Architecture trajectory

The project is converging on a dual-speed governance system, not a single model trick.

The near-term direction is clear now. The flat path handles fast lookup and cheap routing. The curved path handles anomaly separation, drift, and governance. The existing 14-layer runtime stays intact. The work is to harden the front door, route ambiguity correctly, and connect the research lanes to the production lanes without pretending every benchmark promotion is already earned.

Open verified results Math flow Benchmark code

What exists now

1. Runtime core

The repo already has a large harmonic, governance, crypto, orchestration, and API surface. The core issue is not “missing architecture.” It is uneven coupling between the strongest theory lanes and the practical front-door lane.

2. Current best gate

`E4` is still the best clean detector on the current benchmark surface: high detection, zero holdout false positives, and stable reruns. That makes it the current anchor rather than an experimental side branch.

3. Geometric edge

The helix result is repeatable and useful, but it is a representation win, not a direct replacement for the detector. It belongs under the gate as a separation engine, not as a headline retrieval system.

4. Null-space role

Null-space is now correctly scoped: it is a boost feature for missed attacks, but not a universal gate. It belongs in the uncertain zone, where extra scrutiny is worth the cost.

The architecture that is emerging

The project is trending toward a hybrid runtime with three speeds: a cheap fast path, a governed curved path, and a quarantine analysis path. That lets flat-space speed and hyperbolic safety coexist instead of being forced into the same optimization target.

Fast path Flat / direct / cheap routing for obvious cases. Use when confidence is high and disagreement is low.

Governed curved path Helix or hyperbolic representation feeding the existing 14-layer runtime for separation, drift, and structured anomaly handling.

Uncertain zone Remainder, null-space, spectral, and route disagreement used only when the primary gate cannot resolve the case cleanly.

Promotion path Only features that improve end-to-end detection without unacceptable false positives move from research into the live runtime.

input
  -> flat pass / cheap semantic route
      -> if confidence high and remainder low: allow fast path
      -> else: hyperbolic / helix projection
          -> 14-layer runtime (L4-L14 unchanged)
              -> if clear: allow / deny
              -> if uncertain: null-space + remainder + spectral quarantine lane
                  -> decision record / audit packet / escalation

Project phases from here

Phase 1

Stabilize what is already winning

Keep `E4` as the main benchmark winner.
Use helix as a geometry lane, not as a replacement detector.
Keep null-space in the uncertain zone only.
Publish verification and reproducibility surfaces.

Phase 2

Reconnect the real tongue machinery

Replace the stub front door with the actual tongue/conlang reader.
Feed the resulting tongue state into the existing downstream layers.
Run ablations against the same attack and clean sets.
Only promote if the end-to-end curve actually improves.

Phase 3

Fuse research and product

Public research domain for proofs, reports, and benchmark pages.
Operator runtime for APIs, HYDRA orchestration, and production controls.
A clean product layer that presents capabilities without overclaiming the math.
A repeatable promotion gate from experiment to runtime.

Subsystem map for the direction of travel

Language plane

Moves from heuristic tongue proxies toward actual symbolic tongue reading. This is the front-door repair lane.

Geometry plane

Owns the helix, Poincaré ball, clustering, radius, and separation machinery. Its job is structure and anomaly geometry, not cheap lookup.

Governance plane

Preserves the 14-layer runtime, gate logic, audit decisions, and escalation states. This is where promotion standards stay strict.

Diagnostic plane

Null-space, remainder, flux, and spectral analysis stay here until they prove they improve net system behavior at an acceptable FP cost.

Operator plane

HYDRA, terminal workflows, browsers, APIs, and deployed services. This is where the system becomes a usable product rather than just a benchmark stack.

Research plane

Verification pages, ablations, codebase maps, and benchmark artifacts live here. This is the evidence layer that keeps the project honest.

Decisions already made by the data

Flat space is not the enemy. It is the speed lane. Helix is not the main detector. It is the structured separation lane. Null-space is not a universal gate. It is a conditional boost feature. The project is going toward a hybrid governed architecture where each mathematical surface is used for the job it actually wins.

The thing to avoid now is more scope collapse. Do not downgrade the real tongue system into toy proxies and then talk as if the proxy is the architecture. The trajectory page exists to keep that boundary explicit.