Why a formal model, and why this one

Why this is the most advanced court-backlog simulator we know of

The claim rests on what the published predecessors do and do not do, not on adjectives. Each column below is a documented model from the method note that follows; each row is a capability a backlog model needs if behaviour, not headcount, is the lever.

yes   partly   no. Ratings follow each model's own published methodology, cited in the note below. Every predecessor is a stock-and-flow or micro-simulation model that takes the share of cases going to trial as given; the US Commission says so directly. This is, to our knowledge, the first court-backlog simulator in which that share is decided by the participants, the innocent are protected by construction, and the court's queue answers back. If a reader knows of a published model that does all three, we would like to see it.

What that lets a decision-maker do that no predecessor allows

Ask the question the backlog actually poses. A judges-per-case model can say how many judges clear 4.6 crore cases in five years; it cannot say whether a plea discount, a perjury rule and a two-track docket would clear them with the judges India already has, because in that model nobody changes behaviour. Here the Secretary who doubts a 60% perjury-conviction rate drags it to 25% and sees whether the equilibrium holds; the judge who fears coerced pleas reads the coerced-innocent count per 10,000 for every scenario; the Finance Secretary reads judicial hours saved on fresh inflow against a stated range. The model also tells its own limits: a priced plea with free perjury on a nine-year queue produces today's result, and the full package tips only once a fresh-case list brings disposal below roughly four years.

Why a formal model, and why this one

A note on method and precedent, for readers who will be asked to rely on these numbers.

1. What the model is for

India's backlog is usually described as a shortage of judges. The arithmetic behind that description is a stock-and-flow model: cases filed, cases disposed, cases pending. The Law Commission's 245th Report built its "rate of disposal" method on exactly this, dividing disposals by working judges to find how many more judges would break even and clear arrears [1]. The Economic Survey 2018–19 did the same at national scale: with a case clearance rate of 89% in the district courts and 746 disposals per judge a year, it computed that about 8,152 additional judges would clear the backlog in five years, or a 24.5% productivity gain at full sanctioned strength [2]. Both models hold behaviour constant: the share of cases that go to full trial is treated as a fact of nature. This simulator treats it as the main policy variable. Fewer than one Indian criminal case in 900 is disposed of by plea bargaining under Chapter XXIII; the rest end by trial, compounding, withdrawal, or acquittal when witnesses turn; in comparable systems, on the consultation paper's figures, 85–97% of convictions are by plea. If that gap is an equilibrium of incentives, as the Department of Justice's 2023 field study on plea bargaining suggests [3], then a model that cannot move it cannot find the cheapest way out of the backlog.

2. Sentencing commissions already govern by simulation

Modelling the effect of a sentencing rule before adopting it is routine, and in places a legal duty. The United States Sentencing Commission must consider the effect of guideline changes on the federal prison population under 28 U.S.C. §994(g) and 18 U.S.C. §4047, and does so with a case-by-case micro-simulation that recomputes each sentence under the proposed rule [4]. Minnesota built a prison-population projection model in 1979, before its guidelines took effect, and used it to hold the prison population within capacity; the approach was judged relatively successful once the policy targeted the sentencing decision directly and left little discretion [5]. The Sentencing Council for England and Wales must publish a resource assessment with every draft and definitive guideline under s.127 of the Coroners and Justice Act 2009 [6]. Court systems model their dockets the same way: the NSW Bureau of Crime Statistics and Research built a day-by-day simulation of the District Criminal Court, validated it against the observed backlog, and used it to test judges, adjournments and late guilty pleas against a target before the government chose [7]. The Urban Institute's Prison Population Forecaster put a fifteen-state version of this in the public's hands as an interactive tool [8]. The draft National Sentencing Guidelines, 2026 adopt this practice: s.17(4) requires a member of the Sentencing Authority with demonstrable expertise in mathematical or game-theoretic models of sentencing policy, and s.18(1)(a)(v) directs the Authority to apply game-theoretic modelling to plea-bargaining, witness cooperation and pendency. This page is a working prototype of that function.

3. Where behaviour is the lever, mechanism design has a record of being right

The sentencing simulators above share one limitation, stated plainly in the US Commission's own methodology: the model does not account for changes in prosecutorial decision-making [4]. Where a reform works by changing what participants choose, the record belongs to game theory. Cartel leniency is the closest analogue to Schedule C of this blueprint: a conspiracy in which symmetric liability makes silence stable until the first confessor is offered immunity. After the US Department of Justice revised its leniency programme in 1993, cartel discoveries rose by about 60% at once and then fell about 40% below the old level, the signature of better detection followed by deterrence; a structural estimate puts cartel formation about 59% lower and detection about 62% higher [9], and the theory of when such programmes reduce cartels has since been worked out in detail [10]. Spectrum auctions are the other canonical case: formats derived from auction theory were first used by US authorities in 1994, have spread worldwide, and earned the 2020 Nobel Prize in Economic Sciences [11]. The inequality at the centre of this simulator, IC_G, is the expected-utility deterrence condition of Becker (1968) [12] and Allingham and Sandmo (1972) [13], which has guided tax-compliance and enforcement design for fifty years: certainty and speed of a sanction do more than its severity, and a credible, calculable reward for truthful disclosure changes the choice.

4. What this simulator does that the precedents do not

It joins the two traditions. The Guidelines' quintile engine (Base Score 5, Rules 5–9, Rule 13) prices the plea the way a sentencing commission's worksheet would. A behavioural layer then lets four kinds of accused, every material witness and every co-conspirator choose the cheaper strategy under heterogeneous beliefs. The complainant is a player too: how long a victim stays in the case decides who is left to testify, so the conviction probability the accused faces is derived from the wait rather than assumed. A queue closes both loops: pleas free trial hours, the wait shortens, fewer complainants give up, more evidence reaches the court, more plead. That feedback is what the Economic Survey's model lacks, and it is why the simulator finds a tipping point (a break-even trial duration of roughly four years at a 75% conviction probability) rather than a linear judges-per-case answer. It is also why the simulator can show the partial-reform trap: a priced plea with free perjury and a nine-year queue produces the status quo.

5. What a model can and cannot establish

A model of this kind proves internal consistency, identifies which assumptions carry the result, and locates thresholds. It does not forecast. The NSW authors note that their results depend heavily on assumptions and validate the model against a prediction interval before use [7]; this page follows the same discipline. Every constant is listed with its status and source on the Assumptions page; the status-quo scenario is checked against the observed 0.11% plea rate with no target set; the headline is shown as a range across pessimistic and optimistic bounds; and the scenario fingerprint lets two readers confirm they are looking at the same inputs. The right use of the page is the one the US and UK commissions make of theirs: to stress-test a draft rule before it is adopted, and then to replace assumed values with pilot data, beginning with the s.138 cheque cases of Stage 1.