manifest 168 -> 169: L-019 retracted (rule 6), L-162/L-169 consensus-frame scope split (mean+rotation vs reference-member)

loss_manifest.json

@@ -660,7 +660,7 @@
   {
    "id": "L-019",
    "name": "Expert Soup composite (InfoNCE + MSE + BCE + Procrustes + CV + spread)",
-   "role": "PRIMARY",
+   "role": "RETRACTED",
    "primitive": "CE",
    "accum": "A7",
    "substrate": "mixed",
@@ -675,17 +675,21 @@
     "L": 0
    },
    "s_raw": 6,
-   "rating": 5,
+   "rating": 1,
    "rating_hi": null,
    "rules_fired": [
+    "retraction=1",
     "single-seed<=6"
    ],
    "cite": [
-    "canon/constellation_forms.md:35-39"
+    "canon/constellation_forms.md:35-46 (retirement annotation)",
+    "repos/bertenstein.md",
+    "canon/shared_pass_contrastive_leak.md",
+    "inventory/SCALE_MANIFEST.md DEMOTED 2026-07-31"
    ],
-   "headline": "mAP .84 (data-limited), 1000/1000 positive volumes, eff-dim 76.9 for COCO-80",
+   "headline": "RETRACTED 2026-07-31 with its source system: the Form-2 / Bertenstein numbers (mAP .84, 1000/1000 positive volumes, eff-dim 76.9) were never independently audited, and the system carrying them was falsified — R@1 was a shared-forward-pass leak, Procrustes cos_after sat at/below a no-relationship null, CV .20 at d=1024 was collapse",
    "gauge": "mAP",
-   "note": "Six terms; no ablation isolates any of them. Included for completeness, not as a template.",
+   "note": "Six terms; no ablation isolates any of them. RETRACTION NOTE 2026-08-02: retired with repos/bertenstein.md — rule 6 floor-breaker. What survives is the FORM, not this composite: whitened-Procrustes pre-alignment with held-out fit discipline, and per-expert projectors into a REFERENCE MEMBER's frame. Its validated descendant is L-162 (consensus NCE+MSE), which ablates cleanly and is measured 2/2 on two modalities.",
    "formula": "InfoNCE + MSE + BCE + Procrustes + CV + spread (6-term, never ablated)",
    "impl": "external (Expert Soup, Form 2)"
   },
@@ -5171,8 +5175,8 @@
    "flags": [],
    "family": "dist",
    "sub": {
-    "R": 2,
-    "P": 2,
+    "R": 3,
+    "P": 1,
     "D": 2,
     "C": 2,
     "I": 2,
@@ -5184,11 +5188,12 @@
    "rules_fired": [],
    "cite": [
     "tools/dist_runs/",
-    "canon/clip_mini_distillation.md"
+    "canon/clip_mini_distillation.md",
+    "canon/cc12m_pod_matrix.md"
    ],
    "headline": "zs-C10 .2693/.2657 (matrix best) and agree_cos .714/.718 - but C100 .062 LOSES to infonce (+3.2pt) and R@1 .025 is 2.2x below, 2/2 seeds",
    "gauge": "zs-C10/C100 + COCO R@1/R@5 + agree_cos + cv16/erank",
-   "note": "CLIP-KD's champion, confirmed ONLY on the coarse gauge: mimicry buys text-head alignment, not separability. The gauge-split half that amends L-038. erank 122 vs infonce 184.",
+   "note": "CLIP-KD's champion, confirmed ONLY on the coarse gauge: mimicry buys text-head alignment, not separability. The gauge-split half that amends L-038. erank 122 vs infonce 184. | CC12M 2/2 (88k steps, 2.05 ep of 10.97M): zs .4429/.4826 r1 .0760/.0716 agree .7345/.7379 - the COCO gauge-split claim COLLAPSED at scale (infonce sweeps all task gauges); mimicry keeps only the agreement set-point. P downgraded accordingly, R up (2nd substrate).",
    "formula": "mean((z_s - z_t)^2), targets L2-normalized",
    "impl": "tools/dist_bed.py loss_of feature_mse (loss_forms.prim_sq)"
