AETERNA v0.9.37

A local non-LLM cognitive architecture with HDC memory, recurrent SNN control, typed execution, and replayable evidence.

Comment

What it does

AETERNA runs locally without an LLM at runtime. It combines structured hyperdimensional representations, recurrent spiking state, persistent goals, bounded program execution, and protected-outcome evaluation.

The Build Week package provides two judge paths:

  1. a one-executable Windows GUI showing a live integrated Cognitive Mission and a deterministic synthetic solar-and-battery control scenario;
  2. a source-verifiable frozen GatePack with artifact hashes, fixed seeds, deterministic replay, and a machine-readable verdict.

Why it matters

Adaptive systems are easy to overclaim when thresholds, seeds, controls, or training data can move after results are known. AETERNA makes the claim path part of the architecture: contract, primary, rerun, independent gate, and ledger. Negative results remain visible instead of being rewritten as success.

Existing foundation vs Build Week extension

AETERNA was an existing research project. Before the event window opened it already contained the core HDC/SNN runtime, persistent-goal machinery, typed Micro-VM, reward firewall, frozen v0.7 GatePack, and H1.0-P certificate. This submission claims only the meaningful extension built with Codex and GPT-5.6 during July 13–21, 2026:

  • 5f8b4b1 (July 13, 23:47 CEST): frozen T0.4h transfer-audit, scorer, provenance, and public-only environment infrastructure;
  • d75e870 (July 14): T0.4h wire implementation, with its later invalid outcome retained rather than rewritten;
  • a064771, 957070e, and bd643db (July 15): complete-cycle preflight and behavior-neutral predictive-schema qualification path;
  • 14db8b8, 8ac778b, and a5770eb (July 16): lexical HDC/FHRR grounding, structurally nondegenerate public projections, and action-gated phase trace;
  • 8d5717d (July 17): phase-only predictive-readability certificate;
  • f73ce789 and f898cc4 (July 18): frozen contract and exact D-H12 V3 certificate;
  • 051d8f7 (July 19): Build Week verification/documentation workflow;
  • a84b7c6 through 582413b (July 21): no-build Windows package, one-click GUI, Cognitive Mission, and synthetic solar integration experience.

The dated Git history and primary Codex /feedback Session ID 019f48ce-4fe6-7612-a0e3-7b00e883fd01 are the event-period evidence. That long-running root task began on July 9 and continued through July 21; only its post-start commits are attributed to Build Week. The entry does not claim that the entire AETERNA codebase was created during the event.

Evidence

  • Persistent-goal live blind certificate with one locked recurrent carrier: 765/788 spatial, 778/788 physical-state, and 788/788 discourse-state actions correct. Reset, cue-flip, and current-only controls collapse the relevant contrasts.
  • D-H12 V3 prospective causal-boundary certificate: 12/12 co-primary endpoints and 12/12 modality-by-stretch coverage cells passed; primary and rerun are byte-identical.
  • Phase-only predictive readability: 91/96 RAW-ALIAS and 253/256 RAW-CLEAN replicas passed complete conditional tables; each of four causal controls produced 0 complete tables in both strata.
  • Frozen v0.7 GatePack: twelve deterministic episodes, artifact verification, replay fingerprint, and passed=true in BlindFrozen mode.

These are separate narrow certificates. Their presentation in one GUI does not constitute a new joint certification.

How to run

No-build Windows showcase

  1. Download and extract dist/AETERNA_v0.9.37_BuildWeek_Windows_x64.zip.
  2. Double-click AETERNA_Showcase.exe.
  3. Run Cognitive Mission and Practical Solar Edge Experiment.

Source verification

git clone https://github.com/kelbas2007/aeterna-nollm-build-week-judges.git
cd aeterna-nollm-build-week-judges
.\scripts\verify_build_week.ps1

The repository is private and must first be shared with the two competition testing accounts listed in the official rules.

Architecture

observation -> HDC/FHRR memory -> recurrent SNN temporal state
            -> persistent goal / bounded model -> typed Micro-VM action
            -> protected outcome -> frozen primary/rerun/gate evidence

Human and Codex/GPT-5.6 responsibilities

GPT-5.6 was used through Codex from July 13–21 as the primary engineering and verification partner. Codex inspected the repository, implemented bounded changes under direction, proposed causal controls, ran falsification and exact rerun workflows, checked provenance/privacy, and packaged the judge experience.

The human owner selected the research objective, chose which hypotheses to test, approved thresholds before protected outcomes, rejected invalid or over-broad interpretations, selected the demonstration scope and license, and authorized all final external actions. Codex/GPT-5.6 did not decide what scientific result should count as success. The AETERNA runtime contains no LLM dependency.

Limitations

  • AETERNA is a research prototype, not AGI or a production agent.
  • D-H12 V3 is certified on one frozen synthetic pulse process; it does not establish the normalized lag law, language grounding, cold-start autonomy, or open-world transfer.
  • The solar microgrid is a deterministic constructed simulation, not a physics-validated digital twin, real sensor deployment, or safety certification.
  • The GUI composes live mechanisms and separately frozen evidence; the composition itself is not a new scientific endpoint.
  • The Windows executable is an unsigned competition build.

Links

The final upload is Public, under three minutes, visibly shows the GUI working, and includes audio explaining both the project and the Codex/GPT-5.6 contribution.

Primary /feedback Codex Session ID: 019f48ce-4fe6-7612-a0e3-7b00e883fd01.

License

The current revision is proprietary. The AETERNA Proprietary Evaluation License v1.0 authorizes OpenAI Build Week and Devpost judges to download, compile, execute, and inspect the project solely for competition evaluation. It is not an open-source license. Earlier Apache-licensed revisions retain their original terms.

Built With

  • codex
  • gpt-5.6
  • hdc-/-fhrr
  • izhikevich-spiking-neural-networks
  • powershell
  • rust
  • sha-256-artifact-verification
  • typed-micro-vm
Share this project:

Updates