's floor. The market goes on Watch Negative immediately and the downgrade confirms after one further computation. Accelerated downgrades are suppressed within 14 days of the end date, which is exactly the terminal window. Upgrades are never accelerated.

**Watch.** A market is placed on Watch (Negative or Positive) on the first computation that starts any action clock. If the composite returns inside the band, the confirmation count resets but the Watch stays open. A further move in the same direction continues it; a move in the opposite direction replaces it. A Watch unresolved after **14 days** is removed, not confirmed. There is no Watch in Terminal; a Watch open at Terminal entry is closed as superseded by the terminal state. Watch is the only forward-looking marker. There are no Outlooks: a grade computed from a trailing 30-day window has no forward model to support one.

**Operations.** A missed daily computation is computed on recovery against its original date. Confirmation clocks count computations, not calendar days, so an outage extends a clock and never satisfies it. Late actions carry both their effective and published times. One market's computation failure never blocks another's.

## Corrections, versioning, and history

**The record is append-only.** Every computation writes a record, including gated and suppressed ones, with the reason stated. Every rating action and Watch event is kept permanently in the rating-actions record with its driver, the caps in force, the clock start, the evidence cutoff, its effective and published times, and links to any correction.

**Corrections add a revision.** A computation error is corrected by publishing revision `n + 1` for the same date; the original stays exactly as published, and consumers take the highest revision. Corrupted inputs are flagged and recomputed where possible; while a grade cannot be corrected, the market is NR-Withdrawn. Late data that closes a disclosed gap is not a correction. Full recomputation from raw observations is guaranteed for grades up to 60 days old; older corrections operate on the stored evidence record, and the correction says so.

**Resolved records are immutable.** After resolution, any flag is applied as an annotation on the permanent record with its evidence, never as a letter change.

**Versioning is prospective.** Every grade is stamped with its methodology version and calibration version. A change applies from its effective date. Published grades are never restated or recomputed under a later version. Where a change moves letters, the migration is published as a letter-migration table and the re-seated letters are recorded as a disclosed re-baseline at the version boundary, not as rating actions.

**Pre-launch records stay without letters.** Grades computed before the public launch were never published opinions. They remain in the record as dated entries with their letters withheld permanently and are not backdated into the published history.

**Comments on the methodology.** Send comments on the methodology, including a factual error about a grade's inputs or computation, through the [Contact page](https://beliefsystems.xyz/contact). Substantiated errors are corrected as above, and the dispositions of error reports are published quarterly. There is no fee, no preview, and no negotiation of outcomes.

## Beta status and validation roadmap

Market Grades are in beta. Grades publish under methodology mg-1.2 and calibration mg-cal-1.0, and the reviews that test them run in public on the schedule below.

**What a grade measures.** A grade describes a market's quality over a trailing 30-day window. It covers how much cash a trader can deploy near the mid price, what a \$1,000 reference trade costs including the venue's fee, whether trading activity corroborates the displayed depth, how widely trading is spread across participants, and how often the order book is one-sided, crossed or empty.

**What a grade does not claim.** A grade describes the market as measured, not a forecast of how it will trade next. A pre-registered study tests whether grades anticipate order-book quality over the following 60 days. Grades carry no predictive claim until that study reports.

**Schedule**

* **From October 25, 2026:** the second stage of the initial calibration review, covering the rules that govern rating actions and Watch.
* **By December 23, 2026:** the predictive-validation results.
* **January 19, 2027:** the first post-launch review, 90 days after launch. It re-examines every calibration parameter, how often rating actions occur, and the predictive results.
* **Every September:** the annual calibration review.

Each result is published on this page whatever it shows, including no change. A change to how grades are computed is announced in the [Governance Log](/trust/governance-log.md) at least five calendar days before it takes effect. It applies from that date and does not revise published grades.

### Known limitation: band descriptions are not individually enforced

Each letter band describes what a professional can do in a market graded in it. The grade is a composite: strength on one measure can offset weakness on another, limited only by caps for a weak pillar or thin trading evidence. A letter can therefore sit above what a single measure supports:

* an A-range grade where the $1,000 reference trade costs more than 1% of the price, where corroborated deployable depth is below $50,000, or where fewer than 30 effective participants trade
* a B-range grade where the reference trade costs more than about 5% of the price
* a C-range grade where the reference trade does not reliably fit in the book, the D band's condition
* a D or C− grade where there is no deployable cash within 5¢ of the mid and no corroborating trades, the F band's condition

At the September 25, 2026 calibration review, no market met the A range's 1% cost condition. A revision capping each band at its own conditions is under review and would be announced in the Governance Log at least five calendar days before taking effect.

## Review schedule

The methodology and calibration are reviewed on a fixed schedule, and every review is published whatever it finds, including no change. The dated reviews for the beta are listed under [Beta status and validation roadmap](#beta-status-and-validation-roadmap): the first post-launch review falls 90 days after launch and is published regardless of outcome.

Beyond that, calibration is reviewed **every September**. The month is fixed so that a review date cannot be chosen around the results. A review outside that cycle requires governance approval. Every recalibration ships with an impact study, including a letter-migration table, and applies prospectively.

Methodology and calibration changes follow the public notice rules in the [benchmark governance policy](/trust/benchmark-governance.md): at least five calendar days' notice before a change takes effect, recorded in the [Governance Log](/trust/governance-log.md).

## Governance and versions

| Item | Current |
| - | - |
| Methodology version | `mg-1.2` |
| Calibration version | `mg-cal-1.0` |
| Grading window | 30 days |
| Computation time | Daily, 00:00 UTC |

Each published grade carries the methodology and calibration versions it was computed under, together with its episode, lifecycle state, Watch status, data status, binding caps, evidence tier, flags, as-of date and window, coverage class, and the three disclosure lines above. A letter is never published without them.

Methodology changes, corrections, and cessations are recorded in the [Governance Log](/trust/governance-log.md). Ceasing to grade a category, a venue, or the family is a noticed governance act with a published effective date, and the published history remains citable.

## Further reading

* [Benchmark governance](/trust/benchmark-governance.md) – how methodology changes are noticed.
* [Governance log](/trust/governance-log.md) – the dated record of methodology changes, corrections, and cessations.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.