Skip to content

kolibri1 TT b2ii: ratify and report the band gate as the P150 standard - #3483

Merged
lu-zero merged 1 commit into
localai-org:mainfrom
lu-zero:row/TT-B2II-BAND-GATE
Oct 20, 2026
Merged

lu-zero merged 1 commit into
localai-org:mainfrom
lu-zero:row/TT-B2II-BAND-GATE

Conversation

@lu-zero

@lu-zero lu-zero commented Oct 19, 2026 •

Copy link
Copy Markdown
Collaborator

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

The ratified standard

The developer ratified (2026-10-19) 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 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 change

test_kolibri1_tt_b2ii.cpp 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 greedy chain,
-gap <= kNatBand (band 2.5), tests/vllm/models/
test_kolibri1_tt_b2ii.cpp:1222 — is untouched.

Three-run verdict (origin/main 61d333e, P150, device recipe)

Byte-identical flip sets across all three runs per arm; the verdict is
stable, no flapping.

Arm Band gate (ON sequences) Exact token Worst in-band Margin to edge
HOST-FREE ON (default) 3/8 ×3 runs 3/11 (3 near-ties, 5 hard) −1.747 nats 0.753 nats
HOST-FREE OFF (VT_TT_HOST_FREE_DECODE=0) 6/8 ×3 runs 12/20 (6 near-ties, 2 hard) −2.157 nats 0.343 nats

Both exact-token counts reproduce the recorded 3/11 and 12/20 exactly.
No classifier defect: no flip beyond the band is recorded as a near-tie
on either arm (checked programmatically across the flip records).

What becomes gateable

Under the ratified standard the b2ii token gate has a STABLE, dated,
measured ON/OFF verdict (3/8 ON, 6/8 ON) instead of an exact-token
coincidence count. The gate is not yet "all sequences ON": the
hard-flip sequences (ON arm: Australia, Wissen, x1=, Ein gutes,
counting; OFF arm: der Mond, counting) stay owed, and the three
pre-existing assertion failures remain, documented: :1263
(argmax_matches >= 141), :1264 (hard_flips == 0), :1419
(instr_hard_flips == 0).

Records

  • ISSUE-LOCAL-01M4P21TZ1Q9CA7P66XBH0BQTX — the 2026-10-19 ratification
    and rescore entry beside the recorded 3/11 / 12/20 numbers.
  • .agents/specs/kolibri-tt.md ## Now — the lifecycle note.

Gates

  • b2ii device token gate: 3 runs per arm, stable verdicts and
    byte-identical flip sets both arms.
  • Host kolibri1 suite: 18/21; test_kolibri1_dequant_cache,
    test_kolibri1_w3, test_kolibri1_ngram_spec fail identically on
    pristine origin/main (verified on a clean 61d333e build) —
    pre-existing on this host.
  • scripts/check-agent-record.py: OK.
  • scripts/agent-preflight.sh --staged: green except
    check-windows-portability.py, which fails identically on pristine
    origin/main (pre-existing, environment).
  • scripts/check-commit-style.py and the trailer suites: OK.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:zai/glm-5.3-flash [maki]

The developer ratified (2026-10-19) 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 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: X/Y
sequences ON, with the worst in-band divergence against the 2.5-nat
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, -gap <= kNatBand, tests/vllm/models/
test_kolibri1_tt_b2ii.cpp:1222) is untouched.

THREE-RUN VERDICT on origin/main 61d333e, P150 non-fleet under the
device recipe, byte-identical flip sets across all three runs per arm:

- HOST-FREE ON: BAND GATE 3/8 sequences ON, worst in-band -1.747 nats
  (margin 0.753); exact-token 3/11 (3 near-ties, 5 hard).
- HOST-FREE OFF: BAND GATE 6/8 sequences ON, worst in-band -2.157 nats
  (margin 0.343); exact-token 12/20 (6 near-ties, 2 hard).

No classifier defect: no flip beyond the band is recorded as a near-tie
on either arm. The verdict is stable and now a measured, dated
quantity. The three pre-existing assertion failures remain documented:
:1263 (argmax_matches >= 141), :1264 (hard_flips == 0), :1419
(instr_hard_flips == 0); they still fire and the hard-flip sequences
stay owed. Records: the 2026-10-19 entry on
ISSUE-LOCAL-01M4P21TZ1Q9CA7P66XBH0BQTX and the row spec's
`.agents/specs/kolibri-tt.md` `## Now`.

Gates: the b2ii device gate 3x stable per arm; host kolibri1 suite
18/21 (test_kolibri1_dequant_cache, test_kolibri1_w3,
test_kolibri1_ngram_spec fail identically on pristine origin/main —
pre-existing on this host); scripts/check-agent-record.py OK;
preflight --staged green except check-windows-portability.py, which
fails identically on pristine origin/main (pre-existing, environment).

This change follows .agents/AGENTS.md — an open canonical local issue
owns the work, the record edits ride the change that made them true,
the gate evidence above is dated, and the independent fresh review
passed before merge.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:zai/glm-5.3-flash [maki]
@lu-zero
lu-zero force-pushed the row/TT-B2II-BAND-GATE branch from b81046d to 7a4d712 Compare October 19, 2026 20:51
@lu-zero
lu-zero merged commit 24dfdf2 into localai-org:main Oct 20, 2026
23 of 28 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant