feat(007): trained inference path for seed-morph78

CHANGELOG.md

@@ -2,6 +2,11 @@
 
 ## Unreleased
 
+- **Trained inference (Spec 007):** `hyperlexical.infer --model-dir` runs a local
+  checkpoint (lazy torch, local trunk, fail-closed exit 2) and emits a
+  `MODEL_EMBEDDING` inference packet; only the approved pin carries
+  `hyperlex-structure-149m`. Default CLI stays the offline stub.
+
 - **Card rename (Spec 007):** Hub card `hyperlex-structure-149m` for `seed-morph78`.
   `name_gate.py` pins the approved checkpoint; `eval_unbind` sets `name_gate`
   true only for a trunk-forward eval of that pin; eval schema field is boolean.

NEXT_MOVES_007.md

@@ -7,6 +7,7 @@
 - morph78 named phrases settled · force/hard **236/277** · soft_ceiling train complete.
 - **PROMOTE_BEST** (ceiling_escape): broad **0.9883** > PRIOR morph65 **0.8867** n=256 + E2 PASS.
 - Card rename: Hub card `hyperlex-structure-149m`; eval packets set `name_gate` true only for a trunk-forward eval of `seed-morph78`.
+- Real inference: `hyperlexical.infer --model-dir ~/.hyperlex/models/BEST` → `MODEL_EMBEDDING` packet (`hyperlex-structure-149m`), verified on Spark.
 - Danny **`flip name_gate`** → `seed-morph78` may be called **Hyperlexical** (receipt `specs/007-hyperlexical-model/receipts/20260924-name-gate-yes-morph78.md`).
 
 ## Gate
@@ -15,9 +16,10 @@ soft_ceiling **spent** for this climb (morph78 now PRIOR/BEST). Further climbs n
 
 ## Next (each a separate operator sentence)
 
-1. Real `seed-morph78` inference path (trunk-forward infer emitting the inference packet) — in progress.
-2. Hub upload — still not authorized.
-3. Optional T13 (promote `scripts/shadow/hyperlexical/` into `src/hyperlex/`).
-4. HOLD morph78; no new climb without a new card.
+1. Hub upload — still not authorized (decide publish vs explicit local-only).
+2. Optional T13 (promote `scripts/shadow/hyperlexical/` into `src/hyperlex/`).
+3. HOLD morph78; no new climb without a new card.
+
+Open findings (receipt `receipts/20260924-infer-model-morph78.md`): classify overconfident on val-settled phrases; trunk+heads = 150.55M vs A1 150M (trunk 149.0M).
 
 Qwen stays stopped unless re-enabled.

ROADMAP.md

@@ -102,6 +102,7 @@ The skill remains the operator surface while the encoder stays SHADOW.
 See [SPEC.md](SPEC.md) and [DESIGN.md](DESIGN.md) for the historical spine. Current operator snapshot: [STATUS.md](STATUS.md).
 
 ## Recent
+- 2026-09-24: Real inference path — `infer --model-dir` on `seed-morph78` emits `MODEL_EMBEDDING` packets as `hyperlex-structure-149m` (Spark-verified).
 - 2026-09-24: Card rename — Hub card `hyperlex-structure-149m`; eval packets set `name_gate` true only for trunk-forward eval of `seed-morph78`.
 - 2026-09-24: Danny `flip name_gate` — `seed-morph78` named **Hyperlexical** (amendment A6). Card rename, Hub, T13 still separate.
 - 2026-09-24: morph78 PROMOTE_BEST (soft_ceiling ceiling_escape; broad 0.9883 > 0.8867 n=256 + E2) after named-phrase settle. `name_gate` still false.

docs/ROADMAP.md

@@ -102,6 +102,7 @@ The skill remains the operator surface while the encoder stays SHADOW.
 See [SPEC.md](SPEC.md) and [DESIGN.md](DESIGN.md) for the historical spine. Current operator snapshot: [STATUS.md](STATUS.md).
 