   },
@@ -5202,24 +5207,25 @@
    "flags": [],
    "family": "dist",
    "sub": {
-    "R": 2,
+    "R": 3,
     "P": 3,
     "D": 2,
     "C": 1,
     "I": 2,
     "L": 0
    },
-   "s_raw": 10,
-   "rating": 8,
+   "s_raw": 11,
+   "rating": 9,
    "rating_hi": null,
    "rules_fired": [],
    "cite": [
     "tools/dist_runs/",
-    "canon/clip_mini_distillation.md"
+    "canon/clip_mini_distillation.md",
+    "canon/cc12m_pod_matrix.md"
    ],
    "headline": "R@1 .0580/.0544 = 2.2x feature_mse 2/2; zs-C100 .0939/.0847 wins 2/2 (+3.2/+2.2pt); erank 184; cv16 lands the S^15 band NATIVELY",
    "gauge": "zs-C10/C100 + COCO R@1/R@5 + agree_cos + cv16/erank",
-   "note": "THE SHIP ARM. The A7 grid in its sanctioned place (output-space distillation, no address paths). C=1: needs in-batch negatives at batch 256; single batch size tested (P6 confound disclosed).",
+   "note": "THE SHIP ARM. The A7 grid in its sanctioned place (output-space distillation, no address paths). C=1: needs in-batch negatives at batch 256; single batch size tested (P6 confound disclosed). | CC12M 2/2: zs .5208/.4979 r1 .1252/.1178 erank 188 - sweeps every raw task gauge; viability 42.0/41.5% of teacher (COCO 21.2%). The raw champion at scale; surpassed only by consensus_nce_mse+rotation.",
    "formula": "0.5*[CE(sim/t, diag) + CE(sim.T/t, diag)], sim = z_s @ z_t.T",
    "impl": "tools/dist_bed.py loss_of infonce (loss_forms.a7_grid_infonce)"
   },
@@ -5233,24 +5239,25 @@
    "flags": [],
    "family": "dist",
    "sub": {
-    "R": 2,
+    "R": 3,
     "P": 1,
     "D": 2,
     "C": 2,
     "I": 1,
     "L": 0
    },
-   "s_raw": 8,
-   "rating": 7,
+   "s_raw": 9,
+   "rating": 8,
    "rating_hi": null,
    "rules_fired": [],
    "cite": [
     "tools/dist_runs/",
-    "canon/clip_mini_distillation.md"
+    "canon/clip_mini_distillation.md",
+    "canon/cc12m_pod_matrix.md"
    ],
    "headline": "zs SPLIT (-4.6pt s0 / +2.3pt s1 vs infonce) but LOSES retrieval 2/2 (R@1 .043/.040 vs .058/.054): coupling wins retrieval at batch 256",
    "gauge": "zs-C10/C100 + COCO R@1/R@5 + agree_cos + cv16/erank",
-   "note": "The zero-coupling control that extends the coupling dose-response to the contrastive family. Per-pair BCE, no partition function = A0 over the grid, NOT A7. I=1: single batch size is the standing confound.",
+   "note": "The zero-coupling control that extends the coupling dose-response to the contrastive family. Per-pair BCE, no partition function = A0 over the grid, NOT A7. I=1: single batch size is the standing confound. | CC12M 2/2: zs .4534/.4573 r1 .0904/.0914 - coupling verdict holds at scale (infonce wins zs AND retrieval 2/2); COCO zs seed-split resolved as small-data noise.",
    "formula": "mean(BCE(sim*t' + b, I)) per pair, no batch denominator",
    "impl": "tools/loss_forms.py siglip_pairwise_loss"
   },
@@ -5264,25 +5271,26 @@
    "flags": [],
    "family": "dist",
    "sub": {
-    "R": 2,
+    "R": 3,
     "P": 2,
     "D": 2,
     "C": 2,
     "I": 1,
     "L": 0
    },
-   "s_raw": 9,
+   "s_raw": 10,
    "rating": 8,
    "rating_hi": null,
    "rules_fired": [],
    "cite": [
     "tools/dist_runs/",
     "canon/clip_mini_distillation.md",
-    "tools/dist_runs/frame_check_s1.json"
+    "tools/dist_runs/frame_check_s1.json",
