clay-borg/specs/Positioning.md

Positioning — what clay-borg is for, and what it does that others do not

Status: draft. A working answer to "where does this sit, and what is the point of it", written against the field rather than in isolation.

Companion to ../INTENT.md (what we are building), Taxonomy.md (the vocabulary), and Ornamentation.md (the boundary). The systems named below have profiles in ../simulators/, each marking what was checked and what is background recollection.

Revised 2026-08-08 after the survey, and §4 changed. Two of the three tracks were described as more open than they are — see §4.0.

---

1. The field, honestly

There are mature systems here and we are not first.

system	what it optimises for	what it does not do
Ludii (Digital Ludeme Project)	generality of representation. Games as ludemes — composable, human-readable concepts. Fast, broad, a game fits in a QR code	it is not built to argue with the designer about the rules, and its evidence is not auditable by a third party
GDL / General Game Playing (Stanford)	agent generality — an agent that plays a game it has never seen	verbose, slow relative to Ludii, and aimed at AI research rather than design support
Machinations	economy and feedback-loop simulation, diagrammatic and immediate	not rules-complete; a model of flows, not of a played game
Tabletop Simulator / boardgame.io	play, human or networked	no analysis; the rules are a script, not a claim
RL self-play stacks	strength	a strong agent tells you what beats what, not why a rule is undecided

Ludii is the closest relative and the right benchmark. If clay-borg's answer to "why not Ludii" is only "ours is in Rust", there is no answer.

2. What we do differently

Four things, and each is already visible in the repo rather than aspired to.

2.1 The measurement is under suspicion by construction

Every other system in §1 treats its own output as data. This one treats its own output as a claim that has to survive an attempt to break it — GameDesign.md §1 admissibility, the finding register, mutation-proof as a condition for a control, and mandatory adversarial review at tier L.

This is not a posture; it is measured. Three review rounds on one measurement found twelve fatal defects, every one in our instrument and none in the game — including two that inverted a verdict already written down. A tool that reports game-design conclusions without that machinery is reporting its own bugs at the same volume as its results.

2.2 Authority is federated, and the simulator is not the authority

The game's owner rules; we report. A rule we cannot execute becomes a finding carried back with the artifact that produced it — never a default we quietly pick.

Ludii's model is a description someone wrote. Ours is a claim about somebody else's artifact, with digests and tripwires against their source of truth, and a queue of open questions that belong to them. That is a different relationship to the designer than any system above has.

2.3 What is not mechanism is modelled too

Ornamentation.md. Materials, ceremony, typography, timing — declared, with falsifiers, and separated from the rules by a test something can run (does the state hash move?).

Every system in §1 either ignores this or treats it as presentation config. We treat "this does not matter" as a claim that can be wrong, because our own register is a list of times it was.

2.4 A hypothesis is a package, not a patch

A design experiment arrives as a selectable variant — a catalog entry, a machine-readable rules_delta, declared success criteria — and is measured against those criteria, with the results reported including the parts that fail. The variant is in the game state, so a recording says which rules it was played under.

3. What that adds up to

Clay-borg is a design-evidence instrument, not a game engine and not a game-playing AI. Its product is an auditable answer to "what does this rule actually do at the table, and how much should you trust that?"

The second half of that sentence is the differentiator. Anyone can produce the number.

4. Where it has to go

Three tracks, in dependency order. None started.

4.0 What the survey changed

Two tracks were overclaimed and are now smaller. Writing the simulators/ profiles found prior art where this document had assumed open ground:

• Track C is not open ground. Browne published 57 criteria for game quality, and Ai Ai already computes designer-facing measures — drama, lead changes, branching factor, completion, duration — from played games. The track moves from invent a theory of fun to adopt an existing vocabulary, credit it, and locate the gap. The gap is plausibly real (co-operative and hidden-role games), and it is a far smaller and more defensible claim.
• Track B probably adopts rather than builds. OpenSpiel implements CFR, best-response and exploitability over games that are simultaneous-move, imperfect-information and cooperative — all four of GROUND's awkward properties. "Does ATTACK ever pay" is a best-response question, and we have been answering it with two hand-written policies.

This is what a survey is for, and it is worth noting that the cost of not doing it earlier was measurable: we spent three review rounds refining a two-policy sweep for a question that has a standard instrument.

Track A — a second game

GROUND is an example, and nothing has tested that claim. Every abstraction here has exactly one instance, which by our own rule (GameDesign §2) means it may have been invented rather than observed.

The work is to take a popular, well-understood game — one where the dynamics are already documented, so our output can be checked against something — and build a clay-borg definition efficiently enough that the cost is itself the finding. The measurement that matters is how much of the kernel turns out to be GROUND-shaped.

Ludii's ludeme decomposition is the obvious prior art to steal from, and the honest question is whether we need our own description language at all or should consume theirs.

Track B — game theory as the lens on dynamics

§4 of Taxonomy.md: our panels measure dynamics, and game theory is the science of exactly that. Today we sample the strategy space with two hand-written policies and a rank parameter. That is a crude instrument for questions that have real theory behind them:

• dominance — is an action dominated? (F17 asks this about ATTACK and answers it with a sweep)
• equilibrium — what does competent play converge to, and is it the play the designer intended?
• cooperative solution concepts — GROUND is co-operative with a defection mechanic; the core and the Shapley value are the standard tools for "is this coalition stable" and "what is a seat contributing", which is precisely what GR-E04 is groping at
• social dilemmas — DARVO is one; the literature on them is large

The proposed shape is a specialised agent, not a library: something that reads a clay-borg game definition and a panel run, and reports what the theory says about the dynamics — with the same admissibility rules as everything else, because a Nash equilibrium quoted without its assumptions is exactly the wrong-subject error in mathematical dress (Taxonomy.md §2.2).

But the first task is not to build it. It is to answer one question: is exploitability meaningful for a co-operative game with a shared threshold? If yes, Track B is mostly integration with OpenSpiel. If no, that is a real research finding and the reason a specialised agent is needed at all.

Track C — assimilated knowledge about why games work

The furthest out, and the one the maintainer named as the point: make what is known about game dynamics available to designers.

MDA's aesthetics vocabulary, Browne's criteria, the balance literature, documented failure modes (runaway leader, kingmaking, analysis paralysis, dominant strategy, degenerate equilibrium). The value is not a database — it is that a finding from a panel run can be named as an instance of a known pattern, and the designer told what usually follows.

And the gap is now visible. The criteria that exist assume a leader: lead change and drama are defined against who is ahead. SHARED GROUND has no individual leader — Modes.csv gives its tiebreak as "Not applicable" — so the canonical vocabulary does not reach the game we have. Whether these measures extend to co-operative and hidden-role play is the open question, and it is a better Track C than the one this document originally proposed.

The risk is obvious and worth stating now: this track is where a tool starts telling designers what is fun, on the strength of numbers that measure dynamics. §4 of the taxonomy is the guard, and it will be under pressure here.

5. Falsifier

Track A is the falsifier for this whole document. If a second game cannot be defined without reshaping the kernel around it, then "GROUND is an example" was never true, clay-borg is a GROUND engine with ambitions, and §2's differentiators are properties of one bespoke tool rather than of a framework.