Archive NexaMat with final qualified model card and evidence (#1)

- initial commit (833976aaa0f1ec8a3ed0fb6aa7107e827e23ff6c)
- chore: upload NexaMat controller smoke marker (fbb7c241c4c376250126cab870f0506f292f138c)
- Add NexaMat controller Qwen3-14B prod candidate checkpoint (84c21ac6d35f5eb2f1fa6709bc8e6e780a8b0b19)
- upload-fft-pilot-merged-model (0744421935b186de1c322bc1d9ea54bb028b8137)
- Remove stale NexaMat controller folder multimodal/controller/nexa_mat_controller_qwen3_14b_stack_prod_candidate_20260518T221530Z (5f1c2124d08e9ccfc7b1f422f3d1c3a656e05d38)
- Publish NexaMat rollout stack manifests and examples (c445e73fe50576f3a20d932c161107c96852eb5a)
- chore: publish safetensors NexaMat stack artifacts (e0030d0f83bd2e1bc38aa86bc27544b311c0f534)
- chore: publish safetensors NexaMat stack artifacts (fd43acc056fa206aab4043ef7380ba64bf96bd69)
- chore: publish safetensors NexaMat stack artifacts (ff103d8636df5709b78f0abc6298ce6171959d68)
- chore: publish safetensors NexaMat stack artifacts (9726ac3a7b67eba59cb6e8b0bb300540b792f419)
- chore: publish safetensors NexaMat stack artifacts (59f68e0f2d2a98d2154e883202f44c9f865b05ca)
- chore: publish safetensors NexaMat stack artifacts (bb06f88b3ef2911a08a801e2b55e009439fa436a)
- chore: publish safetensors NexaMat stack artifacts (3643d86839fa32c33bc63041fa943f55af00edcd)
- chore: publish safetensors NexaMat stack artifacts (fcbb517f5adba61fa415bdad3b4f1ab4c71feaf8)
- chore: publish safetensors NexaMat stack artifacts (8f73c062989cb2ef4496963009d44f1dfbdeadb2)
- Upload NEXA_MAT_FULL_TECHNICAL_PAPER.pdf (9b7d861c36592f79552e8f27f305732bfc506e34)
- Rename NEXA_MAT_FULL_TECHNICAL_PAPER.pdf to Nexa_Mat_Paper.pdf (04f16f8ef5f5fea7ace9ffdb45ac19cbc2818f55)
- Archive NexaMat with final qualified model card and evidence (6697478c63ac320290c3f88f76eb975b612a9bb4)
- Clarify historical metastability evidence and reliability limits (cd981ebe5acdb8df9a71f325deef6efe0a09306e)
- docs: publish sanitized final evidence and figures (3e751267a309c173748593a582b192ee8050fcbb)
- release: add figures/nexamat_profile_yield_vs_targets.png (dbc5d83ba3240168c0bdebb5c90c568730639343)
- release: add figures/nexamat_profile_geometry_rates.png (be43b1fcd1244a7258488533216d68e4530b7a74)
- docs: add sanitized metastability evidence plot (e6b90979ec7e5f34565e1ad0aa3973505ea066f9)
- release: add decoder/native_v78/model.safetensors (5c9e49ace97935ac0b94e66bc8a23456adabb5f5)
- docs: publish plotted NexaMat technical paper (33127d05cce90cb318845fa70113028bcb150710)
- chore: disable LFS tracking for Nexa_Mat_Paper.pdf (d0204ed8d90a2d3b151e7c396924dd94dc93de99)
- fix: patch PR #1 README and gitattributes only to resolve conflicts (b2c27e51c0e57bb68a571607b4a81be095de1a6c)
- fix: align PR #1 .gitattributes with main to clear conflict (39f4e53010bd7d8c75a5a8f5468743c92bb623bd)

decoder/native_v78/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e19e2c6c98c2ba8eb4b539d91be652128fe7cecb9877ff3b94b9ee62433f55c
+size 2381144

decoder/native_v78/model_manifest.json

@@ -0,0 +1,44 @@
+{
+  "schema_version": "nexamat.model_manifest/v1",
+  "model_name": "NexaMat",
+  "project_status": "archived_on_ice_indefinitely",
+  "artifact_role": "shipped_recovered_native_decoder",
+  "artifacts": {
+    "pytorch_weights": {
+      "path": "decoder/native_v78/model_state_dict.pt",
+      "sha256": "75498a20c6a25f2fa91b2a087960f18cbae803de095dc2358cb1c747f867878e",
+      "size_bytes": 2391887
+    },
+    "safetensors": {
+      "path": "decoder/native_v78/model.safetensors",
+      "sha256": "1e19e2c6c98c2ba8eb4b539d91be652128fe7cecb9877ff3b94b9ee62433f55c",
+      "size_bytes": 2381144
+    }
+  },
+  "checkpoint": {
+    "state_dict_key": "model",
+    "top_level_keys": [
+      "format",
+      "model"
+    ],
+    "load_mode": "weights_only_true",
+    "source_provenance": "derived from a verified immutable internal checkpoint; optimizer and private metadata removed"
+  },
+  "state_dict": {
+    "tensor_count": 41,
+    "parameter_elements": 594340,
+    "dtypes": {
+      "float32": 41
+    },
+    "keys_sha256": "5304978f2bc821906e1762835a2dbda5f780e674ccaf7567dd872ce625044144"
+  },
+  "scientific_claim": {
+    "intended_use": "native inorganic CIF candidate generation followed by external validation",
+    "metastability": "partial historical evidence only; not a verified high-yield production rate",
+    "historical_bounded_msun": "2/64 in one prior test; recurrence not fully verified",
+    "historical_low_hull_examples_eV_per_atom": [
+      0.1786,
+      0.1837
+    ]
+  }
+}

