Prism 2.1 Scoring

.github/workflows/images.yml

@@ -21,6 +21,12 @@ on:
       - "Cargo.lock"
       - ".github/workflows/images.yml"
   workflow_dispatch:
+    inputs:
+      prism_pod_only:
+        description: Build the CUDA 13 + Transformer Engine Prism pod
+        required: false
+        type: boolean
+        default: false
 
 permissions:
   contents: read
@@ -36,6 +42,7 @@ concurrency:
 jobs:
   build-push-digests:
     name: build · push · record digests
+    if: ${{ github.event_name != 'workflow_dispatch' || inputs.prism_pod_only != true }}
     runs-on: ubuntu-latest
     timeout-minutes: 90
     strategy:
@@ -135,6 +142,51 @@ jobs:
           path: digests-out/${{ matrix.image_suffix }}.json
           if-no-files-found: error
 
+  build-prism-pod:
+    name: build · push Prism CUDA pod
+    if: ${{ github.event_name == 'workflow_dispatch' && inputs.prism_pod_only }}
+    runs-on: ubuntu-latest
+    timeout-minutes: 120
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Free runner disk for the CUDA base
+        run: |
+          sudo rm -rf /usr/local/lib/android /usr/share/dotnet /opt/ghc /usr/local/.ghcup
+          docker system prune --all --force
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Log in to GHCR
+        uses: docker/login-action@v3
+        with:
+          registry: ghcr.io
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Build and push Prism pod
+        id: build
+        uses: docker/build-push-action@v6
+        with:
+          context: deploy/prism-pod
+          file: deploy/prism-pod/Dockerfile
+          platforms: linux/amd64
+          push: true
+          provenance: false
+          sbom: false
+          tags: |
+            ghcr.io/baseintelligence/base/prism-pod:${{ github.sha }}
+            ghcr.io/baseintelligence/base/prism-pod:v10-cuda13-te
+
+      - name: Report immutable image reference
+        env:
+          DIGEST: ${{ steps.build.outputs.digest }}
+        run: |
+          test -n "$DIGEST"
+          echo "ghcr.io/baseintelligence/base/prism-pod@$DIGEST"
+
   merge-digests:
     name: merge digest manifest
     needs: build-push-digests
@@ -329,7 +381,7 @@ jobs:
               -H "X-GitHub-Api-Version: 2022-11-28" \
               "https://api.github.com/orgs/${OWNER}/packages/container/${enc}/visibility" \
               -d '{"visibility":"public"}' || true)
-            echo "pkg=${pkg} http=${code} body=$(head -c 200 /tmp/vis-${enc}.json || true)"
+            echo "pkg=${pkg} http=${code} body=$(head -c 200 "/tmp/vis-${enc}.json" || true)"
             if [ "$code" = "204" ] || [ "$code" = "200" ]; then
               ok=$((ok + 1))
             fi

crates/challenge-agentic/src/prompts.rs

@@ -1,7 +1,7 @@
 //! System prompt for the agentic anti-cheat loop.
 
 /// Prompt version stamp (audit).
-pub const AGENTIC_PROMPT_VERSION: &str = "agentic_v4";
+pub const AGENTIC_PROMPT_VERSION: &str = "agentic_v5";
 