+    "canon/cc12m_pod_matrix.md"
    ],
    "headline": "FRAME AMBIGUITY PROVEN 2/2: agree_cos -0.021/-0.002 (orthogonal to teacher frame, zs floor .109/.111, erank 175 healthy - the tell) yet ONE fitted rotation recovers agree .501/.497 and zs .240/.218 - geometry learned, gauge free, both seeds",
    "gauge": "zs-C10/C100 + COCO R@1/R@5 + agree_cos + cv16/erank",
-   "note": "Fills the KL-A7 cell. No cross-space term = global gauge freedom; erank stayed 175.5 at 'floor' (the tell). Usable ONLY with a Procrustes post-fit (2,500 pairs). Post-rotation agreement matches across seeds to 0.004 - the mechanism, not a lottery.",
+   "note": "Fills the KL-A7 cell. No cross-space term = global gauge freedom; erank stayed 175.5 at 'floor' (the tell). Usable ONLY with a Procrustes post-fit (2,500 pairs). Post-rotation agreement matches across seeds to 0.004 - the mechanism, not a lottery. | CC12M 2/2: raw floor replicates (zs .1076/.1035, agree -.011/-.017) and the rotation now recovers only PARTIAL geometry (agree ->.304/.301, zs ->.138/.132; erank sags 103/89): the relational form degrades ABSOLUTELY at scale - the frame law's scale scoping, second seed pair.",
    "formula": "KL(softmax(sim_t/t) || log_softmax(sim_s/t)) per row, both frames' own sims",
    "impl": "tools/loss_forms.py affinity_kl_loss; forensics scratchpad frame_check.py -> tools/dist_runs/frame_check.json"
   },
@@ -5329,27 +5337,303 @@
    "flags": [],
    "family": "dist",
    "sub": {
-    "R": 2,
+    "R": 3,
     "P": 2,
     "D": 2,
-    "C": 1,
+    "C": 2,
     "I": 1,
     "L": 0
    },
-   "s_raw": 8,
-   "rating": 7,
+   "s_raw": 10,
+   "rating": 8,
    "rating_hi": null,
    "rules_fired": [],
    "cite": [
     "tools/dist_runs/",
     "canon/clip_mini_distillation.md",
-    "tools/dist_runs/frame_check_s1.json"
+    "tools/dist_runs/frame_check_s1.json",
+    "canon/cc12m_pod_matrix.md"
    ],
    "headline": "P7 REFUTED at the letter (raw zs .154/.161 loses to single-teacher 2/2) BUT post-rotation agree->LAION .7132/.7116 EQUALS feature_mse's in-frame .714/.718 both seeds: the GPA mean shape carries the geometry, loses only the gauge",
    "gauge": "zs-C10/C100 + COCO R@1/R@5 + agree_cos + cv16/erank",
-   "note": "Closes gm3 D18 (multi-teacher cell now RUN). GM3's student-exceeds-teachers did not appear. The L-025 operator's consensus target is geometry-faithful; judge it with a frame fit or it reads as a false floor. Rotation replicated s0+s1 (deltas 0.0016).",
+   "note": "Closes gm3 D18 (multi-teacher cell now RUN). GM3's student-exceeds-teachers did not appear. The L-025 operator's consensus target is geometry-faithful; judge it with a frame fit or it reads as a false floor. Rotation replicated s0+s1 (deltas 0.0016). | CC12M 2/2: raw zs .2738/.2681 but rotated agree .7654/.7629 EXCEEDS mimicry's in-frame .7345/.7379 - the GPA mean carries the geometry better than direct distillation preserves it. Rotated zs .4344/.3938; rotated retrieval MEASURED: R@1 .0456/.0508 — pure-MSE consensus cannot retrieve even rotated; the composite's NCE term is the retrieval payer (anchor dial). | CONSTRUCTION NOTE 2026-08-02: build_gpa aligns towers to a RUNNING MEAN and requires same-dim towers (torch.stack), which is both why the target is frame-free and why it cannot fuse a heterogeneous bank. The reference-member alternative is L-169.",