docs/final_closeout/ARCHIVE_CLOSEOUT_2026-07-24.md

@@ -0,0 +1,152 @@
+# NexaMat Final Project Archive Closeout
+
+Date: 2026-07-24  
+Status: Archived and on ice indefinitely; model shipped with qualified claims  
+Canonical project root: `docs/projects/Nexa_Mat`  
+Archive protocol: NexaMass-style immutable bundle, checksum, manifest, and
+minimal local index
+
+## Executive Verdict
+
+NexaMat is a successful recovery and an incomplete stability program.
+
+The original decoder was brittle: atom identity and novelty were strong, but
+native geometry validity was too sparse to support downstream science. The
+alignment campaign recovered a compact 4.061859M-parameter stack into a useful
+native/no-repair inorganic candidate generator. The final non-alloy panel passed
+its profile-specific static gates, and the strongest SSB lane produced 595
+accepted candidates from 4,096 raw samples.
+
+The metastability campaign reached meaningful but incomplete evidence. In one
+prior bounded 64-candidate test, the cohort mean was 0.8750 eV/atom and two
+individual candidates reached 0.1786 and 0.1837 eV/atom. The correct claim is
+that NexaMat demonstrated the ability to produce isolated candidates near 0.18
+eV/atom while typical evaluated cohort behavior remained closer to the 0.8
+eV/atom regime. That was genuine progress, not a complete miss. Reliability was
+not fully verified: the historical test reported 2/64 M.S.U.N. candidates and
+3/64 strict relaxation convergence, but later cohorts did not establish the
+same rate as repeatable. The campaign's practical ambition was at least 25% and
+ultimately 50%. Later architecture work improved representation AUC, force
+magnitude, and external teacher agreement without making metastable yield
+reliable.
+
+The project stopped because the experiment program had run for weeks and the
+remaining reliability gap no longer justified continuous iteration. This is a
+time, cost, and focus decision, not evidence that NexaMat can never improve.
+
+The honest shipping position is therefore:
+
+> NexaMat is a compact, efficient native inorganic proposal model with sparse
+> demonstrated metastable support. It requires external relaxation and
+> stability screening and is not yet a reliable high-yield metastable
+> generator.
+
+## What Was Achieved
+
+- Recovered native no-repair crystal generation from catastrophic geometry
+  starvation.
+- Separated native, projected, scaffolded, ranked, relaxed, and hull-scored
+  accounting.
+- Preserved raw candidate payloads and explicit failure reasons.
+- Generalized the SSB recovery into profile-conditioned non-alloy inorganic
+  generation.
+- Passed the final static suite for SSB, battery, oxide, halide/chalcogenide,
+  semiconductor, and broad-inorganic profiles.
+- Retained 100% novelty in the final six-profile suite.
+- Produced the first bounded low-hull/M.S.U.N. evidence.
+- Built a source-disjoint trajectory, force, stress, relaxation, and hull
+  evidence stack.
+- Established a real stability representation signal: ROC-AUC 0.70845 and
+  minimum balanced accuracy 0.60048.
+- Established that the frozen MACE-OMat teacher was strong on the comparison
+  bridge: force cosine 0.75450 and force MAE 0.05728 eV/A.
+- Built reliable external gating, provenance, W&B, Wasabi, and failure receipts.
+
+## What Remains On Ice
+
+- High-yield relaxation convergence.
+- Repeatable density near the metastable hull threshold.
+- A native controller that turns local physical signals into globally
+  productive coordinate and lattice motion.
+- Alloys/intermetallics under the same general profile contract.
+- An apples-to-apples public MatterGen/SUN benchmark.
+- A matched generative objective trained directly on a broad low-hull endpoint
+  distribution.
+
+## Final Evidence Table
+
+| Surface | Best verified result | Interpretation |
+| --- | ---: | --- |
+| Full compact stack | 4.061859M parameters | Strong efficiency |
+| Trainable decoder | 1.248145M parameters | Small native generator |
+| V90 SSB native accepted | 595/4096, 14.53% | Useful static yield |
+| Final six-profile accepted | 513, 411, 424, 436, 256, 242 | Non-alloy static panel passed |
+| Final six-profile novelty | 100% per profile | Strong novelty |
+| Full 512 stability screen | 58/512 converged; mean hull 1.2330; 0 M.S.U.N. | Failed stability gate |
+| Best bounded stability anchor | 3/64 converged; mean hull 0.8750; 2 M.S.U.N. | Sparse support, not reliable density |
+| Best individual hull rows | 0.1786 and 0.1837 eV/atom | Desired region is reachable |
+| V32 representation | AUC 0.70845; balanced accuracy 0.60048 | Basin signal exists |
+| V32 controller | negative coordinate, lattice, and pair contraction | Controller failed |
+| Frozen foundation | cosine 0.75450; MAE 0.05728 eV/A | Strong local physical teacher |
+| Residual transfer | approximately 0.75 cosine, no gain | Plateau |
+
+## Canonical Artifact Destinations
+
+### Private Archive
+
+The complete internal project archive is preserved in immutable private object
+storage. Its URI is intentionally withheld from the public release.
+
+The prefix contains:
+
+- the complete project documentation bundle;
+- a SHA-256 sidecar;
+- a machine-readable file manifest;
+- the final model publication manifest;
+- archive and upload verification receipts.
+
+Historical run trees remain in private object storage and are not part of the
+public model package.
+
+### Hugging Face
+
+Canonical model hub:
+
+```text
+https://huggingface.co/AethronPhantom/Nexa_Mat2
+```
+
+The model hub contains the shipped encoder/decoder stack, curated research
+lineage artifacts, final documentation, figures, evidence summaries, and an
+explicit limitations section. It is a model and evidence hub, not a claim that
+every historical checkpoint is production-ready.
+
+## Archival Policy
+
+This closeout does not delete source code or historical evidence. Code remains
+authoritative in git. Heavy experiment artifacts remain authoritative in
+object storage. Hugging Face holds the curated shippable model surface. Local
+project documentation remains a lightweight navigation and recovery index.