 /// Prism-specific anti-cheat rules (metrics / copy / hooks / non-causal leak).
 ///
@@ -16,6 +16,7 @@ pub const PRISM_DOMAIN_RULES: &str = r"Prism domain:
 - Causal LM: next-token CE must not see future/label tokens. Dense TokenMix/t_mix/Linear over time after transpose(1,2) WITHOUT triu/tril/is_causal/attn_mask = non_causal_label_leak.
 - Cheat if: near-identical AST/byte copy of a corpus architecture/touched surface (AST similarity_bps >= 9500; identical patch_sha256 / patch text); hardcoded METRICS_JSON; bpb impossible vs tokens_seen/wall_clock; recipe-v1 bpb < 1.0; eval peeking; network/exfil in the delta; eval-set leakage patterns in the delta; non_causal_label_leak as above.
 - NEVER treat standard modern LM components as plagiarism: RMSNorm, LayerNorm, RoPE/ALiBi, SwiGLU/GeGLU, gated residual, parallel residual, GQA/MQA, Pre-Norm/Post-Norm. Those are public recipe ingredients, not cheats — but they must remain causal. New models/ packages under AutoModel are allowed novelty.
+- Tokenizer is SUBMITTED, not imposed (tree `tokenizer/` files or a `build_tokenizer(ctx)` hook beside build_model). Bringing your own tokenizer is allowed and G1 already scores tokenizer-neutral bits/byte. Read metrics.json -> tokenizer.card (probe_tokens_per_byte, vocab_multiword_frac, vocab_max_token_bytes, probe_roundtrip_ok, flags) plus any build_tokenizer/vocab source in the delta. Cheat (tokenizer_gaming) when the tokenizer is engineered for the metric instead of language modeling: multi-word / answer-phrase single tokens (vocab_multiword_frac flag), vocab stuffed with eval-looking strings, decode() that rewrites or injects output text, or extreme compression built to shortcut token-denominated probes. A merely WEAK tokenizer (byte-level, small vocab, poor merges) is NOT a cheat — judge intent to game, not quality; card flags alone without corroborating source evidence → suspicious at most.
 - suspicious: only for strong unique structural overlap with a champion (AST >= 8500) or inconsistent metrics without a slam-dunk forge. Below AST 8500 with no other cheat signal → clean.
 - Quality/coherence of the model is NOT your job — only anti-cheat.";
 

crates/lium-rent-pool/src/lib.rs

@@ -13,6 +13,12 @@ pub const RECOVERY_WINDOW_MS: u64 = 6 * 60 * 60 * 1000;
 /// True when a failed row should re-enter the rent queue (429 within window).
 #[must_use]
 pub fn should_recover(error_detail: &str, updated_at_ms: u64, now_ms: u64) -> bool {
+    let l = error_detail.to_ascii_lowercase();
+    // HuggingFace / FineWeb 429s are not Lium rent 429s — re-renting an
+    // 8×5090 pod does not fix a dataset CDN throttle.
+    if l.contains("huggingface") || l.contains("fineweb") || l.contains("\"stage\": \"dataset\"") {
+        return false;
+    }
     is_rate_limited(error_detail) && now_ms.saturating_sub(updated_at_ms) <= RECOVERY_WINDOW_MS
 }
 
@@ -70,5 +76,10 @@ mod tests {
         ));
         assert!(should_recover("provision: 429 rate limit", 100, 100 + 1));
         assert!(!should_recover("provision: 429", 0, RECOVERY_WINDOW_MS + 1));
+        assert!(!should_recover(
+            "measure: exec: HTTP Error 429: huggingface.co/datasets/HuggingFaceFW/fineweb-edu",
+            100,
+            100 + 1
+        ));
     }
 }

crates/prism-artifacts/src/receive.rs

@@ -29,14 +29,14 @@ pub const CHECKPOINT_OVERHEAD_NUM: u64 = 3;
 /// Denominator of the 1.5× overhead factor (with [`CHECKPOINT_OVERHEAD_NUM`] → 3/2).
 pub const CHECKPOINT_OVERHEAD_DEN: u64 = 2;
 
-/// Recipe parameter cap (`prism_recipe::MAX_PARAMS` = 350M). Used as the
+/// Recipe parameter cap (`prism_recipe::MAX_PARAMS` = 1B). Used as the
 /// absolute HTTP body ceiling and as the harvest fallback when measured
 /// `n_params` is missing (older harness). Prefer measured `n_params`.
-pub const RECIPE_MAX_PARAMS: u64 = 350_000_000;
+pub const RECIPE_MAX_PARAMS: u64 = 1_000_000_000;
 
 /// Absolute max packed/upload bytes = [`RECIPE_MAX_PARAMS`] × FP32 × 2 × 1.5.
 /// Per-receive budgets are tighter when measured `n_params` is known.
-pub const MAX_CHECKPOINT_BYTES: usize = 4_200_000_000; // 350_000_000 * 12
+pub const MAX_CHECKPOINT_BYTES: usize = 12_000_000_000; // 1_000_000_000 * 12
 
 const _: () = assert!(RECIPE_MAX_PARAMS * 12 == MAX_CHECKPOINT_BYTES as u64);
 