    "formula": "mean((z_s - gpa(z_t1..t5))^2), GPA cached iter-5 align .9401",
    "impl": "tools/dist_bed.py loss_of consensus_gpa (target build_gpa)"
+  },
+  {
+   "id": "L-162",
+   "name": "consensus NCE+MSE composite on a MEAN consensus (CaptionBert Stage-A form, vision; rotation required)",
+   "role": "PRIMARY",
+   "primitive": "CE",
+   "accum": "A7",
+   "substrate": "consensus-embedding",
+   "flags": [],
+   "family": "consensus-frame",
+   "sub": {
+    "R": 3,
+    "P": 3,
+    "D": 2,
+    "C": 2,
+    "I": 2,
+    "L": 0
+   },
+   "rules_fired": [
+    "split"
+   ],
+   "cite": [
+    "tools/dist_runs/consensus_nce_mse_s0.jsonl",
+    "tools/dist_runs/vitbank_s0.jsonl",
+    "canon/clip_mini_distillation.md",
+    "canon/cc12m_pod_matrix.md",
+    "repos/captionbert-v2.md"
+   ],
+   "headline": "THE CROWN AT SCALE: consensus composite + rotation beats direct deployment-teacher distillation on EVERY task gauge 2/2 (CC12M 11x data); local COCO verdict was BETWEEN - scale promoted it",
+   "gauge": "agree_gpa + rotated zs/R@1 + cv16/erank",
+   "note": "Per-SAMPLE MSE (sum(-1).mean = 2(1-cos)) is load-bearing: per-element mean is ~1000x smaller than the NCE term and the 1.0/1.0 weighting silently degenerates to pure NCE. 8000 steps COCO train, pure Adam (house override of the source AdamW). | CC12M CROWN 2/2: + the 2,500-pair rotation SWEEPS infonce on every task gauge (zs-C10 .5412/.5708 vs .5208/.4979; C100 .2210/.2487; R@1 .1322/.1360; R@5 .3140/.3300). Viability 44.7/48.1% of teacher. The GM3 student-exceeds shape in full: distill vs consensus, ship student + 0.5MB rotation. | SCOPE SPLIT 2026-08-02 (rule 9): this row is the objective against a RUNNING-MEAN consensus (build_gpa), which is frame-free and therefore ships a 0.5MB rotation. Its sibling L-169 is the same objective against a REFERENCE-MEMBER consensus, which is in-frame by construction. Do not average the two — the deployment requirement differs, not the loss.",
+   "formula": "a7_grid_infonce(zs, zt_gpa, t=.07) + prim_sq(zs, zt_gpa).sum(-1).mean()",
+   "impl": "tools/dist_bed.py loss_of consensus_nce_mse",
+   "s_raw": 12,
+   "rating": 10,
+   "rating_hi": null
+  },
+  {
+   "id": "L-163",
+   "name": "AlignmentBank 8-term geometry loss (CaptionBert Stage-B, vision port)",
+   "role": "AUX",
+   "primitive": "SQ",
+   "accum": "A2",
+   "substrate": "bank-geometry",
+   "flags": [],
+   "family": "vitbank",
+   "sub": {
+    "R": 2,
+    "P": 0,
+    "D": 1,
+    "C": 1,
+    "I": 2,
+    "L": 0
+   },
+   "rules_fired": [
+    "sub1pct<=5"
+   ],
+   "cite": [
+    "tools/dist_runs/vitbank_s0.jsonl",
+    "tools/dist_runs/vitbank_s1.jsonl",
+    "tools/dist_runs/vitbank_random_ctl.json"
+   ],
+   "headline": "PROBE-REFUTED with FULL CONTROLS 2/2: trained bank BELOW untrained (gap -.0038/-.0045) AND a fully-random bank (random frames+anchors) matches the real-frames control (.4159/.4114 vs .4149/.4107) - the entire +1.4-1.6pt enrichment over raw is generic random-feature capacity; the geometric content contributes ~zero and geometric training subtracts",
+   "gauge": "80-class COCO multi-label linear-probe mAP, trained-vs-untrained-control",