+
+No failed mechanism checkpoint is promoted merely because it is newer. The
+V78 recovered native checkpoint is the shippable decoder lineage; V90 is static
+evaluation evidence; V29 and later checkpoints are research lineage.
+
+## Reopen Decision
+
+Reopen only for one of these reasons:
+
+1. reproduce the July 2 low-hull anchor under frozen lineage;
+2. build a low-hull endpoint or composition-conditioned crystal-structure
+   prediction pilot with a matched diffusion/flow objective;
+3. integrate a stronger external stability model as a transparent product
+   routing stage;
+4. address a concrete product requirement that the shipped static generator
+   does not satisfy.
+
+Do not reopen for another generic loss-weight, learning-rate, router, profile
+adapter, or static-validity sweep.
+
+The project is archived indefinitely rather than terminated permanently. A
+future restart should begin from the frozen evidence and reproduction contract,
+not from the assumption that the metastability campaign was a total failure.

docs/final_closeout/NEXA_MAT_FINAL_MODEL_CARD_2026-07-24.md

@@ -0,0 +1,121 @@
+---
+library_name: pytorch
+tags:
+  - materials-science
+  - crystal-generation
+  - diffusion
+  - inorganic-materials
+  - nexamat
+license: other
+---
+
+# NexaMat2 Model Card
+
+## Model Summary
+
+NexaMat2 is a compact materials-generation stack centered on a frozen physics
+encoder and a native diffusion decoder for inorganic crystal proposals. The
+full compact encoder-decoder stack contains 4,061,859 parameters. The decoder
+contains 1,248,145 trainable parameters.
+
+The model was recovered from a low-yield geometry generator through explicit
+native accounting, profile-conditioned chemistry contracts, radius-aware
+periodic geometry, and bounded external evaluation. It is shipped as a
+candidate generator, not as a substitute for relaxation or thermodynamic
+stability calculations.
+
+## Components
+
+| Component | Parameters | Function |
+| --- | ---: | --- |
+| Physics encoder | 2,813,714 | Physics features and plausibility grounding |
+| Diffusion decoder | 1,248,145 | Native atom, coordinate, and lattice generation |
+| Full compact stack | 4,061,859 | Candidate proposal and physics grounding |
+| Controller pilot | Separate Qwen3-14B artifact | Optional evidence routing and explanation |
+
+## Intended Use
+
+- Generate native, unrepaired inorganic candidate structures.
+- Produce diverse candidate pools for downstream screening.
+- Combine with the encoder for physics-aware ranking.
+- Support research and product prototyping with explicit external validation.
+
+## Out-of-Scope Use
+
+- Treating generated structures as experimentally stable without relaxation and
+  hull evaluation.
+- Claiming SOTA parity with MatterGen from the internal static profile suite.
+- Organic, molecular-crystal, polymer, or MOF generation.
+- Safety-critical or autonomous synthesis decisions without expert review.
+
+## Verified Evidence
+
+The final non-alloy static suite passed profile-specific gates:
+
+| Profile | Raw | Accepted | Charge | LeMat-distance | Strict | Novelty |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| SSB regression | 1,440 | 513 | 100.0% | 35.7% | 60.4% | 100.0% |
+| Battery ionics | 1,248 | 411 | 100.0% | 33.1% | 60.3% | 100.0% |
+| Oxide catalysts | 1,584 | 424 | 100.0% | 26.8% | 59.7% | 100.0% |
+| Halides/chalcogenides | 2,688 | 436 | 100.0% | 16.3% | 55.1% | 100.0% |
+| Semiconductors | 672 | 256 | profile diagnostic | 38.1% | 65.0% | 100.0% |
+| Broad inorganic | 2,688 | 242 | profile diagnostic | 9.0% | 55.1% | 100.0% |
+
+One prior bounded external stability test produced:
+
+- 2/64 M.S.U.N. candidates;
+- mean evaluated `e_above_hull = 0.8750 eV/atom`;
+- two individual candidates at 0.1786 and 0.1837 eV/atom;
+- 3/64 strict relaxation convergence.
+
+This is evidence that NexaMat can produce isolated candidates near 0.18
+eV/atom, while the bounded cohort mean remained 0.8750 eV/atom. It does not
+prove repeatable high-yield metastable generation or establish 2/64 as a
+current production rate. The historical 3.125% result was well below the
+intended 25-50% yield, and later cohorts did not fully verify its recurrence. A
+broader 512-candidate screen converged 58/512, had weighted mean hull 1.2330
+eV/atom, and produced no M.S.U.N. candidates.
+
+## Limitations
+
+- Relaxation convergence and low-hull yield are substantially weaker than
+  static validity.
+- The low-hull anchor did not reproduce reliably enough in later cohorts.
+- Alloys/intermetallics remain outside the passed static profile panel.
+- Internal and public model comparisons use different protocols.
+- The optional 14B controller is a partial pilot and is not part of the compact
+  4.061859M-parameter claim.
+
+## Required Validation
+
+Every generated candidate should pass:
+
+1. profile-specific static validation;
+2. external geometry and post-relax validation;
+3. MLIP or DFT relaxation appropriate to the intended use;
+4. formation-energy and hull evaluation against a compatible reference set;
+5. domain-expert review before synthesis or deployment decisions.
+
+## Artifact Lineage
+
+- Encoder: `encoder/v1/`
+- Original production decoder: `decoder/diffusion_v1/`
+- Recovered native decoder: `decoder/native_v78/`
+- Research initialization: `research/v29_directional_step800/`
+- Final documents: `docs/final_closeout/`
+- Evidence summaries: `eval/final_evidence/`
+
+The V78 checkpoint is the shippable recovered native lineage. V90 is an
+evaluation surface without a train checkpoint at the historically configured
+URI. Later conservative-field checkpoints are archived research evidence and
+are not production promotions.
+
+The metastability program is on ice indefinitely because its duration and cost
+outgrew the reliability gains. It is not classified as a complete scientific
+failure: the archived low-hull candidates provide a real restart anchor.
+
+## Citation
+
+Use the technical paper in
+`docs/final_closeout/NEXA_MAT_FINAL_TECHNICAL_PAPER_2026-07-24.md` and cite the
+exact model-hub revision used for an experiment.

