aprender-test-lib 0.68.2

# perf-matrix.yaml — THE EXPECTED CELL SET, and the ONE place a number lives.
#
# Governing spec: docs/specifications/PP-LLAMA-001-MASTER.md, PP-1 (the expected
# cell set is enumerated HERE and the verdict asserts every one is present) and
# PP-33 (every threshold, floor, ceiling, ratchet direction and phase the gate
# reads lives in this file with `threshold_class` and `author`). A gate that only
# reads the receipts it was handed cannot tell a missing host from a passing one,
# and a threshold typed into a script cannot be changed without a code review of
# the gate.
#
# GROUNDING RULE (PP-33). Every number in this file is one of:
#   policy    — a deliberate product decision, needs an author, NOT a measurement
#   inherited — set by a prior governing spec, precedence recorded
#   ratchet   — today's measurement, monotone thereafter
# There are no invented continuous thresholds. A baseline that has not been
# measured under a conformant protocol is UNMEASURED with an owner and an
# expiry -- never a plausible-looking number.
#
# CELL STATUS VOCABULARY (PP-1, §7.4). A baselines.<host>.<workload> cell is:
#   MEASURED   {receipt, commit, n, interleaved, bands: {c<N>: {agg, dec, prefill}}}
#              the seed Arm A ratchets against. `receipt` must exist in the tree.
#   UNMEASURED {owner, and EXACTLY ONE of expires | expires_after}
#              temporary, COUNTED against the denominator, RED once expired.
#   NA         {reason, decided_by, date}
#              permanent, excluded from the denominator. Never a silent skip:
#              the gate prints the decision and its author on every run.
#
# WHO READS WHAT
#   scripts/perf_gate.sh                 arms, phases, baselines, ladder, hosts,
#                                        witness, stream, expiry_anchors
#   scripts/lib/perf_receipt.py          ladder.declared, protocol
#   scripts/lib/parity_block.py          ladder.declared, arms.L3.delta, derivation
#   scripts/lib/bench_receipt.py         derivation.ratio_tolerance
#   scripts/check_perf_matrix_schema.sh  every rule stated in this header
#   scripts/check_thresholds_in_matrix.sh  refuses a float literal in the readers
schema_version: 2

# ---------------------------------------------------------------- protocol --
# PP-22: the protocol values are part of the JOIN KEY, so two receipts taken
# under different windows are not comparable and must not be quotiented. The
# tree's dual-bound window (max(30, 8c) samples AND 60 s) with 2 warmup requests
# per worker and a 5 s quiesce is NORMATIVE; the 10 s cooldown between lanes is
# added here so an interleaved pair cannot inherit the previous lane's thermals.
protocol:
  window_ms: 60000
  warmup_requests_per_worker: 2
  quiesce_ms: 5000
  cooldown_ms: 10000
  n_predict: 128
  prompt_tokens: 512
  replicates_min: 5
  interleaved: true
  # PP-28: an unpinned sampler is an unpinned measurement. `n_predict` travels
  # on the wire as OpenAI `max_tokens`; the corpus spells it that way too.
  sampler: {temperature: 0.0, seed: 0, ignore_eos: true}
  threshold_class: policy
  author: spec-owner

# ---------------------------------------------------------------- ladder ----
# PP-24. The declared ladder is a REQUEST; the derived one is what both servers
# admitted. The ladder was copied into five files (parity_block.py,
# perf_receipt.py x2, llama_pin.toml, and here); it lives here now and the
# producers read it.
ladder:
  declared: [1, 4, 8, 16]
  derive_from: [slots_admitted.apr, slots_admitted.llama]
  threshold_class: policy
  author: spec-owner