+   "note": "CAPACITY, NOT ADDRESSING - now with the three-way control: raw < random-bank ~= real-frames-untrained > trained, both seeds (tools/dist_runs/vitbank_random_ctl.json). The 538-d signature's probe value does NOT come from the fitted Procrustes frames or consensus anchors; any 128 random nonlinear features buy the same lift. Source recipe's +0.125 NLI gain did not transfer (their bed untested here, not refuted). emb_cv readout-only per the 1e-3 ceiling.",
+   "formula": "1.0*expert_agreement + 1.0*rotation_ortho + 0.5*anchor_spread + 0.1*anchor_entropy + 0.3*cross_expert_var + 0.3*|bank_cv-t| + 0.3*|emb_cv-t| + 0.5*disagree_preserve",
+   "impl": "tools/vit_bank.py VisionAlignmentBank.bank_loss",
+   "s_raw": 6,
+   "rating": 5,
+   "rating_hi": null
+  },
+  {
+   "id": "L-164",
+   "name": "x3 full stack on the vision student (NCE + elem-MSE .5 + BCE .3 + centered-cos align .5 + CV 1e-3 + EmbeddingAutograd)",
+   "role": "PRIMARY",
+   "primitive": "CE",
+   "accum": "A7",
+   "substrate": "consensus-embedding",
+   "family": "x3-arms",
+   "sub": {
+    "R": 2,
+    "P": 2,
+    "D": 1,
+    "C": 1,
+    "I": 2,
+    "L": 0
+   },
+   "rules_fired": [],
+   "headline": "WEAK-ANCHOR SET-POINT 2/2: agreement lands at .4957/.4958 (cross-seed 4e-4) — .17 BELOW the anchored composite — while task gauges stay in-noise and probe drops 1.1pt. Its faster NCE descent (~8%) is UNANCHORED-optimizer speed: the inert elem-MSE leaves the contrastive term unopposed (diag: base agree .593 at step500 vs .432; x3diag_*.jsonl)",
+   "gauge": "agree_gpa + rotated zs/R@1 + probe mAP + per-term diag",
+   "note": "The agreement value is a property of the LOSS STACK, not the seed: three stacks give three deterministic set-points (.41 cbert / .50 x3 / .67 anchored). Align term exonerated by ablation (noalign ≡ full on every curve). AMOE design input: anchor-weight is a clean retention<->plasticity dial.",
+   "formula": "a7_grid_infonce + 0.5*prim_sq.mean() + 0.3*BCE(frozen_head) + 0.5*centered_cos + 1e-3*x3_cv + EA(tang .01, sep 1.0)",
+   "impl": "tools/dist_bed.py loss_of x3_full",
+   "cite": [
+    "tools/dist_runs/x3_queue.log",
+    "tools/dist_runs/vitbank_s0.jsonl",
+    "tools/dist_runs/vitbank_s1.jsonl"
+   ],
+   "flags": [],
+   "s_raw": 8,
+   "rating": 7,
+   "rating_hi": null
+  },
+  {
+   "id": "L-165",
+   "name": "BCE through a frozen teacher-space head (x3 supervised force, isolated)",
+   "role": "AUX",
+   "primitive": "CE",
+   "accum": "A0",
+   "substrate": "frozen-head-logits",
+   "family": "x3-arms",
+   "sub": {
+    "R": 2,
+    "P": 0,
+    "D": 2,
+    "C": 2,
+    "I": 2,
+    "L": 0
+   },
+   "rules_fired": [
+    "sub1pct<=5"
+   ],
+   "headline": "INERT as an add-on 2/2: every gauge within noise of the anchored composite (agree +.008 consistent-sign whisper); supervision distilled through a frozen head does not behave like supervision",
+   "gauge": "agree_gpa + rotated zs/R@1 + probe mAP",
+   "note": "The head was fit on GPA targets then frozen (the x3 soup-head analog). Slight descent drag (~.06 EMA).",
+   "formula": "0.3 * BCE_logits(frozen_linear80(zs), labels)",
+   "impl": "tools/dist_bed.py loss_of x3_bce + _fit_teacher_head",
+   "cite": [
+    "tools/dist_runs/x3_queue.log",
+    "tools/dist_runs/vitbank_s0.jsonl",
+    "tools/dist_runs/vitbank_s1.jsonl"
+   ],
+   "flags": [],
+   "s_raw": 8,
+   "rating": 5,