docs/final_closeout/NEXA_MAT_FINAL_PROJECT_COMPENDIUM_2026-07-24.md

@@ -0,0 +1,260 @@
+# NexaMat Final Project Compendium
+
+Date: 2026-07-24  
+Project state: archived and on ice indefinitely  
+Model state: shipped with qualified claims
+
+## 1. Executive Summary
+
+NexaMat began as a compact but brittle materials generator. Its atom predictions
+were strong, its outputs were novel, and its encoder carried useful physical
+signals, but the decoder placed too little probability mass on valid periodic
+geometry. Early external runs produced only a handful of strict CIFs from
+thousands of samples. This made every downstream stability benchmark appear
+catastrophic because the expensive scientific stack was starved before it could
+evaluate enough usable structures.
+
+The alignment campaign solved that first problem. It introduced honest native
+accounting, explicit failure reasons, profile-conditioned chemistry, radius-aware
+geometry, charge-aware composition logic where appropriate, best-of-N native
+rollouts, and target-gated evaluation. The result is a 4.061859M-parameter stack
+that produces useful native candidate pools across six non-alloy inorganic
+profiles.
+
+The second problem was harder. A statically valid crystal can still sit far from
+a stable energy basin. In one prior bounded test, NexaMat produced two M.S.U.N.
+candidates from 64 and individual hull values near 0.18 eV/atom, while the
+cohort mean remained 0.8750 eV/atom. This shows isolated low-hull capability,
+not a typical 0.18 eV/atom distribution or a verified production yield. Later
+cohorts did not fully verify recurrence of the historical 3.125% result, which
+was already far below the intended 25-50% yield. Several weeks of force,
+trajectory, conservative-energy, tensor, routed-expert, and
+foundation-residual experiments improved proxies without moving the external
+objective reliably. The project is therefore put on ice indefinitely at a
+clear plateau rather than being called a complete failure or being kept alive
+through unproductive iteration.
+
+## 2. What The Model Is
+
+NexaMat is a modular materials stack:
+
+```mermaid
+flowchart LR
+    P["Profile and generation request"] --> D["Native diffusion decoder"]
+    D --> C["Native CIF candidate"]
+    C --> S["Static profile gate"]
+    S --> E["Physics encoder and ranking"]
+    E --> R["External relaxation"]
+    R --> H["Hull and metastability evaluation"]
+    H --> O["Evidence-backed candidate set"]
+```
+
+The decoder jointly proposes atom identities, fractional coordinates, and a
+periodic lattice. The encoder supplies physics-aware representation and scoring.
+External relaxation and hull evaluation remain authoritative for stability.
+
+The compact stack excludes the optional Qwen3-14B controller:
+
+```text
+physics encoder:          2,813,714 parameters
+diffusion decoder:        1,248,145 parameters
+compact full stack:       4,061,859 parameters
+```
+
+## 3. The Original Failure
+
+The initial decoder learned chemistry more readily than global crystal geometry.
+Atom vocabulary and novelty saturated while lattice scale, periodic pair
+distances, charge-compatible composition, and coordinate placement remained
+brittle. The conceptual mistake was treating generic denoising loss and
+teacher-forced validation as adequate substitutes for the density of native
+benchmark-eligible CIFs.
+
+The decoder was therefore good at producing different-looking compositions but
+poor at instantiating them as usable crystals. The encoder became a crutch
+because it had to reject nearly everything.
+
+## 4. The Static-Validity Recovery
+
+The recovery succeeded because it changed the contract, not merely the loss:
+
+- raw candidates were retained;
+- native and repaired surfaces were separated;
+- failure reasons became explicit;
+- SSB chemistry became a profile rather than a global hardcoded identity;
+- ionic and non-ionic profiles used different charge policies;
+- radius-aware site assignment and lattice checks attacked periodic geometry;
+- target-gated early stopping reduced wasted benchmark time;
+- checkpoint selection used external native yield instead of final-step loss.
+
+The strongest SSB lane, V90, reached:
+
+| Metric | Result |
+| --- | ---: |
+| Accepted | 595/4096 |
+| Filter accept | 14.53% |
+| Allowed elements | 100.00% |
+| Charge neutrality | 100.00% |
+| LeMat-distance validity | 14.53% |
+| Strict validity | 54.20% |
+| Formula uniqueness | 41.80% |
+| Novelty | 100.00% |
+
+The final six-profile suite accepted 513 SSB, 411 battery, 424 oxide, 436
+halide/chalcogenide, 256 semiconductor, and 242 broad-inorganic candidates. The
+last two were tolerant passes against profile-specific targets. Alloys and
+intermetallics remained unresolved.
+
+## 5. The Stability Gap
+
+Static validity asks whether a crystal is structurally and chemically plausible
+enough to evaluate. Metastability asks whether relaxation places it close to a
+competitive energy basin. The latter is a much narrower target.
+
+The broad 512-candidate screen showed the gap:
+
+```text
+relaxation convergence:      58 / 512
+weighted mean hull:          1.2330 eV/atom
+M.S.U.N. candidates:         0 / 512
+post-relax validity:         high
+```
+
+The best bounded low-hull anchor improved the scientific picture:
+
+```text
+relaxation convergence:      3 / 64
+mean hull:                    0.8750 eV/atom
+M.S.U.N. candidates:         2 / 64
+best individual hull rows:   0.1786 and 0.1837 eV/atom
+```
+
+The result matters because it proves the generator can reach the desired region.
+It is not enough because the probability mass remains sparse, the observed
+2/64 yield is far below the 16-32/64 practical ambition, and the result did not
+reproduce consistently enough for product reliability.
+
+## 6. Metastability Campaign
+
+The project tested a broad set of hypotheses:
+
+1. force and stress auxiliary supervision;
+2. real relaxation trajectories;
+3. normalized geometry and cell-shape context;
+4. scalar conservative energy surfaces;
+5. bounded mobility and symmetric lattice strain;
+6. rank-2 angular refinement;
+7. profile adapters and FiLM;
+8. routed periodic experts;
+9. source scheduling and profile balancing;
+10. MACE relation and force distillation;
+11. a fresh broad force foundation;
+12. frozen MACE-OMat plus a learned residual potential.
+
+Several internal signals became useful. V32 reached true-label ROC-AUC 0.70845
+and minimum balanced accuracy 0.60048. V30 reached force MAE 0.14260 eV/A and
+low-force MAE 0.07715 eV/A. The frozen MACE-OMat foundation reached force cosine
+0.75450 and MAE 0.05728 eV/A.
+
+The generative controller did not improve accordingly. V32 moved coordinates,
+lattice, and pair distances in the wrong direction and became non-finite. V30
+had negative coordinate and pair contraction. The final residual grid stayed
+near its frozen foundation baseline and did not justify mobility or native
+generation stages.
+
+## 7. Why The Redesigns Plateaued
+
+The redesigns attacked useful intermediate mechanisms, but they did not change
+the native endpoint distribution strongly enough. The system repeatedly asked
+a generator trained primarily for denoising and static geometry to behave like
+a low-hull endpoint model after modest continuation.
+
+Four deeper mismatches remained:
+
+- training data was not consistently filtered and weighted around broad relaxed
+  low-hull endpoints;
+- the native sampler was not always the exact reverse process implied by the
+  training corruption;
+- composition, coordinates, and lattice were not coupled strongly enough by
+  one endpoint objective;
+- internal force, critic, and contraction gates delayed direct external
+  metastability feedback.
+
+This explains how proxy metrics improved while metastable yield stayed flat.
+
+## 8. Technical Innovations
+
+The project produced reusable engineering and scientific mechanisms:
+
+- replay-grade candidate JSONL with raw payloads and failure reasons;
+- strict separation of native, projected, scaffolded, ranked, relaxed, and
+  hull-scored surfaces;
+- profile-conditioned element, charge, radius, and benchmark contracts;
+- charge-balanced composition search;
+- radius-aware periodic geometry checks;
+- generated hard-negative replay;
+- external noisy-state training paths;
+- stability evidence ingestion with provenance-safe joins;
+- source-disjoint trajectory and composition-family splits;
+- bounded symmetric lattice updates with positive volume;
+- target-gated 64-to-128-to-512 external promotion;
+- streaming and resumable benchmark operations;
+- object-storage-first lineage, manifests, receipts, and W&B references.
+
+## 9. Honest Product Position
+
+NexaMat is ready to ship as a compact proposal model when the product preserves
+an explicit screening funnel:
+
+```text
+native generation
+-> static profile validation
+-> encoder/ranker
+-> external relaxation
+-> hull and novelty checks
+-> expert review
+```
+
+It should not present raw candidates as stable materials. Its value is efficient
+candidate generation and unusually strong static coverage for its parameter
+count, not SOTA stability proof.
+
+## 10. Lessons
+
+1. Optimize the final scientific outcome from the beginning.
+2. Keep benchmark surfaces separate so repair cannot masquerade as generation.
+3. A good classifier or force model is not automatically a good actuator.
+4. Sparse success proves support, not distribution-level competence.
+5. Freeze evaluator, sampler, seeds, and lineage before interpreting trends.
+6. Use small external probes early; do not wait for a large campaign to reveal
+   objective mismatch.
+7. Preserve negative results. The failed architectures sharply narrowed the
+   remaining problem.
+8. Stop when repeated redesigns move proxies but not the external objective.
+
+## 11. Reopening Strategy
+
+The best future experiment is a composition-conditioned low-hull endpoint pilot:
+
+1. assemble relaxed primitive structures with reliable hull evidence;
+2. separate near-hull, metastable, and higher-hull slices;
+3. train a mathematically matched wrapped-coordinate diffusion or Riemannian
+   flow over coordinates and lattice;
+4. warm-start useful chemistry embeddings but not the failed controller;
+5. evaluate fixed compositions first;
+6. run the same frozen 64-candidate relaxation/hull protocol.
+
+If fixed-composition generation succeeds, composition is the next bottleneck.
+If teacher denoising succeeds but free sampling fails, the sampler is the
+bottleneck. If fixed-composition geometry fails, the periodic generator
+architecture must be replaced.
+
+## 12. Final Verdict
+
+NexaMat did not reach the reliable high-yield stable-material generator
+originally targeted, but it is not a failed project. It transformed a nearly
+unusable decoder into a compact, auditable, multi-profile inorganic candidate
+generator, demonstrated sparse metastable support, and established the exact
+scientific boundary that remained. The project is archived and placed on ice
+indefinitely with a useful shipped model, durable evidence, and a decisive
+restart contract.