crates/prism-automodel/src/intake.rs

@@ -24,6 +24,15 @@ pub const MEMBER_BASE: &str = "automodel.base";
 pub const MEMBER_PATCH: &str = "automodel.patch";
 /// Optional entry / recipe knobs.
 pub const MEMBER_TOML: &str = "prism.toml";
+/// Optional miner dependency manifest: pip `requirements.txt` installed on
+/// the pod (network-on install phase) before the netns-isolated train/eval.
+/// Delivered to the pod when the miner's patch adds it to the applied tree
+/// (it becomes a touched file, so slim delivery keeps it); the harness
+/// installs any manifest it finds in the tree (see `prismlib/deps.py`).
+pub const MEMBER_REQUIREMENTS: &str = "requirements.txt";
+/// Optional miner dependency manifest: `pyproject.toml` (PEP 621) installed
+/// with `pip install .` on the pod's install phase.
+pub const MEMBER_PYPROJECT: &str = "pyproject.toml";
 
 /// Persisted under the packed tree for `GET …/diff`.
 pub const META_BASE: &str = ".prism/automodel.base";
@@ -149,22 +158,16 @@ pub fn extract_automodel_zip(zip_bytes: &[u8]) -> Result<AutomodelMembers, Intak
         .map_err(|_| IntakeError::Invalid(format!("non-utf8: {MEMBER_BASE}")))?
         .trim()
         .to_owned();
-    if pin_id.is_empty() {
-        return Err(IntakeError::Invalid(format!("empty {MEMBER_BASE}")));
-    }
-    if pin_id.chars().any(char::is_whitespace) {
+    if pin_id.is_empty() || pin_id.chars().any(char::is_whitespace) {
         return Err(IntakeError::Invalid(format!(
-            "{MEMBER_BASE} must be a single-line pin id without whitespace"
+            "{MEMBER_BASE} must be a non-empty single-line pin id without whitespace"
         )));
     }
-    let prism_toml = match files.get(MEMBER_TOML) {
-        Some(b) => Some(
-            std::str::from_utf8(b)
-                .map_err(|_| IntakeError::Invalid(format!("non-utf8: {MEMBER_TOML}")))?
-                .to_owned(),
-        ),
-        None => None,
-    };
+    let prism_toml = files
+        .get(MEMBER_TOML)
+        .map(|b| std::str::from_utf8(b).map(str::to_owned))
+        .transpose()
+        .map_err(|_| IntakeError::Invalid(format!("non-utf8: {MEMBER_TOML}")))?;
     Ok(AutomodelMembers {
         pin_id,
         patch,
@@ -309,6 +312,8 @@ fn slim_delivery_files(mat: &MaterializedAutomodel) -> BTreeMap<String, Vec<u8>>
         META_DIFFSTAT,
         mat.entry.as_str(),
         MEMBER_TOML,
+        MEMBER_REQUIREMENTS,
+        MEMBER_PYPROJECT,
     ]
     .into_iter()
     .chain(mat.diffstat.files.iter().map(|entry| entry.path.as_str()))

crates/prism-automodel/src/lib.rs

@@ -39,8 +39,8 @@ pub use intake::{
     expand_tree_blob_for_pod, extract_automodel_zip, fixture_automodel_zip, intake_automodel_zip,
     materialize, pack_tree_blob, pin_checkout_dir, resolve_pin, submission_id_for_patch,
     zip_is_automodel_layout, AutomodelMembers, ExpandedAutomodel, IntakeError,
-    MaterializedAutomodel, ALLOW_LEGACY_ENV, MEMBER_BASE, MEMBER_PATCH, MEMBER_TOML, META_BASE,
-    META_DIFFSTAT, META_PATCH,
+    MaterializedAutomodel, ALLOW_LEGACY_ENV, MEMBER_BASE, MEMBER_PATCH, MEMBER_PYPROJECT,
+    MEMBER_REQUIREMENTS, MEMBER_TOML, META_BASE, META_DIFFSTAT, META_PATCH,
 };
 pub use pin::{
     fixture_happy_patch_path, fixture_pin_dir, tree_content_sha256, verify_pin_tree, AutomodelPin,

crates/prism-budget/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "prism-budget"
+description = "PRISM dual-cap budget facts + lexicographic budget gates (split out of prism-pipeline::composite for the per-crate LOC cap; re-exported there)"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+repository.workspace = true
+rust-version.workspace = true
+publish = false
+
+[dependencies]
+serde = { version = "1", features = ["derive"] }
+
+[dev-dependencies]
+serde_json = "1"
+
+[lints]
+workspace = true