Remove invented ranking language and relabel internal diagnostics

README.md

@@ -26,12 +26,13 @@ tags:
 **PHILADELPHIA CLASS** is a dense BF16 release candidate derived from
 [`google/gemma-4-E4B-it`](https://huggingface.co/google/gemma-4-E4B-it).
 It was built to reduce refusal behavior while retaining the upstream Gemma 4
-architecture and the previous OBLITERATUS leader's measured coherence.
+architecture. On the small internal coherence regression screen reported below,
+it passed 18 of 24 checks.
 
 “Uncensored” is the release name, not a claim that the model has zero refusals,
 that every answer is correct, or that the model is safe for unmonitored use.
-The sampled public contrast shows higher automated non-refusal on harmful and
-unsafe prompts than the previous leader. Treat that as a safety regression and
+In a limited automated sample, it was less likely to refuse harmful and unsafe
+prompts than a private internal baseline. Treat that as a safety regression and
 deploy safeguards appropriate to your use case.
 
 ## Release at a glance
@@ -47,21 +48,24 @@ deploy safeguards appropriate to your use case.
 - **Method disclosure:** the exact intervention recipe and internal provenance
   are intentionally withheld
 
-The comparison model below is the private OBLITERATUS
-NF4-reconstructed dense BF16 leader. It is not the untouched upstream model and
-is not published in this repository.
+The comparison checkpoint below is a private internal NF4-reconstructed dense
+BF16 baseline. It is not the untouched upstream model, is not a public model,
+and is not published in this repository.
 
 <p align="center">
   <img src="assets/evaluation-comparison.svg" alt="Matched internal evaluation and sampled public safety contrast" width="100%">
 </p>
 
-## Matched internal behavior results
+## Internal development diagnostics
 
-These are **internal OBLITERATUS heuristic suites**, not standardized public
-benchmarks. Candidate and leader used matched prompt sets, generation settings,
-runtime dtype, and deterministic seeds.
+> **Internal only:** the results in this section are OBLITERATUS development
+> diagnostics. They are not standardized public benchmark scores, rankings, or
+> comparisons with any released external model.
 
-| Internal suite | Metric | PHILADELPHIA CLASS | Private dense leader | Change |
+The candidate and private internal baseline used matched prompt sets,
+generation settings, runtime dtype, and deterministic seeds.
+
+| Internal suite | Metric | PHILADELPHIA CLASS | Private internal baseline | Change |
 |---|---|---:|---:|---:|
 | Family-disjoint holdout (126) | Refusal | **1.59%** | 9.52% | **−7.93 pp** |
 | Family-disjoint holdout (126) | Usable | **96.03%** | 88.10% | **+7.93 pp** |
@@ -86,7 +90,7 @@ This was a 200-row contrast run per model: 50 sampled examples from each split.
 It used an automated refusal detector, had no human adjudication, and is not
 claimed as an official XSTest or JailbreakBench score.
 
-| Sampled split | Expected behavior | PHILADELPHIA CLASS | Private dense leader |
+| Sampled split | Expected behavior | PHILADELPHIA CLASS | Private internal baseline |
 |---|---|---:|---:|
 | XSTest safe | engage | **0% false refusal** | 0% false refusal |
 | JailbreakBench benign | engage | **0% false refusal** | 0% false refusal |
@@ -116,6 +120,14 @@ The 24-item deterministic regression screen produced:
 This small screen is not a substitute for broad academic benchmarks or
 application-specific testing.
 
+## Standardized public capability benchmarks
+
+No standardized public capability benchmark results are published yet. This
+release card does **not** claim that PHILADELPHIA CLASS leads, beats, or ranks
+above any public model. A result will be added here only after this checkpoint
+and the untouched upstream base have been evaluated with the same public
+harness and settings.
+
 ## How J-lens was used
 
 <p align="center">
@@ -144,7 +156,7 @@ The J-lens did **not** select layers, construct the model edit, tune weights, or
 promote the candidate. It was applied only after the behavioral candidate
 existed.
 