+   "rating_hi": null
+  },
+  {
+   "id": "L-166",
+   "name": "EmbeddingAutograd gradient operator (radial damping + nearest-anchor separation), isolated",
+   "role": "AUX",
+   "primitive": "SQ",
+   "accum": "NA",
+   "substrate": "gradient-path",
+   "family": "x3-arms",
+   "sub": {
+    "R": 3,
+    "P": 2,
+    "D": 1,
+    "C": 2,
+    "I": 2,
+    "L": 0
+   },
+   "rules_fired": [],
+   "headline": "THE SEED-LOTTERY COLLAPSER (candidate, 2 seeds): pins rotated zs at .2539/.2534 (cross-seed 5e-4) where the baseline swings .232-.264 — means unchanged, VARIANCE collapsed. Mildly retards its own descent (~.03 EMA, it deletes gradient components)",
+   "gauge": "rotated zs cross-seed spread + per-term diag",
+   "note": "Identity forward; argmax-over-anchors in the BACKWARD only (flagged vs the aleph failure class — gradient-side, no forward selection event; run Phil-directed). Variance-stabilization is the interesting property for AMOE expert reproducibility; needs more seeds before claiming. | CC12M 2/2: VARIANCE COLLAPSE REPLICATES - zs seed-spread .0083 (.3398/.3315) vs base .0657 (.3271/.3928), 8x, means equal, agree pin-tight .7422/.7418. Two beds, two scales: EA is a variance-channel operator (P upgraded on the variance claim, not means).",
+   "formula": "backward: g <- (g - radial) + tang*radial; g <- g - sep*(g.nearest_anchor)_+ ; forward identity",
+   "impl": "tools/loss_forms.py EmbeddingAutograd",
+   "cite": [
+    "tools/dist_runs/x3_queue.log",
+    "tools/dist_runs/vitbank_s0.jsonl",
+    "tools/dist_runs/vitbank_s1.jsonl",
+    "canon/cc12m_pod_matrix.md"
+   ],
+   "flags": [],
+   "s_raw": 10,
+   "rating": 8,
+   "rating_hi": null
+  },
+  {
+   "id": "L-167",
+   "name": "CV force at the 1e-3 ceiling on the anchored composite (x3 cv form)",
+   "role": "AUX",
+   "primitive": "SQ",
+   "accum": "A8",
+   "substrate": "pentachoron-volumes",
+   "family": "x3-arms",
+   "sub": {
+    "R": 3,
+    "P": 0,
+    "D": 2,
+    "C": 2,
+    "I": 2,
+    "L": 0
+   },
+   "rules_fired": [
+    "sub1pct<=5"
+   ],
+   "headline": "INERT 2/2, third independent confirmation of the P4 shape: trajectory rides the baseline curve to ~.01 EMA the whole way; gauges in-noise (zs seed-spread if anything wider)",
+   "gauge": "EMA trajectory overlay + full gauge set",
+   "note": "The CV force at the lawful ceiling is DYNAMICALLY INVISIBLE on this substrate — now shown in trajectory form, not just endpoints. | CC12M 2/2: inert THIRD substrate (agree .7434/.7410 = base's .7405/.7420; set-point unmoved).",
+   "formula": "1e-3 * |CV_16sets(emb) - cv_target|  (cv_target = measured consensus CV .164)",
+   "impl": "tools/dist_bed.py loss_of nce_mse_cv (loss_forms.x3_cv_loss)",
+   "cite": [
+    "tools/dist_runs/x3_queue.log",
+    "tools/dist_runs/vitbank_s0.jsonl",
+    "tools/dist_runs/vitbank_s1.jsonl",
+    "canon/cc12m_pod_matrix.md"
+   ],
+   "flags": [],
+   "s_raw": 9,
+   "rating": 5,
+   "rating_hi": null
+  },
+  {
+   "id": "L-168",
+   "name": "CaptionBert Stage-A VERBATIM replication on the vision student (elem-MSE 1.0 + CV 0.1 + AdamW/warmup/cosine/clip)",
+   "role": "PRIMARY",
+   "primitive": "CE",
+   "accum": "A7",
+   "substrate": "consensus-embedding",
+   "family": "x3-arms",
+   "sub": {
+    "R": 3,
+    "P": 2,
+    "D": 2,
+    "C": 1,