 ## Recent
+- 2026-09-24: Real inference path — `infer --model-dir` on `seed-morph78` emits `MODEL_EMBEDDING` packets as `hyperlex-structure-149m` (Spark-verified).
 - 2026-09-24: Card rename — Hub card `hyperlex-structure-149m`; eval packets set `name_gate` true only for trunk-forward eval of `seed-morph78`.
 - 2026-09-24: Danny `flip name_gate` — `seed-morph78` named **Hyperlexical** (amendment A6). Card rename, Hub, T13 still separate.
 - 2026-09-24: morph78 PROMOTE_BEST (soft_ceiling ceiling_escape; broad 0.9883 > 0.8867 n=256 + E2) after named-phrase settle. `name_gate` still false.

docs/shadow-hyperlexical.md

@@ -42,6 +42,8 @@ Danny ~2500 candidate bar: **met**. Hermes 913 / “gap to 2500” is **supersed
 
 ```bash
 PYTHONPATH=scripts/shadow python3 -m hyperlexical.infer --text rizz --offline
+# trained pin (Spark; needs torch + local trunk):
+PYTHONPATH=scripts/shadow python3 -m hyperlexical.infer --text rizz --model-dir ~/.hyperlex/models/BEST
 PYTHONPATH=scripts/shadow python3 -m hyperlexical.ingest_tap --dry-run
 PYTHONPATH=scripts/shadow python3 -m hyperlexical.export
 PYTHONPATH=scripts/shadow python3 -m hyperlexical.export --include-live

scripts/shadow/hyperlexical/infer.py

@@ -1,4 +1,4 @@
-"""SHADOW CLI. Offline stub. No torch. No Hub."""
+"""SHADOW CLI. Offline stub by default. ``--model-dir`` runs a local trained checkpoint. No Hub."""
 
 from __future__ import annotations
 
@@ -17,15 +17,34 @@ def main(argv=None) -> int:
     p.add_argument("--restricted", action="store_true")
     p.add_argument("--role-scheme", default="positional", choices=("positional", "type_slot"))
     p.add_argument("--out", default="")
+    p.add_argument("--model-dir", default="", help="local train-out dir (e.g. ~/.hyperlex/models/BEST); needs torch + trunk")
+    p.add_argument("--trunk-dir", default="", help="local ModernBERT-base snapshot (default HYPERLEX_TRUNK_DIR)")
     args = p.parse_args(argv)
     if os.environ.get("HYPERLEX_OFFLINE") == "0" and not args.offline:
         print("abort: online infer is out of U1", file=sys.stderr)
         return 2
-    packet = build_packet(
-        args.text,
-        restricted=args.restricted,
-        role_scheme=args.role_scheme,
-    )
+    if args.model_dir:
+        if args.role_scheme != "positional":
+            print("abort: trained infer supports role_scheme=positional only", file=sys.stderr)
+            return 2
+        from .infer_model import InferModelError, infer
+
+        try:
+            packet = infer(
+                args.text,
+                model_dir=args.model_dir,
+                trunk_dir=args.trunk_dir or None,
+                restricted=args.restricted,
+            )
+        except InferModelError as exc:
+            print(json.dumps({"abort": True, "error": str(exc), "brier": None}), file=sys.stderr)
+            return 2
+    else:
+        packet = build_packet(
+            args.text,
+            restricted=args.restricted,
+            role_scheme=args.role_scheme,
+        )
     text = json.dumps(packet, indent=2, sort_keys=True)
     if args.out:
         with open(args.out, "w", encoding="utf-8") as fh:

scripts/shadow/hyperlexical/infer_model.py

@@ -0,0 +1,159 @@
+"""Trained-model inference for a pinned checkpoint. Lazy torch. Local files only.
+
+Emits ``hyperlex.hyperlexical.inference.v0.1`` with ``MODEL_EMBEDDING``. Lineage
+is the classify head on the CLS state, as in train. Unbind treats each
+whitespace token as a positional atom and reads role/filler heads on its pooled
+token states, as in train val. Brier stays null; this is not a forecast.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import os
+from pathlib import Path
+
+from .layout import HIDDEN, MAX_LEN
+from .name_gate import PUBLIC_CARD_NAME, name_gate_for, pin_of
+from .packet import SCHEMA, build_packet, sha256_hex, validate_packet
+
+WEIGHT_FILE = "model.safetensors"
+
+
+class InferModelError(RuntimeError):
+    pass
+
+
+def default_trunk_dir() -> Path:
+    return Path(
