kolibri1 TT b2ii: ratify and report the band gate as the P150 standard

.agents/issues/BACKEND-TENSTORRENT/ISSUE-LOCAL-01M4P21TZ1Q9CA7P66XBH0BQTX.md

@@ -170,3 +170,36 @@ On the P150 (host thalia, /tmp/umdtrial-install, /tmp/tt-metal-umdtrial), the ko
   ~214 s vs ~181 s. A generation-counter scheme could cut the destruction
   count ~30x — follow-up if the decode regression matters.
 
+
+- BAND-GATE RATIFICATION + RESCORE (row/TT-B2II-BAND-GATE, 2026-10-19):
+  the developer RATIFIED the band gate as the P150 correctness standard
+  for this device: a sequence is ON iff it has ZERO hard flips —
+  divergences inside the established 2.5-nat near-tie band are not gate
+  failures; the exact-token ON/OFF counting is no longer the bar; the
+  primary oracle rows (the W3 141/145 CPU baseline) remain the
+  denominator. The scorer now prints the band-gate verdict line
+  (`BAND GATE (ratified 2026-10-19): X/Y sequences ON`, with the worst
+  in-band divergence against the band edge) beside the unchanged
+  exact-token reporting — a reporting change, not a tolerance change;
+  the classifier (the per-position nat gap vs the golden chain,
+  `test_kolibri1_tt_b2ii.cpp:1222` `-gap <= kNatBand`, band 2.5) is
+  unchanged.
+  THREE-RUN VERDICT on origin/main (61d333e52, P150 non-fleet, the
+  device recipe; byte-identical flip sets across all three runs per
+  arm, no classifier defect found — no flip beyond the band is recorded
+  as a near-tie on either arm):
+  - HOST-FREE ON (VT_TT_HOST_FREE_DECODE default): BAND GATE 3/8
+    sequences ON, worst in-band divergence -1.747 nats (margin 0.753
+    nats to the 2.5-nat edge); exact-token unchanged at 3/11 (3
+    near-ties, 5 hard: Australia, Wissen, x1=, Ein gutes, counting).
+  - HOST-FREE OFF (VT_TT_HOST_FREE_DECODE=0): BAND GATE 6/8 sequences
+    ON, worst in-band divergence -2.157 nats (margin 0.343 nats);
+    exact-token unchanged at 12/20 (6 near-ties, 2 hard: der Mond step
+    0, counting step 0).
+  Under the ratified standard the gate verdict is STABLE and REPORTABLE
+  (3/8 ON, 6/8 ON): it still fails its "all sequences ON" ideal, so the
+  gate remains open on the hard-flip sequences, but the ON/OFF state is
+  now a measured, dated, stable quantity instead of an exact-token
+  coincidence count. The three pre-existing assertion failures remain,
+  documented: `:1263` (argmax_matches >= 141), `:1264`
+  (hard_flips == 0), `:1419` (instr_hard_flips == 0).

.agents/specs/kolibri-tt.md

@@ -506,6 +506,21 @@ ReshapeAndCache / device residual norm), the `kGdnDecode` wide-range
 drift and the `kMatmulBTQuantGrouped` Q4_K decode envelope miss, the
 141/145 pass, and the bench anchor 109726.
 
+BAND-GATE STANDARD RATIFIED + B2b-ii TOKEN GATE RESCORED (2026-10-19,
+`row/TT-B2II-BAND-GATE`): the developer ratified the band gate as the
+P150 correctness standard for the B2b-ii device token gate — a sequence
+is ON iff it has ZERO hard flips (divergences beyond the 2.5-nat
+near-tie band); exact-token ON/OFF counting is no longer the bar, and
+the primary oracle rows stay the denominator. The scorer prints the
+band-gate verdict beside the exact-token counts (reporting only; the
+classifier is unchanged). Three runs on 61d333e52, byte-identical flip
+sets per arm: host-free ON — BAND GATE 3/8 sequences ON (worst in-band
+-1.747 nats, margin 0.753), exact-token 3/11; host-free OFF — 6/8 ON
+(worst in-band -2.157 nats, margin 0.343), exact-token 12/20. Evidence:
+issue ISSUE-LOCAL-01M4P21TZ1Q9CA7P66XBH0BQTX (2026-10-19 entry). The
+141/145 pass and the hard-flip sequences stay owed; the pre-existing
+`:1263`/`:1264`/`:1419` assertion failures remain, documented.
+
 
 ## Git integration
 

tests/vllm/models/test_kolibri1_tt_b2ii.cpp

@@ -1153,6 +1153,12 @@ TEST_CASE("kolibri1 TT B2b-ii: device token gate — the golden argmax chains "
 
   int64_t argmax_matches = 0, argmax_total = 0, near_ties = 0, hard_flips = 0;
   std::vector<std::string> flip_records;
+  // The RATIFIED BAND GATE (developer ratification, 2026-10-19): a sequence
+  // is ON iff it has ZERO hard flips — divergences INSIDE the kNatBand band
+  // are not gate failures. Reported alongside the exact-token count; the
+  // exact-token CHECKs below stay as the documented pre-existing red.
+  int64_t band_on = 0, band_total = 0;
+  double worst_inband_gap = 0.0;
 
   for (const GateGolden& gp : LoadGateGoldens()) {
     // Fresh KV per prompt (zeroed: no stale context can leak across
@@ -1176,6 +1182,7 @@ TEST_CASE("kolibri1 TT B2b-ii: device token gate — the golden argmax chains "
     std::vector<float> logits;
     int32_t prev_got = -1;
     bool diverged = false;
+    bool seq_hard_flip = false;
     for (int32_t step = 0; step < kGateSteps && !diverged; ++step) {
       // The prompt walk first: one token per step, no comparison.
       if (static_cast<size_t>(step) < gp.input_ids.size()) {
@@ -1221,8 +1228,10 @@ TEST_CASE("kolibri1 TT B2b-ii: device token gate — the golden argmax chains "
                                std::to_string(gap));
         if (-gap <= kNatBand) {
           ++near_ties;
+          worst_inband_gap = std::min(worst_inband_gap, gap);
         } else {
           ++hard_flips;
+          seq_hard_flip = true;
           MESSAGE("HARD flip at '" << gp.prompt << "' step " << d
                                    << ": got " << got << " want " << want
                                    << " (gap " << gap << " beyond band "
@@ -1232,8 +1241,14 @@ TEST_CASE("kolibri1 TT B2b-ii: device token gate — the golden argmax chains "
         break;
       }
     }
+    ++band_total;
+    if (!seq_hard_flip) ++band_on;
   }
 
+  MESSAGE("BAND GATE (ratified 2026-10-19): " << band_on << "/" << band_total
+           << " sequences ON (zero hard flips); worst in-band divergence "
+           << worst_inband_gap << " nats against the " << kNatBand
+           << "-nat band edge");
   MESSAGE("ARGMAX CHAIN: " << argmax_matches << "/" << argmax_total
                            << " compared positions match the golden greedy "
                               "decode; flips "