+    "I": 2,
+    "L": 0
+   },
+   "rules_fired": [],
+   "headline": "THE REPLICATION ANCHOR 2/2 (prototype-replication law's first yield): the faithful recipe is WORSE than the adaptation on nearly every gauge — agree .4113/.4120 vs .672/.670, zs_rot .213/.221 vs .232/.264, probe .390/.386 vs .401/.394 — so the bank refutation was NOT adaptation damage; the adaptation is the stronger Stage-A",
+   "gauge": "full gauge set vs the adapted composite",
+   "note": "REPLICATION EXCEPTION flags carried: CV 0.1 (100x ceiling) + AdamW wd .01 + warmup/cosine/clip, all verbatim-sanctioned by the law. Mechanism of the deficit: inert elem-MSE (no absolute anchor) + CV spending gradient on regularity + decay damping. Also the weak-anchor set-point pattern: agree cross-seed 7e-4. | CC12M 2/2: verbatim set-point .5314/.5300 (cross-seed .0014) vs adapted .741 - ADAPTED>VERBATIM replicates at 11x; the set-point itself is budget-modulated (.41 local -> .53 at scale).",
+   "formula": "a7_grid_infonce + prim_sq.mean() + 0.1*x3_cv_loss; AdamW(3e-4, wd .01) + LinearLR(1000) + cosine(1e-6) + clip 1.0",
+   "impl": "tools/dist_bed.py loss_of cbert_full + train_arm sched branch",
+   "cite": [
+    "tools/dist_runs/x3_queue.log",
+    "tools/dist_runs/vitbank_s0.jsonl",
+    "tools/dist_runs/vitbank_s1.jsonl",
+    "tools/dist_runs/cbert_full_s0.jsonl",
+    "tools/dist_runs/cbert_full_s1.jsonl",
+    "canon/cc12m_pod_matrix.md"
+   ],
+   "flags": [],
+   "s_raw": 10,
+   "rating": 8,
+   "rating_hi": null
+  },
+  {
+   "id": "L-169",
+   "name": "consensus NCE+MSE composite on a REFERENCE-MEMBER consensus (captionbert-8192-v2/-b, text; NO rotation)",
+   "role": "PRIMARY",
+   "primitive": "CE",
+   "accum": "A7",
+   "substrate": "consensus-embedding",
+   "flags": [],
+   "family": "consensus-frame",
+   "sub": {
+    "R": 3,
+    "P": 3,
+    "D": 2,
+    "C": 2,
+    "I": 2,
+    "L": 0
+   },
+   "s_raw": 12,
+   "rating": 10,
+   "rating_hi": null,
+   "rules_fired": [
+    "split"
+   ],
+   "cite": [
+    "repos/captionbert-v2.md",
+    "repos/captionbert-b.md",
+    "canon/consensus_distillation_text.md",
+    "history/CAMPAIGN_RECORD_2026-08-02_captionbert.md"
+   ],
+   "headline": "THE SIBLING OF L-162 AND THE CHEAPER DEPLOYMENT: same objective, but the consensus is aligned to a REFERENCE MEMBER (bert-base) instead of a running mean, so the frame gap collapses monotonically to +.0010 and NO rotation ships. Student beats its best teacher (.6077 vs distilbert .5883) at 13% of the teachers' combined params; replicated on a second trunk (-b .6031, a null on 19% more data)",
+   "gauge": "8-task STS mean + erank + self_cos + frame gap",
+   "note": "Two modalities now carry this objective: vision (L-162, mean consensus, rotation) and text (here, reference-member, none). The loss is identical; the deployment requirement is set by the TARGET'S CONSTRUCTION. Rank ceiling is an AGREEMENT limit — the consensus target uses 28.7 of 768 directions, and 19% more corpus moved nothing.",
+   "formula": "a7_grid_infonce(zs, zt_ref, t=.07) + prim_sq(zs, zt_ref).sum(-1).mean(); zt_ref = normalized centroid of teachers whitened-Procrustes-mapped into a REFERENCE MEMBER's frame",
+   "impl": "tools/captionbert_8192_v2.py (stage 1 maps + stage 3 loss)"
   }
- ]
+ ],
+ "amendments": []
 }