# --------------------------------------------------------------- witness ----
# PP-26. A batch-invariance witness with no minimum length agrees on nothing.
# `authors:` names the author of EACH number, because these two were decided by
# different people; the block-level `author` is the default for any key it omits.
witness:
  # PP-26 (v3.1): batch invariance is measured where it exists -- every slot
  # of an m=c batch of identical prompts must agree with every other slot to
  # min_agree_tokens, and no slot may emit one token id for max_constant_run
  # steps (#2753's signature ran for 116). The m=1 reference's agreement with
  # the batch is RECORDED per band (`divergence_at`), not gated: on lambda it
  # measures the fp divergence between kernel families, each of which is
  # batch-size invariant to the end (evidence/perf041/lambda/). It becomes a
  # gate when the top-2 margin instrument lands (master §12 row 22).
  min_agree_tokens: 64
  max_constant_run: 16
  m1_agreement: report
  max_age_days: 7
  threshold_class: policy
  author: spec-owner
  authors: {min_agree_tokens: spec-owner, max_constant_run: spec-owner, max_age_days: perf-gate}
  provenance: "[M] 2026-09-02 lambda RTX 4090 (evidence/perf041/lambda/witness.json, m1-vs-m4-three-prompts.txt): intra-batch and cross-m agreement within a kernel family reach n_predict; m=1-vs-batch agreement is kernel-family fp divergence, recorded not gated"

# ---------------------------------------------------------------- stream ----
# PP-27. A replayed stream reports the whole response at the end, so
# client_ttft/e2e approaches 1. At or above this the stream is not live.
stream:
  live_ttft_over_e2e_max: 0.95
  threshold_class: policy
  author: spec-owner

# ---------------------------------------------------- comparator template ----
# PP-20 / PP-15. The comparator lane is a QUANTITY per band, never a boolean
# flag: `-np {c}`, `-c {c} * n_ctx_slot`. A numeric batch size <= 1 switches
# llama.cpp's batching OFF and is REFUSED rather than tuned.
comparator_template:
  n_ctx_slot: 1024
  n_ctx_slot_min: 640
  n_batch_min: 2
  refuse_n_batch_le: 1
  kv_type: f16
  fa: auto
  threshold_class: policy
  author: spec-owner

# ------------------------------------------------------------- derivation ----
# The tolerance a DERIVED ratio must reproduce its own samples within. It is a
# policy number like any other: it used to live as `RATIO_TOLERANCE = 0.01` in
# scripts/lib/bench_receipt.py, where no matrix change could reach it.
derivation:
  ratio_tolerance: {value: 0.01, threshold_class: policy, author: perf-gate}
  # The bound above which a ratio is likelier a measurement error than a win.
  # It used to be `CEILING = 1.50` in scripts/lib/parity_block.py, beside a
  # `STRETCH = 1.50` that gated nothing at all.
  sanity_ceiling: {value: 1.50, threshold_class: policy, author: perf-gate}

hosts:
  lambda:
    accelerator: rtx-4090
    compute_class: cuda
    comparator: llamacpp
    reachable_by: [cuda]
    ci_runner: none
  gx10:
    accelerator: gb10
    compute_class: cuda
    comparator: llamacpp
    reachable_by: [cuda]
    ci_runner: none
  intel:
    accelerator: cpu
    compute_class: cpu
    comparator: llamacpp
    reachable_by: [cpu, wgpu]
    ci_runner: self-hosted
  # PP-16 / #2841. `compute_class: metal` named a path NO BUILD IN THIS TREE CAN
  # REACH -- crates/aprender-serve pins trueno-gpu with features ["cuda"] and
  # there is no Metal inference path at all -- so every receipt this host could
  # produce would have declared a class its binary never took. The honest cell
  # is NA with a decider and a date, not a class nothing implements.
  mini:
    accelerator: m4
    compute_class: null
    comparator: llamacpp
    reachable_by: [cpu]
    ci_runner: none
    status: NA
    reason: >-
      apr has no Metal inference path; W1 cannot be measured under any class
      this build reaches (#2841).
    decided_by: spec-owner
    date: '2026-09-02'
    github_issue: 2841

  # yoga joined the fleet 2026-09-09 (paiml/infra#486, YOGA-NIGHTLY-001). It is
  # declared here for ONE reason and its cells say so: PP-19's host-isolation
  # rule reads its host vocabulary from this block, so cuda-nightly's `ada-yoga`
  # job could not name a valid `perf-yoga` group until yoga existed here.
  # scripts/check_perf_concurrency_groups.sh found that job ungrouped the hour
  # it was written. The jobs that name it are 67-D1 cuda-unit and 67-D2, the
  # sm_89 parity host; rack-mounted, permanent, RTX 4060 Laptop (sm_89, 8 GB),
  # driver 595.
  #
  # NO PERFORMANCE CELL, deliberately, and this is not an oversight to be filled
  # in later (YOGA-NIGHTLY-001 §7, N-4): SILICON_FLOORS is keyed on compute
  # capability, and yoga answers `8.9` — the same key as the 24 GB / ~450 W
  # desktop 4090 the floor was derived from, from an 8 GB / 100 W laptop. The
  # UNCALIBRATED-SILICON refusal cannot fire because the key MATCHES while the
  # device does not. Until that is re-keyed (R-7), any number measured here
  # would be compared against a floor from another machine.
  yoga:
    accelerator: rtx-4060-laptop
    compute_class: cuda
    comparator: llamacpp
    reachable_by: [cuda]
    ci_runner: self-hosted
    status: NA
    reason: >-
      correctness-only host. SILICON_FLOORS is keyed on compute capability and
      yoga shares 8.9 with the 4090 the floor came from, so no throughput number
      taken here is comparable to anything (YOGA-NIGHTLY-001 §7, R-7, N-4).
    decided_by: spec-owner
    date: '2026-09-09'
workloads: [W1, W2]

arms:
  # ---------------------------------------------------------------- Arm A ----
  # PP-31. The arm used to ratchet scaling_efficiency(c) = (agg(c)/agg(1))/c
  # UP-ONLY, which FALLS when agg(1) improves: a faster single-client path made
  # the gate redder. Self-regression ratchets the QUANTITIES, per band, against
  # the value the last MEASURED receipt on protected main achieved.
  # scaling_efficiency and overhead_share are REPORTED and never gated.
  A:
    name: self-regression
    phase: release
    comparator_required: false
    threshold_class: ratchet
    direction: down-is-fail
    statistic: lcb95
    n_min: 5
    author: perf-gate
    metrics: {agg: [1, 4, 8, 16], dec: [1, 4, 8, 16], prefill: [1]}
    reported: [scaling_efficiency, overhead_share]
  # --------------------------------------------------------------- Arm L3 ----
  # PP-3 / PP-17 / §7.5. B1 (aggregate floor 0.80, authorless) and B2 (decode
  # floor 1.00, `inherited_from` a document that never existed in any ref) are
  # REPLACED by one non-inferiority arm. The gated set is ASYMMETRIC by band
  # because the two metrics measure different things: at c=1 the aggregate IS
  # the decode, and at c>1 the aggregate is the only number a deployment feels.
  # PASS iff ratios.<metric>.lcb95 >= 1 - delta. An (cell, band, metric) absent
  # from `armed_by` is REPORTING -- a gate arms when a measurement arms it,
  # never on a date (P-6).
  L3:
    name: parity
    phase: release
    comparator_required: true
    statistic: lcb95
    threshold_class: policy
    author: spec-owner
    gated: {c1: [dec_ratio, prefill_ratio], c_gt_1: [agg_ratio]}
    reported: [agg_ratio@c1, dec_ratio@c_gt_1, prefill_ratio@c_gt_1, ttft,
               itl_p95, scaling_efficiency, overhead_share]
    delta: {dec_ratio: 0.0, prefill_ratio: 0.0, agg_ratio: 0.0}
    # Written by the first PASS: armed_by.<host>.<wl>.c<c>.<metric> = {receipt, commit}
    armed_by: {}
    rising_decode_trap: >-
      decode_ratio RISING while agg_ratio FALLS is the serialization signature,
      not a win: each request owns the whole device in turn. The measurement
      that named this trap is recorded in evidence/parity/LEDGER.md -- see the
      2026-08-25 row for commit 53062e7f3, whose validity_by_band marks c>1
      INVALID-BUILD. The digits are deliberately NOT quoted here: PP-12 forbids
      publishing a c>1 aggregate from that run, and this file used to quote two
      of them.
  # ---------------------------------------------------------------- Arm C ----
  C:
    name: integrity
    phase: both
    comparator_required: false
    # The integrity rules (completed == requested, timeouts fatal, tokenization
    # stated, drain recorded, zero-token bands refused) are unchanged from the
    # governing spec this one supersedes; the precedence is recorded rather than
    # re-derived, and scripts/check_perf_matrix_schema.sh checks the path exists.
    threshold_class: inherited
    inherited_from: docs/archive/perf-2026-09-01/APR-PERF-GATE-001-v2.2.md
    author: spec-owner
  # -------------------------------------------------------------- Arm L1 -----
  # The v3 wire rules that need no timing: server_config presence, the derived
  # ladder, the stream witness, the batch-invariance witness, the sampler pin,
  # the raw sample rows, the join key and the roofline. Static, so `both`.
  L1:
    name: v3-schema-rules
    phase: both
    comparator_required: false
  # ------------------------------------------------------------ expiry -------
  # The UNMEASURED-cell clock. Split out of Arm A (PP-6): Arm A is a RELEASE
  # arm, but "this cell has been UNMEASURED past its deadline" is a static fact
  # about the matrix that a merge can and should see. Keeping it inside Arm A
  # would have moved eleven armed rules to release in the same commit that
  # demoted Arm A, which is a disarm wearing the shape of a refactor.
  expiry:
    name: baseline-clock
    phase: both
    comparator_required: false
  # ------------------------------------------------------------- Arm C-sig ---
  C_sig:
    name: signature-freshness
    phase: release
    comparator_required: false
  # --------------------------------------------------------- cell set --------
  cells:
    name: cell-completeness
    phase: release
    comparator_required: false
  # ---------------------------------------------------------------- Arm D ----
  D:
    name: memory-efficiency
    phase: reporting
    promote_with: PERF-001
    # The NOTE threshold: refusing work while memory sits reserved-and-empty is
    # the contiguous-allocation signature this arm exists to catch. It was a
    # bare `0.5` in perf_gate.sh with no author and no matrix row.
    note_kv_utilization_below: 0.5
    threshold_class: policy
    author: perf-gate
  # ---------------------------------------------------------------- Arm E ----
  E:
    name: prefill-decode-interference
    phase: reporting
    workload: W2
    promote_with: PERF-001

# RATCHET BASELINES. Arm A is comparator-free, so it is the primary gate -- but
# a ratchet needs a measurement to ratchet FROM. Both committed 745fa8588 runs
# are SPENT and withdrawn (evidence/parity/LEDGER.md rows 1 and 2: the subject
# binary was built with continuous batching compiled out), so there is still
# nothing conformant to seed from. Committing a number anyway would be exactly
# the fabricated-baseline class this file exists to remove.
#
# THE FOUR W1 CELLS NO LONGER EXPIRE BY CALENDAR. They carried
# `expires: '2026-09-25'`, which would have turned Arm A into a date-armed FAIL
# on four hosts with no conformant measurement path in existence -- P-6 says no
# gate arms by date. The deadline is now the §12 row that BLOCKS the
# measurement, through the `expires_after` anchor mechanism below.
baselines:
  lambda:
    W1:
      status: UNMEASURED
      owner: perf-gate
      reason: >-
        Blocked on PP-LLAMA-001 row 18 (the reference measurement). The
        745fa8588 lambda run is SPENT and its ratios are withdrawn.
      expires_after: {anchor: PP-LLAMA-001-row-18, days: 0}
    W2: &w2_expires_with_perf_001
      status: UNMEASURED
      owner: perf-gate
      # NOT A CALENDAR DATE, deliberately. All four W2 cells carried
      # `expires: '2026-09-25'`, and under this file's own GROUNDING RULE that
      # number is none of policy / inherited / ratchet -- it is an invented
      # continuous threshold, in the one file that says there are none. It also
      # contradicted the section it implements: W2's expiry is dated from an
      # EVENT ("PERF-001 merge + 30 days"), not from the calendar, because
      # "a serialising server has no batch for a long prefill to interfere
      # with, so W2 measures nothing today and a blocking W2 would be
      # permanently red -- a gate people learn to walk past".
      #
      # The clock is therefore ARMED BY THE ANCHOR below, not by a date typed
      # here. Until PERF-001 merges the cell REPORTS and says so, naming what
      # has to happen; the PR that lands batching sets `merged_on` and the
      # 30-day clock starts.
      expires_after: {anchor: PERF-001, days: 30}
      expiry_decided_by: '@noah'
      expiry_cites: PP-LLAMA-001 §12
      expiry_decided_on: '2026-08-29'
      reason: >-
        W2 is REPORTING and non-blocking. Its expiry is "PERF-001 merge + 30
        days", owner @noah. The 2026-09-25 this cell used to carry had no
        author (PERF-056, #2777).
    D1:
      # 0.65.2 post-publish dogfood parity protocol (scripts/llama_pin.toml#protocol.http,
      # qwen2.5-coder-1.5b-instruct q4_k_m, ladder 1/4/8/16). NOT W1/W2: perf_gate.sh
      # --workload accepts only W1|W2, so this cell is a RECORDED seed (P-6, seeded at
      # achieved) that no arm consumes yet; nothing here is ARMED (armed_by untouched).
      # Ledger: evidence/parity/LEDGER.md row 7. Decision: decided_by noah 2026-09-05,
      # these receipts are the 0.66 baseline; discovery baseline 8e1e9ad40.
      status: MEASURED
      receipt: evidence/dogfood/0.65.2/lambda.json
      commit: 8e1e9ad40
      n: 5
      interleaved: true
      subject: 'published 0.65.2, default features (cpu lane; parity.lanes[0])'
      bands:
        c1: {agg: 23.177905280153, dec: 44.1002587776968, prefill: 47.74679967996131}
        c4: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}
        c8: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}
        c16: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}
    D1-cuda:
      # 0.65.2 post-publish dogfood parity protocol (scripts/llama_pin.toml#protocol.http,
      # qwen2.5-coder-1.5b-instruct q4_k_m, ladder 1/4/8/16). NOT W1/W2: perf_gate.sh
      # --workload accepts only W1|W2, so this cell is a RECORDED seed (P-6, seeded at
      # achieved) that no arm consumes yet; nothing here is ARMED (armed_by untouched).
      # Ledger: evidence/parity/LEDGER.md row 8. Decision: decided_by noah 2026-09-05,
      # these receipts are the 0.66 baseline; discovery baseline 8e1e9ad40.
      status: MEASURED
      receipt: evidence/dogfood/0.65.2/lambda.json
