AURORA-M0 fit snapshot — 2026-05-02-pubgrade (publication-grade)
Pickled aurora.m0.M0Config snapshot from the AURORA-M0 consolidation pass on 2026-05-02. This is the AURORA leg for the H1 evaluation — the event-first model whose energy-per-useful-task ratio is compared against the LoRA-Llama-8B baseline (Anthril/lora-llama8b-aurora-baseline-2026-05-02).
This is not a HuggingFace
transformersmodel. It is an AURORA-M0 fit snapshot governed by ADR-0037 (AURORA-M0 fit contract). It is loaded byaurora.m0.M0Configfrom a pickle artefact, not byfrom_pretrained. The HuggingFace Hub is used here as a versioned artefact store, not as an inference surface.
Repository contents
| Path | Role |
|---|---|
aurora_config.pkl |
Pickled aurora.m0.M0Config instance (~507 KB). Loaded by AuroraRunner.prepare() at H1 evaluation time. |
fit-manifest.json |
ADR-0037 fit contract: corpus SHA, microfield activation counts, host id, git HEAD, wall-clock, etc. |
Headline activation (passes the ADR-0037 gate)
"microfield_observation_counts": {
"ActorIntentMicrofield": 1072,
"CausalLinkMicrofield": 11813,
"EntityStateMicrofield": 1166,
"LinguisticSurfaceMicrofield": 1000,
"RiskMicrofield": 1000,
"RoutineMicrofield": 10777,
"TemporalMicrofield": 23224
}
All seven Stage-1 microfield archetypes have ≥ 1 observation. The H1 launcher's ADR-0037 activation gate accepts this fit. Contrast the pre-enrichment 2026-04-30 fit, where six of seven archetypes had zero observations — that fit is permanently ineligible for H1 publication under ADR-0037.
Fit provenance (from fit-manifest.json)
| Field | Value |
|---|---|
fit_kind |
aurora-m0-consolidation |
n_episodes_target |
1 000 |
n_episodes_materialised |
1 000 |
n_episodes_failed |
0 |
n_events_total |
25 426 |
synthetic_events_per_episode |
2 |
episodic_index_size |
1 024 |
gyms |
household_scheduling, inventory_business, code_maintenance, supplier_substitution, apprenticeship_lda |
seed |
0 |
fit_started_at |
2026-05-02 T 10:47:00 Z |
fit_finished_at |
2026-05-02 T 11:22:38 Z |
fit_wallclock_s |
2 136.9 |
energy_consumed_j |
176 099.4 |
host_id |
amd64-john_oconnor |
python_version |
3.14.2 |
git_head_sha |
6df834a6ab669f24f9c1f4094dcf131614edaf3b |
corpus_sha256 |
76b66f7f47312f86ca5607a07fea5fa9a33f21601a0d38c6bbaa488a16bfe0f0 |
enrichment_script_sha256 |
bda253b2759063ab51761074051483dbb80177f09e15269428e7f6df37d36bff |
aurora_config_sha256 |
49fec3a8072dcbaf9693501eda1e0dded8bfd475589be9f3bee421fcf4836b3a |
spec_anchor |
engineering-spec/decisions/0037-aurora-fit-contract.md |
How to use (replication)
git clone https://huggingface.co/Anthril/aurora-m0-fit-2026-05-02-pubgrade
# In an AURORA checkout:
python scripts/run-stage1-h1-evaluation.py \
--aurora-fit aurora-m0-fit-2026-05-02-pubgrade/ \
--lora-adapter <path-to-lora-baseline>
The H1 launcher reads fit-manifest.json, validates every required field per ADR-0037 §Change 3, validates archetype activation per ADR-0037 §Change 4, and refuses to launch if either check fails (exit code 78).
Energy-source publication acceptance
Per ADR-0035, an H1 PASS/FAIL verdict requires notes["energy_source"] >= 1.0 on every per-episode result during evaluation — measured-energy via NVML + psutil. The fit itself records energy_consumed_j = 176 099.4 from the host's energy meter; the evaluation-time energy gate is enforced separately by the comparator. Replication partners on hosts without NVML will see verdict INCONCLUSIVE_SYNTHETIC_ENERGY regardless of the quality of this snapshot.
Limitations and intended use
- Research only. This snapshot is published to enable AURORA Stage-1 H1 evaluation runs.
- Custom format. Loading requires the
auroraPython package at the same git revision (6df834a6). The snapshot is not portable across breaking ADR revisions of the M0 contract.
License
CC-BY-4.0. The snapshot, manifest, and AURORA framework are AURORA-original.