docs/final_closeout/NEXA_MAT_FINAL_TECHNICAL_PAPER_2026-07-24.md

@@ -0,0 +1,386 @@
+# NexaMat: Recovering a Compact Native Crystal Generator and Locating the Metastability Boundary
+
+## Abstract
+
+NexaMat is a compact 4.061859M-parameter materials stack comprising a
+2.813714M-parameter physics encoder and a 1.248145M-parameter diffusion decoder.
+The initial decoder learned atom identity and novelty substantially faster than
+periodic geometry, producing very low native benchmark yield. We report a
+multi-stage recovery that introduced replay-grade candidate accounting,
+profile-conditioned chemistry policies, charge-aware composition handling,
+radius-aware periodic geometry, native best-of-N rollouts, and external
+target-gated evaluation. The final native, no-repair suite passed profile-specific
+static gates for six non-alloy inorganic domains and preserved 100% novelty.
+
+We then investigated whether the recovered generator could be aligned to
+relaxable low-hull basins. One prior bounded experiment produced two M.S.U.N.
+candidates from 64 selected structures and individual hull values of 0.1786
+and 0.1837 eV/atom, while the cohort mean remained 0.8750 eV/atom. This
+demonstrates sparse capability near 0.18 eV/atom, not typical performance at
+that level or a verified production yield. Only 3/64 structures met strict
+relaxation convergence, later cohorts did not fully verify recurrence of the
+historical 3.125% result, a 512-candidate screen produced zero M.S.U.N.
+candidates, and subsequent force, trajectory, conservative-field, tensor,
+routed-expert, and foundation-residual redesigns did not produce repeatable
+external improvement. We conclude that NexaMat is a useful compact native
+candidate generator with partial metastability evidence, but not a proven
+high-yield metastable generator. The central remaining issue is generative
+endpoint density rather than static validity or local force magnitude.
+
+## 1. Introduction
+
+Crystal generation couples discrete composition, periodic coordinates, and
+lattice geometry. A model can perform well on atom prediction while failing to
+place those atoms into a physically useful cell. It can also emit a statically
+valid crystal that relaxes into a poor or chemically uncompetitive energy basin.
+These are separate scientific problems and require separate evidence.
+
+NexaMat exposed both boundaries. Its first failure was native geometry density:
+only a tiny fraction of raw samples could be exported as strict CIFs. Its second
+failure emerged after that problem was repaired: static-valid candidates did
+not reliably relax into near-hull structures.
+
+This paper documents the complete system, the alignment mechanisms that worked,
+the metastability experiments that did not, and the evidence boundary of the
+shipped model.
+
+## 2. Model
+
+### 2.1 Physics encoder
+
+The encoder contains 2,813,714 parameters and was trained on mixed relaxed,
+rattled, and off-equilibrium inorganic structures. It provides physics-aware
+representations and plausibility signals. The encoder is frozen during the
+shipped native-generation workflow.
+
+### 2.2 Diffusion decoder
+
+The decoder contains 1,248,145 trainable parameters and jointly predicts:
+
+- atom identities;
+- fractional coordinate updates;
+- lattice updates;
+- optional static and stability diagnostics.
+
+Periodic neighbor information is represented through distance and direction
+features. Later research variants added scalar energy, force, stress, bounded
+mobility, symmetric strain, rank-2 angular state, routed experts, and foundation
+residuals. These variants are research lineage, not shipped promotions.
+
+### 2.3 Candidate surfaces
+
+The final accounting contract preserves distinct surfaces:
+
+1. native candidate;
+2. static-qualified candidate;
+3. projected or scaffolded candidate;
+4. critic-ranked candidate;
+5. relaxed candidate;
+6. hull-scored candidate.
+
+Only native candidates count toward native-generation claims. Projection,
+scaffolding, ranking, or relaxation cannot overwrite native metrics.
+
+## 3. Data And Evidence
+
+The project used canonical crystal structures, profile-specific datasets,
+generated candidate replay, relaxation trajectories, force and stress labels,
+and external hull evidence. Sources included SSB regression data, broad
+inorganic rows, OMat-style force data, MPTrj trajectories, JARVIS/LeMat-style
+evidence where available, and generated NexaMat failures.
+
+Each stability row was designed to preserve:
+
+- source and material identity;
+- composition family;
+- atomic numbers, fractional coordinates, and lattice;
+- trajectory frame and time;
+- force, stress, and energy labels with masks;
+- convergence and post-relax validity;
+- hull evidence;
+- provenance hashes.
+
+Splits were source- and material-disjoint where the mechanism experiments
+required transfer evidence.
+
+## 4. Native-Validity Alignment
+
+### 4.1 Failure diagnosis
+
+The original benchmark showed a severe mismatch: chemical plausibility and
+novelty were much stronger than allowed-element and periodic-distance validity.
+Generic denoising loss did not measure usable native CIF density, and final-step
+checkpoint selection could prefer a model whose free sampler had regressed.
+
+### 4.2 Recovery mechanisms
+
+The recovery introduced:
+
+- candidate JSONL containing raw atoms, coordinates, lattice, and explicit
+  failure reasons;
+- hard-negative replay from generated failures;
+- allowed-element and charge pressure;
+- profile-specific element, charge, and radius policies;
+- radius-aware periodic geometry;
+- charge-balanced composition search for ionic profiles;
+- formula-diverse selection;
+- native best-of-N rollouts;
+- external checkpoint selection by accepted yield and geometry metrics.
+
+### 4.3 Static results
+
+The strongest V90 SSB diagnostic accepted 595/4096 native samples, with 100%
+allowed elements, 100% charge neutrality, 54.20% strict validity, 41.80%
+selected formula uniqueness, and 100% novelty.
+
+The final multi-profile suite produced:
+
+| Profile | Raw | Accepted | Filter accept | LeMat-distance | Strict | Novelty |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| SSB regression | 1,440 | 513 | 35.6% | 35.7% | 60.4% | 100.0% |
+| Battery ionics | 1,248 | 411 | 32.9% | 33.1% | 60.3% | 100.0% |
+| Oxide catalysts | 1,584 | 424 | 26.8% | 26.8% | 59.7% | 100.0% |
+| Halide/chalcogenides | 2,688 | 436 | 16.2% | 16.3% | 55.1% | 100.0% |
+| Semiconductors | 672 | 256 | 38.1% | 38.1% | 65.0% | 100.0% |
+| Broad inorganic | 2,688 | 242 | 9.0% | 9.0% | 55.1% | 100.0% |
+
+![Figure 1. Native candidate yield by profile](../../figures/nexamat_profile_yield_vs_targets.png)
+
+*Figure 1. Native accepted counts against profile-specific diagnostic targets.
+The targets are not a shared public leaderboard; they are the production
+tolerances used by the final internal profile suite.*
+
+![Figure 2. Geometry validity rates by profile](../../figures/nexamat_profile_geometry_rates.png)
+
+*Figure 2. LeMat-radius and strict-geometry rates. The plot shows why accepted
+yield alone is insufficient: broad-inorganic and halide/chalcogenide candidates
+remain weaker on radius-compatible geometry even when the static gate is
+otherwise usable.*
+
+These metrics use profile-specific gates and are not public MatterGen-equivalent
+stability metrics.
+
+## 5. External Stability Evaluation
+
+The external funnel consisted of native generation, static qualification,
+selection, short relaxation, post-relax validation, and hull/SUN-style scoring.
+
+The broad 512-candidate selector screen produced:
+
+```text
+converged:                 58 / 512
+weighted mean hull:        1.2330 eV/atom
+M.S.U.N.:                  0 / 512
+post-relax validity:       high
+```
+
+One prior bounded stable-manifold probe produced:
+
+```text
+converged:                 3 / 64
+mean hull:                 0.8750 eV/atom
+M.S.U.N.:                  2 / 64
+best hull rows:            0.1786, 0.1837 eV/atom
+```
+
+A follow-up force-basin probe reported mean hull 0.8165 eV/atom and retained
+2/64 hits but converged only 4/64 structures. Later reproduction cohorts did not
+establish stable recurrence. These historical results show that isolated
+low-hull candidates are reachable, but they do not establish a reliable current
+metastable distribution or production yield.
+
+![Figure 3. Historical stability evidence](../../figures/nexamat_metastability_evidence.svg)
+
+*Figure 3. Cohort mean energy above hull and the two best historical individual
+samples. The low individual points are not cohort means. Cohorts were evaluated
+at different campaign stages and should be read as bounded evidence, not as a
+single calibrated production estimate.*
+
+## 6. Mechanism Experiments
+
