forge

Partial. The runtime delivers this, with a documented gap against what the reference describes. The gap is named in the compiler's own ledger rather than left to be discovered.

Since v1.12.0 (stub); v2.41.0 (v2.41.0, real implementation) · Used inside a declaration

Grammar

forge <Name>(seed: "<text>") -> <Type> {
    mode:        combinatorial | exploratory | transformational   # optional (default exploratory)
    novelty:     <0.0..1.0>                                        # optional (default 0.5)
    depth:       <int ≥ 1>                                         # optional (default 1) — incubation iterations
    branches:    <int ≥ 1>                                         # optional (default 1) — illumination best-of-N
    constraints: <AnchorRef>                                       # optional — verification anchor + coherence floor
}

forge is Directed Creative Synthesis — the primitive that lets an LLM create, not merely interpolate. It compiles to structured IR metadata that the runtime executes as an orchestrated Poincaré-Hadamard four-phase pipeline (the documented structure of mathematical invention — Poincaré 1908, Hadamard 1945, Wallas 1926), under a measured, fail-closed novelty guarantee: a forge returns a value ONLY if it provably clears a novelty floor; a derivative result is never passed off as creative.

This is not a prompt template. Novelty is a measured quantity, not an assertion, and the pipeline fails closed when it isn't met.

Surface

forge is a flow-body block (sibling of step), not a top-level declaration.

anchor GoldenRatio {
    require: aesthetic_harmony
    confidence_floor: 0.70
}

flow CreateVisualConcept(brief: String) -> Visual {
    forge Artwork(seed: "aurora borealis over ancient ruins") -> Visual {
        mode:        transformational
        novelty:     0.85
        constraints: GoldenRatio
        depth:       4
        branches:    7
    }
}

The four phases

Phase	What it does	Temperature
1. Preparation	Expand the seed into its obvious/conventional reading B — the baseline novelty is measured against.	low (0.3)
2. Incubation	depth speculative iterations pushing far past the obvious.	τ_eff = τ_base·(0.5 + 0.5·novelty)
3. Illumination	branches crystallizations (best-of-N); each branch's novelty is measured as ν = NCD(B, branch).	τ_base(mode)
4. Verification	Select the argmax-utility branch and enforce the novelty floor fail-closed.	~0

Fields

mode: — Boden's creativity taxonomy

A closed catalog (axon-T868) from Margaret Boden's The Creative Mind (1990), each mapping to a sampling profile (τ_base, freedom, rule_flexibility):

Mode	τ_base	freedom	rule_flex	Character
combinatorial	0.9	0.8	0.3	novel recombination of known ideas
exploratory (default)	0.7	0.6	0.5	structured navigation of a possibility space
transformational	1.2	1.0	0.9	rule-breaking synthesis, new paradigms

novelty: — the measured floor

A value in [0.0, 1.0] (axon-T869). It (a) blends the incubation temperature (τ_eff = τ_base·(0.5 + 0.5·novelty)) and (b) sets the fail-closed novelty floor the final output must clear.

How novelty is measured (honest mathematics). Kolmogorov complexity K(x) is uncomputable, so novelty cannot be computed exactly. forge uses the Normalized Compression Distance — NCD(x,y) = [C(xy) − min(C(x),C(y))] / max(C(x),C(y)) with a real compressor C — the standard computable approximation of the Normalized Information Distance, a universal metric grounded in Kolmogorov complexity (Li, Chen, Li, Ma, Vitányi, IEEE TIT 2004). Novelty ν(O) = NCD(baseline, O): "how much of the output is NOT already implied by the obvious reading of the seed." Genuine creation ⇔ the output does not compress away given the baseline ⇔ ν ≥ floor.

depth: / branches:

Incubation iterations and illumination branches, each ≥ 1 (axon-T870).

constraints: — the verification anchor

An optional reference to a declared anchor carrying a confidence_floor (axon-T871). The verification phase uses it as the coherence gate.

The fail-closed guarantee

A forge returns its typed value only if the winning synthesis clears the measured novelty floor. Otherwise it fails with a structured error — never a silent, derivative, or empty result:

• forge.novelty_floor_breached — the best branch was too derivative of the obvious reading of the seed (ν < floor).
• forge.no_feasible_branch — no branch satisfied the constraints: anchor.

This is the load-bearing property: creativity you can verify, or a loud failure.

Honest limits

• forge synthesizes a typed concept/specification, not a rendered image/audio (v1). "Novelty" is novelty-relative-to-the-obvious-baseline — a computable proxy, not a claim of absolute unprecedentedness.
• The runtime hard gate in v1 is the measured novelty floor; per-branch live anchor-confidence judging is a follow-up (the anchor is statically validated).

What this primitive is NOT

• Not a prompt with a high temperature. The novelty floor is measured and enforced; a temperature knob asserts nothing and fails open.
• Not a step. A step is one cognitive operation; forge is a multi-phase, verified creative pipeline.
• Not a top-level declaration. It lives inside a flow body.

See also

• axon://primitives/anchor — the constraints: verification predicate.
• axon://primitives/step — single-operation counterpart.
• axon://primitives/flow — the parent of every forge block.