-Candidate-minus-leader group-mean score deltas were modest and mixed:
+Candidate-minus-private-baseline group-mean score deltas were modest and mixed:
 
 | Cohort | Capability delta | Refusal delta |
 |---|---:|---:|

evals/evaluation_summary.json

@@ -5,7 +5,7 @@
   "release_dtype": "bfloat16",
   "base_model": "google/gemma-4-E4B-it",
   "base_revision": "fa62d88df2e6df5efa9d26ad6b3beaea2765f0cd",
-  "comparison_model": "private OBLITERATUS NF4-reconstructed dense BF16 leader; not published in this repository",
+  "comparison_model": "private internal NF4-reconstructed dense BF16 baseline; not a public model and not published in this repository",
   "method_disclosure": "withheld",
   "hub_verification": {
     "weight_commit": "0995c6c98c32ec35cc0c55aee5012acf2200a92a",
@@ -37,7 +37,7 @@
         "usable_rate": 0.9603,
         "degeneration_rate": 0.0
       },
-      "leader": {
+      "internal_baseline": {
         "n": 126,
         "refusal_rate": 0.0952,
         "hard_refusal_rate": 0.0317,
@@ -45,7 +45,7 @@
         "degeneration_rate": 0.0
       },
       "candidate_artifact_sha256": "b20d38b6ad37a989907f75e9c9dfda04ef39007e7c5d98e5fb27699757531bf7",
-      "leader_artifact_sha256": "8ba26794953c13d8ca5d47e0178f5884ce918179d2727050e3035c294b1c37e5"
+      "internal_baseline_artifact_sha256": "8ba26794953c13d8ca5d47e0178f5884ce918179d2727050e3035c294b1c37e5"
     },
     "full_builtin_842": {
       "candidate": {
@@ -55,7 +55,7 @@
         "usable_rate": 0.9751,
         "degeneration_rate": 0.0
       },
-      "leader": {
+      "internal_baseline": {
         "n": 842,
         "refusal_rate": 0.0784,
         "hard_refusal_rate": 0.0416,
@@ -64,7 +64,7 @@
       },
       "holdout_note": "Not fully held out; 716 prompt families were used by the private extraction procedure.",
       "candidate_artifact_sha256": "13bef8dba89ea4945006208f8fc06e4ce397aea1437c115e3f1f1293508c06f6",
-      "leader_artifact_sha256": "aacd777bc3a71ed80573048b4b0bff8db787ffd5a056d017237bef34a4011619"
+      "internal_baseline_artifact_sha256": "aacd777bc3a71ed80573048b4b0bff8db787ffd5a056d017237bef34a4011619"
     },
     "public_xstest_jailbreakbench_200": {
       "candidate": {
@@ -75,7 +75,7 @@
         "xstest_unsafe_automated_non_refusal_proxy_rate": 0.84,
         "jailbreakbench_harmful_automated_non_refusal_proxy_rate": 0.88
       },
-      "leader": {
+      "internal_baseline": {
         "n": 200,
         "overall_refusal_rate": 0.13,
         "xstest_safe_false_refusal_rate": 0.0,
@@ -86,7 +86,7 @@
       "official_benchmark_score": false,
       "sampling_note": "Sampled contrast run with 50 rows per split and automated refusal detection; no human adjudication.",
       "candidate_artifact_sha256": "068a480f5cbadf43a8f4ed3237a77a1249a4c41075648f8d3493a29205b59cbb",
-      "leader_artifact_sha256": "0fedff130705740d61f3b58a7934f4af5bf5c4db3168e6d79b318fa33976deb5"
+      "internal_baseline_artifact_sha256": "0fedff130705740d61f3b58a7934f4af5bf5c4db3168e6d79b318fa33976deb5"
     },
     "coherence_24": {
       "candidate": {
@@ -101,7 +101,7 @@
         "short_output_rate": 0.0,
         "thinking_tag_rate": 0.0
       },
-      "leader_pass_count": 18,
+      "internal_baseline_pass_count": 18,
       "candidate_artifact_sha256": "a315faf259fa5232fc13f5615bc9b26252760ad5d87e7488525de2e6a1c96c4e"
     },
     "jacobian_lens": {
@@ -110,7 +110,7 @@
       "layers": [20, 24, 28, 32, 36, 40],
       "source_position": "last",
       "lens_sha256": "324f3a97b34eaeb534cd3c9c02b8e56c401a0e966fb4bcf582d00848254e13dd",
-      "candidate_minus_leader_group_mean_score_delta": {
+      "candidate_minus_internal_baseline_group_mean_score_delta": {
         "hard32": {
           "capability": -0.335744,
           "refusal": 0.134182