Sema: An Executable Reconstruction of Match Composition
A public laboratory for disjoint proposal batches, temporal fairness, deterministic replay, and honest approximation.
CENTRAL QUESTION
How much of production matchmaking judgment can be made inspectable without pretending to reproduce a proprietary production system?
WORKING THESIS
A useful public reconstruction should encode invariants and counterexamples, expose its approximation budget, and clearly separate demonstrated semantics from product-specific policy.
PUBLICATION BOUNDARY
What this piece can honestly claim.
GROUNDED IN
A public implementation shaped by transferable questions from matching-systems work, with conformance tests, an exhaustive oracle, replay, and explicit bounded search.
DOES NOT CLAIM
Sema is not a copy of an employer system, a production benchmark, evidence of title-specific outcomes, or proof that its default policy is optimal for a real game.
DRAFT ARCHITECTURE
The argument, section by section.
- 01
State the reconstruction boundary first
The project reconstructs a class of allocation problems, not a confidential system or a universal game policy.
EVIDENCEPublish a claim table before discussing architecture or performance.
- 02
Encode proposals as a conflict graph
Once proposals share tickets or backfill slots, selecting one changes the feasible set for the rest.
EVIDENCEVisualize one batch and verify disjointness independently from preference ordering.
- 03
Keep policy ordering legible
Backfill, waiting time, coverage, and match quality are not naturally one scalar objective.
EVIDENCEShow explicit lexicographic or Pareto decisions and the alternatives they reject.
- 04
Use an oracle to test semantics
Optimized or indexed selection is trustworthy only while it agrees with a slower specification on bounded inputs.
EVIDENCECompare the production-shaped path with exhaustive enumeration and differential fuzzing.
- 05
Expose approximation as output
A bounded search that exhausts its budget has produced a qualified answer, not an unqualified optimum.
EVIDENCEInclude the budget, truncated frontier, replay seed, and best-known alternatives in the result.
- 06
Return calibration to the product
The system can protect invariants and explain tradeoffs, but it cannot invent the product's definition of a good match.
EVIDENCEList the parameters that require experiments, domain owners, and real outcome data.
EDITORIAL GATE
Evidence required before publication.
Conformance cases
Small named cases for disjointness, starvation, backfill, freshness, and concurrent reservation.
Oracle comparison
An exhaustive Pareto reference for small inputs and differential checks against bounded strategies.
Replay artifact
A deterministic input and decision trace that makes a surprising batch inspectable.
Claim table
For every conclusion, mark whether it is proven by code, illustrated by a model, or left to product calibration.