+### 6.1 Trajectory and force supervision
+
+Real trajectory windows, force, stress, energy descent, and endpoint displacement
+were added to the training contract. V30 reached force MAE 0.14260 eV/A and
+low-force MAE 0.07715 eV/A, but force cosine remained 0.30881, coordinate
+contraction was -2.486%, and pair contraction was -14.555%.
+
+### 6.2 Stability representation
+
+V32 corrected the true-label stability policy and reached ROC-AUC 0.70845 and
+minimum balanced accuracy 0.60048. This demonstrated that the shared
+representation contained a transferable basin-quality signal.
+
+The controller failed. Held-out force MAE was 0.35973 eV/A, force cosine was
+0.23114, one-step coordinate/lattice/pair contractions were all negative, and
+the four-step higher-order rollout became non-finite.
+
+### 6.3 Broader conservative fields
+
+The project tested profile adapters, FiLM, periodic residual experts, tensor
+angular refinement, source scheduling, MACE force mixing, MACE relation
+distillation, and a fresh force foundation. The custom NexaMat force field
+plateaued near 0.306 force cosine on the frozen bridge, while MACE-OMat reached
+approximately 0.776 on the audit surface.
+
+The final foundation-residual campaign used frozen MACE-OMat plus a trainable
+equivariant residual. The frozen baseline reached cosine 0.75450 and MAE 0.05728
+eV/A. Activating the residual gradient path and sweeping residual scale and
+learning rate did not improve held-out direction. The best residual result was
+0.75322, and no controller or native screen was justified.
+
+### 6.4 Representation-manifold audit
+
+The representation audit extracted node-pooled and shared stability-context
+states for 1,024 source/target pairs. It projected source, intermediate, and
+relaxed states with PCA, t-SNE, and UMAP, then measured stability separation,
+cross-profile source-to-target direction agreement, and profile-level
+contraction.
+
+![Figure 4. Node-pooled PCA manifold](../../figures/manifold/node_pooled_pca.png)
+
+*Figure 4. The node-pooled PCA manifold contains broad profile structure but
+does not cleanly separate relaxation stage, force, convergence, or contraction.
+Cross-profile stability AUC was 0.5183 and minimum balanced accuracy was
+0.4216.*
+
+![Figure 5. Stability-context PCA manifold](../../figures/manifold/stability_context_pca.png)
+
+*Figure 5. The stability-context representation organizes domains more clearly
+than the raw pooled state, but stable and unstable rows still overlap. Mean
+cross-profile AUC was 0.5378 and minimum balanced accuracy was 0.3627 in this
+V31 audit.*
+
+The source-to-target directions were also inconsistent. Mean cross-profile
+direction cosine was 0.0856 for the node-pooled state and -0.0077 for the
+stability context; the corresponding minima were -0.1763 and -0.4849. These
+values explain why a representation could encode profile or local-force
+information without providing one transferable global relaxation direction.
+
+![Figure 6. Node-pooled UMAP families](../../figures/manifold/node_pooled_umap_families.png)
+
+*Figure 6. Family envelopes in the node-pooled UMAP. Most families would require
+large contraction to become non-overlapping, indicating a diffuse local
+representation.*
+
+![Figure 7. Stability-context UMAP families](../../figures/manifold/stability_context_umap_families.png)
+
+*Figure 7. Family envelopes in the stability-context UMAP. Several families are
+well separated, while others remain broad. HDBSCAN identified nine families
+and classified 25.2% of rows as sparse or unassigned.*
+
+The manifold evidence changed the diagnosis. The model did not lack all
+structure: it learned domains, local neighborhoods, and some stability-related
+features. It lacked a domain-general mechanism that mapped those features to a
+consistent global coordinate and lattice update. V32 later improved the
+true-label stability classifier to ROC-AUC 0.70845, but the controller still
+failed, confirming that representation quality and generative control were
+separate bottlenecks.
+
+## 7. Discussion
+
+### 7.1 Why static validity improved
+
+Static recovery used direct, localizable signals and explicit sampler contracts.
+The model could learn allowed chemistry, charge policy, radius compatibility,
+lattice bounds, and candidate rejection. These constraints substantially
+increased benchmark-eligible density.
+
+### 7.2 Why metastability plateaued
+
+Metastability is an endpoint-distribution problem. A local force model may know
+a downhill direction without teaching a generator to initialize near a globally
+competitive basin. A stability critic may rank structures without providing an
+accurate actuator. A controller may improve teacher-forced trajectories while
+failing under free native rollouts.
+
+The campaign retained a decoder whose pretraining contract emphasized denoising
+and conservative lattice behavior. Later continuation attempted to reshape that
+distribution using relatively small trajectory and stability objectives. The
+resulting improvements were real but mostly proxy-level.
+
+The likely remaining requirements are:
+
+- a broad low-hull endpoint training distribution;
+- a corruption and reverse process matched exactly at sampling time;
+- tighter joint modeling of composition, coordinates, and lattice;
+- direct and frequent external endpoint feedback;
+- fixed-composition ablations that separate geometry from composition failure.
+
+## 8. Efficiency
+
+The compact stack is approximately 8.7% of MatterGen's reported 46.8M parameter
+count. This is a meaningful engineering result for static candidate generation.
+It must not be converted into a stability-performance claim because the models
+were not evaluated under the same full protocol.
+
+The optional 14B controller is separate and must not be included in the compact
+parameter-efficiency claim.
+
+## 9. Limitations
+
+- No repeatable high-yield metastable or SUN/MSUN result.
+- No identical public MatterGen benchmark.
+- Alloys/intermetallics remain unresolved.
+- Hull values depend on evaluator, reference data, and relaxation coverage.
+- Sparse low-hull hits are sensitive to checkpoint, sampler, seeds, and
+  evaluator lineage.
+- Recent force-field research checkpoints are not shippable native generators.
+- Custom readiness scores and qualitative public-model matrices are excluded
+  from the scientific evidence figures because they are not shared-protocol
+  measurements.
+
+## 10. Conclusion
+
+NexaMat demonstrates that a very small materials stack can recover strong native
+inorganic candidate generation when the benchmark and sampler contracts are
+made explicit. It also demonstrates that static validity, representation
+quality, and force accuracy do not automatically produce a metastable
+distribution.
+
+The project closes with a useful shipped model and a clear scientific boundary.
+It is placed on ice indefinitely because the multi-week campaign's remaining
+reliability gap no longer justified continuous iteration. Future work should
+not extend the controller lineage incrementally. It should first reproduce the
+low-hull anchor and then test a matched low-hull endpoint generator under
+fixed-composition and free-composition settings.
+
+## Appendix A. Complete Manifold Projection Set
+
+The main text uses the PCA and family-envelope views for interpretation. The
+remaining t-SNE and UMAP projections are included here to make the visual
+evidence complete without assigning scientific meaning to any single
+two-dimensional projection.
+
+![Figure 8. Node-pooled t-SNE](../../figures/manifold/node_pooled_tsne.png)
+
+*Figure 8. Node-pooled t-SNE colored by relaxation stage, profile, force,
+energy reduction, convergence, and coordinate contraction.*
+
+![Figure 9. Node-pooled UMAP](../../figures/manifold/node_pooled_umap.png)
+
+*Figure 9. Node-pooled UMAP using the same evidence surfaces.*
+
+![Figure 10. Stability-context t-SNE](../../figures/manifold/stability_context_tsne.png)
+
+*Figure 10. Stability-context t-SNE. Profile organization is visible, but
+convergence and contraction do not form one transferable low-dimensional
+direction.*
+
+![Figure 11. Stability-context UMAP](../../figures/manifold/stability_context_umap.png)
+
+*Figure 11. Stability-context UMAP. This view supports the same conclusion as
+the PCA and family analysis: structured domain information is present, while
+global actuator direction remains weak.*
+
+## Artifact Availability
+
+Curated model artifacts and final documentation:
+
+```text
+https://huggingface.co/AethronPhantom/Nexa_Mat2
+```
+
+The complete project archive and historical run lineage are retained in
+private immutable storage. Their locations and operational identifiers are
+intentionally withheld from the public release.

