trident 0.2.0 — the soft3 release

One sentence: trident compiles to nox by default with soft3 as its infrastructure — strata algebra and hemera hashes in the compiler, and a new warrior joy (the cyber battlefield, built the way trisha is built) doing execute · prove · verify · deploy — so trident prove produces a zheng proof. 0.2.0 ships only when the whole scope, zheng openings and bbg included, is done. The proof is the meat; no gates, no --experimental.

Supersedes and absorbs: nox-target.md (Phase 1 executed), nox-api-alignment.md (done on branch feat/nox-api-alignment). Both get deleted on sign-off.

Current state (audited 2026-09-07)

The integration is not a graft — the design anticipated it. What's missing is completion, proving, identity — and the removal of every parallel implementation: trident accumulates soft3 as default.

Milestones

M0 · land the alignment (done, needs merge)

PR feat/nox-api-alignment → master. Uncommitted Cargo.toml/lock tidied first.

M1 · full language surface → nox (3–4 sessions)

NoxCompiler Phase 2: bounded for (unroll when bound is static; recursive core otherwise), multi-function calls (battery in subject), structs/arrays/tuples (cons-trees), mutable assignment (subject edit), sponge/Merkle builtins → nox hash pattern + jets. Builtin-sync rule holds across the 4 places (reference/language.md · typecheck · tir/tree · cost). Unit tests per mapping; tests/nox_surface.rs compiles every language feature.

M2 · honest cost model (1–2 sessions)

[cost] tables=["reductions"] becomes real: static reduction/jet counts per builtin, trident build --target nox --cost prints the proving bill. trident bench grows a nox column (unit: reductions, not cycles — no fake equivalence). target.toml flips cost_model=true, tests=true, level 4.

M3 · joy is born — run + verify, end to end (2–3 sessions)

New warrior joy (~/cyber/joy, sibling of trident, built exactly the way trisha is: thin Rust binary, trident via path = "../trident", implements the Runner/Prover/Verifier/Deployer traits over ProgramBundle). Battlefield: cyber — terrain nox. Deps live in joy, not in trident-lang: cyber-nox, zheng, bbg, cyber-hemera, strata-nebu. trident's CLI already delegates run/prove/verify to the resolved warrior — joy slots into the battlefield registry beside trisha. M3 scope: joy run (nox reduce with Tracer) and joy verify by re-execution. Differential harness: every benches/references/ program runs on both Triton (trisha) and nox (joy) — outputs must agree. This is the release's correctness anchor.

M4 · zheng prover — the meat (4–8 sessions, no gate)

joy prove: nox trace → zheng witness → zheng proof; joy verify checks the proof without re-execution. The blocker is the work: axis/hash/look opening derivation in zheng gets solved inside this release — it is the release. Order of attack: wire the pipeline for Layer-1-only programs first (forces the axis/hash openings), then look (meets M6's bbg serde). 0.2.0 does not ship until joy verify accepts a zheng proof for every differential- harness program. No --experimental, no honest-dashes fallback: the dash today IS the release note; 0.2.0 replaces it with the number.

M5 · soft3 replaces field/ — one algebra, one hash (3 sessions)

No parallel implementations survive. trident's field/ is replaced by the stack, not cross-tested against it:

• goldilocks.rs → deleted; Field is strata-nebu (already a dep).
• poseidon2.rs → deleted; hashing is cyber-hemera everywhere — compiler internals, content addressing, the hash builtin, proof params. One Poseidon2 on the planet, and it is hemera's.
• babybear.rs / mersenne31.rs → stay: they are foreign-target properties (risczero, sp1), not stack algebra. Our thing is Goldilocks; strata never learns these fields. The PrimeField trait shrinks to what foreign targets need; the Goldilocks path speaks strata's trait directly.
• migration is bit-exact-or-explicit: any digest that changes value is a breaking change called out in the CHANGELOG, not silently absorbed.

Then identity lands on the real hash: content-addressed code graduates from "hash of normalized AST" to the hemera particle of the normalized AST — trident deploy emits particle + cyberlinks (name→hash is a link, certificate→hash is a link). Memoization comes free: nox ask() computes order_axon = H(formula, object) — deployed trident functions are globally memoizable by construction.

M6 · bbg state — the look pattern opens (2 sessions + small bbg PR)

nox pattern 17 (look) reads bbg with proof, in joy. Needs upstream: serde for QueryProof/Commitment/Opening in bbg/lens (small, unblocks the soft3 SDK too). trident side: os.state.read lowers to look; bundle declares state deps. In-scope for 0.2.0, no slipping — the scope is discrete and completes in this release.

M7 · truth in public (1 session)

• README: nox is the default target — "produces a zheng proof; STARK on Triton via --target triton"; quick start shows the default path; the "minimal dependencies" principle rewritten: soft3-native by default, nothing outside the stack.
• trident.pink hero line gains the same dual-proof truth (only after M3 green).
• soft3 status.md: trident row state, ".tri → .nox" becomes verified fact; zheng row updates per M4 outcome.
• CHANGELOG + GitHub release 0.2.0; conformance note in soft3.

Architecture decision (settled by owner)

The trisha pattern, faithfully: a new warrior — joy — for the cyber battlefield. trident stays the weapon (compile, cost, audit; its algebra and hash become strata + hemera by default, nox the default target so trident build main.tri emits .nox). joy is the warrior: a thin external binary owning execute · prove · verify · deploy on nox, carrying the heavy deps (cyber-nox for execution, zheng, bbg, cyber-hemera). Triton keeps trisha; cyber gets joy; the ProgramBundle boundary stays clean. cargo install trident-lang + cargo install joy is the full kit; the README's "minimal dependencies" principle is rewritten in M7: minimal means no dependency outside the soft3 stack.

Out of scope (0.3.0+)

os/cyber union target (os.signal/os.neuron via mudra+cybergraph), radio/cell integration, foculus, jali/FHE path, nox native compile (wasm/arm64 — separate plan in nox repo).

Verification (per repo rules)

Every milestone: cargo check zero warnings · cargo test · trident bench no regressions · trident audit holds · differential triton×nox agreement (from M3). Branches feat/*, PRs, no direct master commits.

Estimate

Full scope M0–M7, no gates: ~16–22 sessions. The spread is M4 — zheng openings are research-shaped; everything else is engineering-shaped. 0.2.0 ships when the last milestone is green, and not before.

Settled

• Warrior joy for the cyber battlefield, trisha-style — not features in trident. (owner, 2026-09-07)
• 0.2.0 ships only with M4 complete — the proof is the meat, no exceptions. openings + bbg are inside 0.2.0; the scope is discrete. (owner, 2026-09-07)
• Version: 0.2.0. (owner, 2026-09-07)
• babybear/mersenne31: foreign-target fields (risczero, sp1), inert, never enter strata — investigated, nothing wrong. (2026-09-07)