+        os.environ.get("HYPERLEX_TRUNK_DIR")
+        or Path.home() / ".hyperlex" / "models" / "trunks" / "ModernBERT-base"
+    )
+
+
+def model_identity(model_dir: Path) -> tuple[str, str | None]:
+    """(model_id, model_version). Only an approved pin carries the public card name."""
+    pin = pin_of(model_dir)
+    if name_gate_for(model_dir):
+        return PUBLIC_CARD_NAME, pin
+    return Path(os.path.realpath(model_dir)).name, pin
+
+
+def model_packet(
+    text: str,
+    *,
+    model_id: str,
+    model_version: str | None,
+    vector: list[float],
+    param_count: int,
+    families: list[str],
+    family_probs: list[float],
+    atoms: list[dict],
+) -> dict:
+    """Pure packet assembly from model outputs. No torch."""
+    if len(families) != len(family_probs) or not families:
+        raise InferModelError("family probabilities do not match families")
+    best = max(range(len(families)), key=lambda i: family_probs[i])
+    vector_blob = ",".join(f"{v:.6f}" for v in vector).encode("utf-8")
+    unbind_ok = bool(atoms) and all(a["filler_pred"] == a["filler"] for a in atoms)
+    packet = {
+        "schema": SCHEMA,
+        "model_id": model_id,
+        "model_version": model_version,
+        "embed_mode": "MODEL_EMBEDDING",
+        "input_hash": sha256_hex(text),
+        "vector_hash": hashlib.sha256(vector_blob).hexdigest(),
+        "selection_proxy": "unspecified",
+        "param_count": int(param_count),
+        "surface": text,
+        "payload_ref": None,
+        "lineage_family": families[best],
+        "lineage_confidence": round(float(family_probs[best]), 6),
+        "typology": [],
+        "stage": None,
+        "role_scheme": "positional",
+        "unbind_ok": unbind_ok,
+        "unbind": {"atoms": atoms},
+        "routes_claimed": ["form", "lexical"],
+        "restricted_intent_suspected": False,
+        "brier": None,
+        "forecast_eligible": False,
+        "auto_fire": False,
+        "class": "INFERRED",
+    }
+    return validate_packet(packet)
+
+
+def infer(text: str, *, model_dir: Path, trunk_dir: Path | None = None, restricted: bool = False) -> dict:
+    if restricted or "__RESTRICTED_FIXTURE__" in (text or ""):
+        return build_packet(text, restricted=True)
+    if not (text or "").strip():
+        raise InferModelError("empty text")
+    model_dir = Path(model_dir)
+    trunk_dir = Path(trunk_dir) if trunk_dir else default_trunk_dir()
+    weight = model_dir / WEIGHT_FILE
+    if not weight.is_file():
+        raise InferModelError(f"missing {weight}")
+    if not trunk_dir.is_dir():
+        raise InferModelError(f"missing trunk snapshot {trunk_dir}")
+
+    from .eval_unbind import TrunkForwardError
+    from .eval_forward import _import_torch, _load_maps, _load_trunk, _offsets, apply_encoder_trainable
+    from .align import atom_token_index, pool_indices
+    from .save_pretrained import split_weight_tensors
+
+    try:
+        torch, nn = _import_torch()
+        from safetensors.torch import load_file
+
+        split = split_weight_tensors(load_file(str(weight), device="cpu"))
+        maps = _load_maps(model_dir, None)
+        tok, encoder = _load_trunk(trunk_dir, model_dir)
+    except (TrunkForwardError, ImportError) as exc:
+        raise InferModelError(str(exc)) from exc
+    applied = apply_encoder_trainable(encoder, split["encoder"])
+    if applied["present"] and applied["loaded"] == 0:
+        raise InferModelError("encoder tensors present but none matched trunk keys")
+    heads = {}
+    for name, n_out in (
+        ("classify", len(maps["families"])),
+        ("role_head", len(maps["role_vocab"])),
+        ("filler_head", len(maps["filler_vocab"])),
+    ):
+        state = split.get(name) or {}
+        if "weight" not in state or tuple(state["weight"].shape) != (n_out, HIDDEN):
+            raise InferModelError(f"{name} shape does not match layout")
+        head = nn.Linear(HIDDEN, n_out)
+        head.load_state_dict(state)
+        heads[name] = head.eval()