docs/final_closeout/PROJECT.md

@@ -0,0 +1,92 @@
+# NexaMat Archived Project Record
+
+Project: NexaMat / NexaMat2  
+Status: Archived and on ice indefinitely; model shipped as a qualified research and candidate-generation artifact  
+Archive date: 2026-07-24  
+Model hub: `https://huggingface.co/AethronPhantom/Nexa_Mat2`  
+
+## Purpose
+
+NexaMat is a compact materials model stack for proposing and evaluating inorganic
+crystal candidates. The project recovered a diffusion decoder that initially
+produced very low-yield geometry into a useful native, no-repair candidate
+generator across six non-alloy inorganic profiles. It then investigated the
+harder problem of repeatable metastable generation.
+
+## Shipped Claim
+
+The shipped model is suitable for:
+
+- profile-conditioned native inorganic candidate generation;
+- auditable candidate filtering and ranking;
+- chemistry-aware proposal workflows using the frozen physics encoder;
+- research, screening, and application prototyping where external relaxation
+  and stability evaluation remain mandatory.
+
+The shipped model is not claimed to be:
+
+- a high-yield metastable or SUN/MSUN generator;
+- a replacement for relaxation, DFT, MLIP, or hull evaluation;
+- a universally validated generator for alloys, organics, MOFs, polymers, or
+  molecular crystals;
+- directly comparable to MatterGen under an identical public stability protocol.
+
+## Canonical Model State
+
+| Component | Parameters | Role | Status |
+| --- | ---: | --- | --- |
+| Physics encoder | 2,813,714 | Plausibility and physics grounding | Shipped, frozen |
+| Diffusion decoder | 1,248,145 | Native atom, coordinate, and lattice proposal | Shipped |
+| Full compact stack | 4,061,859 | Encoder plus decoder | Shipped |
+| Multimodal controller | Separate 14B pilot | Evidence routing and explanation | Optional, partial pilot |
+
+The production decoder lineage is the recovered V78 native checkpoint. V90 is
+an evaluation artifact produced from a provided checkpoint; its recorded run
+prefix does not contain a trainable `best_diffusion_checkpoint.pt`. Later
+V26-V36 conservative-field experiments are archived as research evidence and
+must not be substituted for the shipped native decoder.
+
+## Final Scientific State
+
+- Static native generation: useful and production-tolerable on SSB, battery,
+  oxide, halide/chalcogenide, semiconductor, and broad-inorganic profiles.
+- Novelty: 100% in the final six-profile static suite.
+- Best bounded metastability anchor: one prior 64-candidate test reported 2/64
+  M.S.U.N. candidates and mean evaluated `e_above_hull = 0.8750 eV/atom`.
+  Individual candidates reached 0.1786 and 0.1837 eV/atom. Those examples show
+  that NexaMat can reach approximately 0.18 eV/atom; they do not mean its
+  typical output lies at 0.18 eV/atom.
+- Repeatability: not fully verified. The same prior test converged only 3/64
+  structures at the strict relaxation threshold, and later cohorts did not
+  establish reliable recurrence. Its historical 2/64 result is evidence of
+  capability, not a current production-rate estimate. It was also well below
+  the intended 25-50% yield.
+- Final campaign verdict: meaningful partial progress followed by a scientific
+  plateau. The project was paused because the campaign had dragged on for weeks
+  and the remaining reliability gap no longer justified continuous iteration,
+  not because the model had no useful metastable support.
+
+## Canonical Documents
+
+1. `ARCHIVE_CLOSEOUT_2026-07-24.md`
+2. `results/papers/NEXA_MAT_FINAL_PROJECT_COMPENDIUM_2026-07-24.md`
+3. `results/papers/NEXA_MAT_FINAL_TECHNICAL_PAPER_2026-07-24.md`
+4. `results/papers/NEXA_MAT_FINAL_MODEL_CARD_2026-07-24.md`
+5. `results/papers/NEXA_MAT_ALIGNMENT_PIPELINE_KNOWLEDGE_PAPER.md`
+6. `results/papers/NEXA_MAT_METASTABLE_ARCHITECTURE_AND_TRAINING.md`
+7. `results/benchmarks/2026-07-24_foundation_residual_metastability_closeout.md`
+
+## Reopen Contract
+
+If the project is resumed, do not restart with another static-validity,
+force-loss, router, adapter, or controller sweep. First reproduce the July 2
+low-hull anchor from the exact checkpoint, sampler, seeds, candidate rows,
+relaxer, and hull evaluator. If it reproduces, train a mathematically matched
+low-hull endpoint generator and measure recurrence density. If it does not,
+classify the result as lineage or evaluator drift before changing the model.
+
+The curated shippable model surface and public evidence are preserved in this
+Hugging Face repository. Private experiment history remains separately
+archived. The project is on ice indefinitely, not abandoned; the retained
+artifacts are sufficient to support a controlled restart if the stability work
+becomes worthwhile again.

eval/FINAL_EVIDENCE_SUMMARY.md

@@ -0,0 +1,98 @@
+# NexaMat Final Public Evidence Summary
+
+This report is the public, sanitized evidence record for the archived NexaMat
+release. It contains aggregate scientific results and artifact-integrity
+information only. Private storage locations, infrastructure identifiers,
+experiment-control commands, account URLs, and internal dataset fingerprints
+are intentionally omitted.
+
+## Scope And Claims
+
+NexaMat is released as a compact native inorganic crystal candidate generator.
+It demonstrated useful static validity and novelty across a six-profile
+non-alloy panel. It also produced isolated low-hull candidates in one bounded
+historical test, but did not demonstrate repeatable high-yield metastable or
+M.S.U.N. generation.
+
+The full documented stack contains 4,061,859 parameters. The separately
+published V78 state is the recovered diffusion-decoder artifact, not the full
+encoder-decoder stack; its manifest reports 594,340 serialized parameter
+elements across 41 tensors.
+
+## Static Native Evidence
+
+All rows below are native/no-repair diagnostic outputs. Profile gates differ,
+so accepted counts should be interpreted within each profile rather than as a
+single shared leaderboard.
+
+| Profile | Raw candidates | Accepted | LeMat-radius valid | Strict geometry | Novelty |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| SSB regression | 1,440 | 513 | 35.7% | 60.4% | 100.0% |
+| Battery ionics | 1,248 | 411 | 33.1% | 60.3% | 100.0% |
+| Oxide catalysts | 1,584 | 424 | 26.8% | 59.7% | 100.0% |
+| Halides/chalcogenides | 2,688 | 436 | 16.3% | 55.1% | 100.0% |
+| Semiconductors | 672 | 256 | 38.1% | 65.0% | 100.0% |
+| Broad inorganic | 2,688 | 242 | 9.0% | 55.1% | 100.0% |
+
+The plots in `figures/` visualize these measured counts and rates without
+introducing a composite readiness score.
+
+## Stability Evidence
+
+These cohorts were produced at different stages and are not interchangeable.
+The individual low-hull values are examples, not cohort averages or a
+reproducible production-rate claim.
+
+| Evidence cohort | Pool | Strict convergence | Mean evaluated hull | M.S.U.N. evidence | Interpretation |
+| --- | ---: | ---: | ---: | ---: | --- |
+| Broad stability screen | 512 | 58/512 | 1.2330 eV/atom | 0 | Broad-screen failure |
+| Historical bounded screen | 64 | 3/64 | 0.8750 eV/atom | 2/64 | Partial support; low reliability |
+| Follow-up force-basin screen | 64 | 4/64 | 0.8165 eV/atom | 2/64 reported | Directional improvement; recurrence not established |
+
+The historical bounded screen included individual candidates at 0.1786 and
+0.1837 eV/atom. These values show that the model reached the low-hull region on
+isolated samples. They do not establish that typical output lies near 0.18
+eV/atom. The intended metastable-yield target was 25-50%; the historical 2/64
+result was materially below it.
+
+## Mechanism Evidence
+
+The final research phase separated representation quality from controller
+quality:
+
+| Metric | Result | Reading |
+| --- | ---: | --- |
+| True-label stability ROC-AUC | 0.70845 | Representation gate passed |
+| Minimum balanced accuracy | 0.60048 | Representation gate passed |
+| Held-out force MAE | 0.35973 eV/A | Controller gate failed |
+| Held-out force cosine | 0.23114 | Controller gate failed |
+| Coordinate/lattice/pair contraction | Negative | Controller gate failed |
+| Frozen physical-field force cosine | 0.75450 | Strong frozen baseline |
+| Frozen physical-field force MAE | 0.05728 eV/A | Strong frozen baseline |
+
+The evidence supports a specific conclusion: the model learned useful
+stability-related representations and local physical signals, but the trained
+actuator did not reliably convert those signals into globally improving native
+rollouts across domains. Later residual experiments did not improve the frozen
+physical-field baseline.
+
+## Limitations
+
+- No shared-protocol SOTA comparison is claimed.
+- The stability cohorts differ in selection and stage, so trends are
+  directional rather than a single calibrated production estimate.
+- The 2/64 historical M.S.U.N. result was not reproduced enough times to claim
+  a stable rate.
+- Alloys/intermetallics, organics, MOFs, polymers, and molecular crystals are
+  outside the qualified panel.
+- Generated candidates require independent relaxation, hull evaluation, and
+  expert review before scientific or synthesis decisions.
+
+## Release Decision
+
+The project is archived and on ice indefinitely because the metastability
+campaign consumed substantial time and compute without establishing the
+desired reliability. This is not a claim that the model is useless: it remains
+a compact, novel, static-valid candidate generator with partial low-hull
+support. The public release preserves that useful state while keeping the
+scientific limitations explicit.