+    encoder.eval()
+
+    tokens = text.split()
+    with torch.no_grad():
+        enc = tok([text], truncation=True, max_length=MAX_LEN, return_tensors="pt")
+        states = encoder(**enc).last_hidden_state[0]
+        cls = states[0]
+        probs = torch.softmax(heads["classify"](cls), dim=-1).tolist()
+        offs = _offsets(tok, text)
+        atoms = []
+        for i, token in enumerate(tokens):
+            idxs = pool_indices(states.size(0), atom_token_index(text, token, offs))
+            h = states[idxs].mean(0)
+            role = maps["role_vocab"][int(heads["role_head"](h).argmax())]
+            filler = maps["filler_vocab"][int(heads["filler_head"](h).argmax())]
+            atoms.append(
+                {"role": f"pos_{i}", "role_pred": role, "filler": token.lower(), "filler_pred": filler}
+            )
+    param_count = sum(p.numel() for p in encoder.parameters()) + sum(
+        p.numel() for h in heads.values() for p in h.parameters()
+    )
+    model_id, model_version = model_identity(model_dir)
+    return model_packet(
+        text,
+        model_id=model_id,
+        model_version=model_version,
+        vector=cls.tolist(),
+        param_count=param_count,
+        families=list(maps["families"]),
+        family_probs=probs,
+        atoms=atoms,
+    )

specs/007-hyperlexical-model/HYPERLEXICAL-PRODUCT-PLAN.md

@@ -80,7 +80,7 @@ Item 4 is done. Hub card name is `hyperlex-structure-149m`; local train-out path
 - [ ] operator reviews this reconciled plan
 - [x] Danny explicit `flip name_gate` (2026-09-24)
 - [x] card rename → `hyperlex-structure-149m` + eval packet `name_gate` true only for trunk-forward eval of `seed-morph78`
-- [ ] real `seed-morph78` inference path (trunk-forward infer → inference packet)
+- [x] real `seed-morph78` inference path (`infer --model-dir`; receipt `receipts/20260924-infer-model-morph78.md`)
 - [ ] finalize public model card from pinned receipts
 - [ ] separately authorize Hub upload, or explicitly close the cycle with no Hub
 - [ ] optional T13 in a separate PR with tests
@@ -97,7 +97,7 @@ Q1 status is not a substitute for Spec 007 product completion, and Spec 007 prog
 
 ## Default recommendation
 
-**HOLD morph78.** NAME is done. Resolve the inference path and PUBLISH before authorizing another training climb.
+**HOLD morph78.** NAME is done. Resolve PUBLISH (and the two inference findings) before authorizing another training climb.
 
 ## Explicit non-goals
 
@@ -119,7 +119,7 @@ Q1 status is not a substitute for Spec 007 product completion, and Spec 007 prog
 - [ ] operator product review
 - [x] Danny `name_gate` decision — yes for `seed-morph78`
 - [x] card rename
-- [ ] real inference path
+- [x] real inference path
 - [ ] Hub decision: upload authorized, or explicit no-Hub this cycle
 - [ ] optional T13 decision
 

specs/007-hyperlexical-model/NEXT_MOVES_007.md

@@ -7,6 +7,7 @@
 - morph78 named phrases settled · force/hard **236/277** · soft_ceiling train complete.
 - **PROMOTE_BEST** (ceiling_escape): broad **0.9883** > PRIOR morph65 **0.8867** n=256 + E2 PASS.
 - Card rename: Hub card `hyperlex-structure-149m`; eval packets set `name_gate` true only for a trunk-forward eval of `seed-morph78`.
+- Real inference: `hyperlexical.infer --model-dir ~/.hyperlex/models/BEST` → `MODEL_EMBEDDING` packet (`hyperlex-structure-149m`), verified on Spark.
 - Danny **`flip name_gate`** → `seed-morph78` may be called **Hyperlexical** (receipt `specs/007-hyperlexical-model/receipts/20260924-name-gate-yes-morph78.md`).
 
 ## Gate
@@ -15,9 +16,10 @@ soft_ceiling **spent** for this climb (morph78 now PRIOR/BEST). Further climbs n
 
 ## Next (each a separate operator sentence)
 
-1. Real `seed-morph78` inference path (trunk-forward infer emitting the inference packet) — in progress.
-2. Hub upload — still not authorized.
-3. Optional T13 (promote `scripts/shadow/hyperlexical/` into `src/hyperlex/`).
-4. HOLD morph78; no new climb without a new card.
+1. Hub upload — still not authorized (decide publish vs explicit local-only).