file_manifest.json

@@ -0,0 +1,120 @@
+{
+  "schema_version": "nexamat.public_file_manifest/v1",
+  "files": [
+    {
+      "path": "Nexa_Mat_Paper.pdf",
+      "sha256": "14d030e6acb4db728419fefcb4750c5a5a96399423d8a82e094fd7f9d353a924",
+      "size_bytes": 6705702
+    },
+    {
+      "path": "README.md",
+      "sha256": "c7f163204e66fdec5b7974cc6da3c6ab7cf7d616dc08f4d712a84573e1d9b87e",
+      "size_bytes": 5514
+    },
+    {
+      "path": "decoder/native_v78/model.safetensors",
+      "sha256": "1e19e2c6c98c2ba8eb4b539d91be652128fe7cecb9877ff3b94b9ee62433f55c",
+      "size_bytes": 2381144
+    },
+    {
+      "path": "decoder/native_v78/model_manifest.json",
+      "sha256": "1f00d511c91ceb268e6fca7f1bbc7a427c09fda3c3c788c943113b3a765275a0",
+      "size_bytes": 1482
+    },
+    {
+      "path": "decoder/native_v78/model_state_dict.pt",
+      "sha256": "75498a20c6a25f2fa91b2a087960f18cbae803de095dc2358cb1c747f867878e",
+      "size_bytes": 2391887
+    },
+    {
+      "path": "docs/final_closeout/ARCHIVE_CLOSEOUT_2026-07-24.md",
+      "sha256": "be69c238dffa65e0a9a8611945a5028118dd76de98c313afa43c428ee35c2f9a",
+      "size_bytes": 6995
+    },
+    {
+      "path": "docs/final_closeout/NEXA_MAT_FINAL_MODEL_CARD_2026-07-24.md",
+      "sha256": "1662db0e7418fa179ddbe49731605fbecaf2bd9e2427225674a439d5cf0ecbc9",
+      "size_bytes": 4907
+    },
+    {
+      "path": "docs/final_closeout/NEXA_MAT_FINAL_PROJECT_COMPENDIUM_2026-07-24.md",
+      "sha256": "cd179a8a468ebf09b98a3e446fa06b2ea04ca54f7cd42d1666350860c2e69bd3",
+      "size_bytes": 10851
+    },
+    {
+      "path": "docs/final_closeout/NEXA_MAT_FINAL_TECHNICAL_PAPER_2026-07-24.md",
+      "sha256": "085b622acad2981c152bfa462947245d99d0c579543efd3824d875f727a8dc2e",
+      "size_bytes": 17212
+    },
+    {
+      "path": "docs/final_closeout/PROJECT.md",
+      "sha256": "8c70e54d6ae32708838f46ab3746c0ee73b57b35ef010705c767254144f9edfc",
+      "size_bytes": 4633
+    },
+    {
+      "path": "eval/FINAL_EVIDENCE_SUMMARY.md",
+      "sha256": "7bdd7bd1b2e82274655778e6fc808567dd28979d72f16338b308ab6596aa8c3f",
+      "size_bytes": 4751
+    },
+    {
+      "path": "figures/manifold/node_pooled_pca.png",
+      "sha256": "0857ca06309a891eb643b7438cdbb028ec9882d64cada7ac24f53b61341f542d",
+      "size_bytes": 1154853
+    },
+    {
+      "path": "figures/manifold/node_pooled_tsne.png",
+      "sha256": "4cb13d469143ed1a882ce6071069308bf09570ecaceb2aea9595eb78790f59d1",
+      "size_bytes": 917262
+    },
+    {
+      "path": "figures/manifold/node_pooled_umap.png",
+      "sha256": "6c2b9f8d5086269b89d4bbb0ef12086c94b4a8e954017184b4182d6b878185b9",
+      "size_bytes": 486157
+    },
+    {
+      "path": "figures/manifold/node_pooled_umap_families.png",
+      "sha256": "fda85949a83855899eb7ee9e064710442ba84eb1dd955ec2522fe35e00d242f0",
+      "size_bytes": 702478
+    },
+    {
+      "path": "figures/manifold/stability_context_pca.png",
+      "sha256": "6d8817f4a1ea05f47a0828626d369ff3d2771ddad5e197e43cb77153162d1825",
+      "size_bytes": 1029049
+    },
+    {
+      "path": "figures/manifold/stability_context_tsne.png",
+      "sha256": "2b9a44e9f9ee3a12e5a3c5652d21fb2e71203817e0a969406281f9f38fb6d32a",
+      "size_bytes": 920866
+    },
+    {
+      "path": "figures/manifold/stability_context_umap.png",
+      "sha256": "dbd2c4d8e6f59be2ca5e229f8492f0e1338d8809364f0465da5a9be33b59aa4b",
+      "size_bytes": 499473
+    },
+    {
+      "path": "figures/manifold/stability_context_umap_families.png",
+      "sha256": "80a6d97907b09cc7273fd53c8305ef628d26395459dc154472a81a0e6e915bd0",
+      "size_bytes": 387952
+    },
+    {
+      "path": "figures/nexamat_metastability_evidence.svg",
+      "sha256": "d8d0c2cdb9abd61de9c17dda5c79a31819df6f7807f0d11bca0c5d3144d621b1",
+      "size_bytes": 3814
+    },
+    {
+      "path": "figures/nexamat_profile_geometry_rates.png",
+      "sha256": "24e242889d65c9ec60a20454694a55324addb1b34252a7b3fc0c1e4eb04748f5",
+      "size_bytes": 115263
+    },
+    {
+      "path": "figures/nexamat_profile_yield_vs_targets.png",
+      "sha256": "609984d3796152b68c11504906541c829ae2525277f680a858ff4ddff4275636",
+      "size_bytes": 123667
+    },
+    {
+      "path": "final_publication_manifest.json",
+      "sha256": "39eaa17c087e7dd7a06ec7246cd85c04d66fb14559cbffa1af6e7c76baed49ac",
+      "size_bytes": 1648
+    }
+  ]
+}

final_publication_manifest.json

@@ -0,0 +1,40 @@
+{
+  "schema_version": "nexamat.final_publication/v2",
+  "project": "NexaMat",
+  "archive_status": "archived_on_ice_indefinitely",
+  "publication_date": "2026-07-25",
+  "model_hub": "AethronPhantom/Nexa_Mat2",
+  "compact_stack_parameters": 4061859,
+  "encoder_parameters": 2813714,
+  "decoder_parameters": 1248145,
+  "published_v78_state_parameter_elements": 594340,
+  "shipped_decoder_lineage": "v78_recovered_native_decoder",
+  "technical_paper": {
+    "path": "Nexa_Mat_Paper.pdf",
+    "sha256": "14d030e6acb4db728419fefcb4750c5a5a96399423d8a82e094fd7f9d353a924",
+    "pages": 16,
+    "embedded_figures": 11
+  },
+  "public_figure_policy": {
+    "included": "direct experiment metrics and representation-manifold diagnostics",
+    "excluded": "composite readiness scores and cross-model charts without a shared evaluation protocol"
+  },
+  "metastability_evidence": {
+    "historical_bounded_msun_count": 2,
+    "historical_bounded_pool_size": 64,
+    "historical_strict_relaxation_converged": 3,
+    "historical_mean_e_above_hull_ev_per_atom": 0.875,
+    "historical_best_e_above_hull_ev_per_atom": [
+      0.1786,
+      0.1837
+    ],
+    "interpretation": "partial low-hull support with low reliability; below the intended 25-50 percent yield"
+  },
+  "claim": "compact native inorganic candidate generator with external screening required",
+  "excluded_claims": [
+    "repeatable high-yield metastable or SUN/MSUN generation",
+    "shared-protocol SOTA parity",
+    "experimental stability without external relaxation and hull evaluation"
+  ],
+  "private_lineage": "preserved separately; storage locations and operational identifiers withheld"
+}