+2. Optional T13 (promote `scripts/shadow/hyperlexical/` into `src/hyperlex/`).
+3. HOLD morph78; no new climb without a new card.
+
+Open findings (receipt `receipts/20260924-infer-model-morph78.md`): classify overconfident on val-settled phrases; trunk+heads = 150.55M vs A1 150M (trunk 149.0M).
 
 Qwen stays stopped unless re-enabled.

specs/007-hyperlexical-model/receipts/20260924-infer-model-morph78.md

@@ -0,0 +1,28 @@
+# Trained inference path — `seed-morph78` (2026-09-24)
+
+**Scope:** first real (non-stub) inference for Hyperlexical. `hyperlexical.infer --model-dir` loads a local trained checkpoint and emits `hyperlex.hyperlexical.inference.v0.1` with `MODEL_EMBEDDING`. No Hub, no T13, no training. Default CLI stays the offline stub.
+
+**Run:** Spark, container `lmsysorg/sglang:dev-qwen38-27b-dflash2`, branch code staged in `~/hlx-infer-test` (the Spark `~/Hyperlex` checkout was not touched). Model `~/.hyperlex/models/BEST` → `seed-morph78`. Outputs in `infer-model-morph78-20260924/`.
+
+## OBSERVED
+
+| Input | model_id / version | lineage (conf) | unbind_ok |
+|---|---|---|---|
+| `rizz` | `hyperlex-structure-149m` / `seed-morph78` | brainrot-aura (0.936) | true |
+| `no cap fr` | same | brainrot-aura (0.880) | true |
+| `touch grass` | same | gaming-meta (0.978) | true |
+| `bet that up` | same | none (1.000) | true |
+| `have fun staying poor` | same | none (1.000) | true |
+| `rizz` on morph65 dir | `hyperlex-encoder-modernbert-base-seed-morph65` / `seed-morph65` | brainrot-aura (0.984) | true |
+
+Trunk-forward E2 on `BEST` with this branch: `e2_pass: true`, unbind_exact 1.0, n_unbind_eval 24, **`name_gate: true`** (card-rename PR #104 behaviour confirmed on the real pin).
+
+## Findings (recorded, not acted on)
+
+1. **Classify overconfidence on val-settled phrases.** `bet that up` and `have fun staying poor` were settled OBSERVED `split=val` for unbind; the classify head returns `none` at confidence ≈1.0 instead of their card lineages (kinship-address, crypto-degen). Lineage confidence is not calibrated; treat it as INFERRED only.
+2. **Parameter count vs A1.** Inference counts **150,546,889** parameters: trunk 149,014,272 + heads 1,532,617 (filler vocab 1,972). A1 says "T1 ≤ 150M". The trunk is within the ceiling; trunk + heads exceeds it by ~0.55M. Whether A1 counts heads is an operator reading; the card name `hyperlex-structure-149m` matches the trunk.
+
+## NOT_COMPUTABLE
+
+- Calibrated lineage accuracy of the inference path (no held-out classify receipt for this pin; E1 still NOT_COMPUTABLE).
+- `type_slot` inference (CLI refuses it for `--model-dir` in this cut).

specs/007-hyperlexical-model/receipts/infer-model-morph78-20260924/e2-best.json

@@ -0,0 +1,32 @@
+{
+  "brier": null,
+  "device": "cuda",
+  "e2_pass": true,
+  "encoder_trainable_loaded": 48,
+  "encoder_trainable_present": 48,
+  "forecast_eligible": false,
+  "heads_loaded": true,
+  "model_dir": "/home/morpheus/.hyperlex/models/BEST",
+  "model_id": "hyperlex-encoder-modernbert-base-seed",
+  "model_swap": 1.0,
+  "n_test": 12,
+  "n_unbind_eval": 24,
+  "name_gate": true,
+  "note": "Trunk-forward unbind_exact + token/slot F1 from /home/morpheus/.hyperlex/models/BEST/model.safetensors vs 004 probe_swap_min. Stub/digest leaves F1 null (no civilian filler lists). name_gate is true only for an approved pin (A6). Applied 48 saved encoder trainable tensors.",
+  "probe_positional_swap": 1.0,
+  "probe_schema": "abraxas.recoverable_structure.probe.v0.1",
+  "probe_swap_min": 0.5,
+  "probe_type_slot_swap": 0.5,
+  "schema": "hyperlex.hyperlexical.eval_unbind.v0.1",
+  "stub_swap": 0.0,
+  "trunk": "answerdotai/ModernBERT-base",
+  "trunk_dir": "/home/morpheus/.hyperlex/models/trunks/ModernBERT-base",
+  "trunk_forward": true,
+  "trunk_loaded": true,
+  "unbind_exact": 1.0,
+  "unbind_slot_f1": 1.0,
+  "unbind_token_f1": 1.0,
+  "unbind_token_precision": 1.0,
+  "unbind_token_recall": 1.0,
+  "weight_file": "model.safetensors"
+}