      commit: 8e1e9ad40
      n: 5
      interleaved: true
      subject: 'published 0.65.2 installed with --features cuda (parity.lanes[1])'
      bands:
        c1: {agg: 326.77673844031267, dec: 345.44981138263495, prefill: 5268.553650727174}
        c4: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}
        c8: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}
        c16: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}
  gx10:
    W1: {status: UNMEASURED, owner: perf-gate, expires_after: {anchor: PP-LLAMA-001-row-15, days: 0}}
    W2: *w2_expires_with_perf_001
    D1:
      # 0.65.2 post-publish dogfood parity protocol (scripts/llama_pin.toml#protocol.http,
      # qwen2.5-coder-1.5b-instruct q4_k_m, ladder 1/4/8/16). NOT W1/W2: perf_gate.sh
      # --workload accepts only W1|W2, so this cell is a RECORDED seed (P-6, seeded at
      # achieved) that no arm consumes yet; nothing here is ARMED (armed_by untouched).
      # Ledger: evidence/parity/LEDGER.md row 9. Decision: decided_by noah 2026-09-05,
      # these receipts are the 0.66 baseline; discovery baseline 8e1e9ad40.
      status: MEASURED
      receipt: evidence/dogfood/0.65.2/gx10.json
      commit: 8e1e9ad40
      n: 5
      interleaved: true
      subject: 'published 0.65.2, default features; block REFUSED, c1 from parity_attempt'
      bands:
        c1: {agg: 3.53, dec: 6.78, prefill: 7.18}
        c4: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}
        c8: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}
        c16: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}
  intel:
    W1: {status: UNMEASURED, owner: perf-gate, expires_after: {anchor: PP-LLAMA-001-row-18, days: 0}}
    W2: *w2_expires_with_perf_001
    D1:
      # 0.65.2 post-publish dogfood parity protocol (scripts/llama_pin.toml#protocol.http,
      # qwen2.5-coder-1.5b-instruct q4_k_m, ladder 1/4/8/16). NOT W1/W2: perf_gate.sh
      # --workload accepts only W1|W2, so this cell is a RECORDED seed (P-6, seeded at
      # achieved) that no arm consumes yet; nothing here is ARMED (armed_by untouched).
      # Ledger: evidence/parity/LEDGER.md row 10. Decision: decided_by noah 2026-09-05,
      # these receipts are the 0.66 baseline; discovery baseline 8e1e9ad40.
      status: MEASURED
      receipt: evidence/dogfood/0.65.2/intel.json
      commit: 8e1e9ad40
      n: 5
      interleaved: true
      subject: 'published 0.65.2, default features; block REFUSED, c1 from parity_attempt'
      bands:
        c1: {agg: 14.32, dec: 27.82, prefill: 28.99}
        c4: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}
        c8: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}
        c16: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}
  # An NA host has no denominator cells. Both are NA with a decider and a date;
  # `cell-completeness` passes an NA cell with no receipt at all, and says so.
  # yoga is NA for the same STRUCTURAL reason mini is, with a different cause:
  # mini has no Metal path to measure, and yoga has no floor its numbers could
  # be compared against — 8.9 is the 4090's key (YOGA-NIGHTLY-001 §7, R-7).
  yoga:
    W1: {status: NA, reason: 'SILICON_FLOORS keyed on 8.9, shared with the 4090 (R-7)', decided_by: spec-owner, date: '2026-09-09'}
    W2: {status: NA, reason: 'SILICON_FLOORS keyed on 8.9, shared with the 4090 (R-7)', decided_by: spec-owner, date: '2026-09-09'}
  mini:
    W1: {status: NA, reason: 'no Metal inference path (#2841)', decided_by: spec-owner, date: '2026-09-02'}
    W2: {status: NA, reason: 'no Metal inference path (#2841)', decided_by: spec-owner, date: '2026-09-02'}
    D1:
      # 0.65.2 post-publish dogfood parity protocol (scripts/llama_pin.toml#protocol.http,
      # qwen2.5-coder-1.5b-instruct q4_k_m, ladder 1/4/8/16). NOT W1/W2: perf_gate.sh
      # --workload accepts only W1|W2, so this cell is a RECORDED seed (P-6, seeded at
      # achieved) that no arm consumes yet; nothing here is ARMED (armed_by untouched).
      # Ledger: evidence/parity/LEDGER.md row 11. Decision: decided_by noah 2026-09-05,
      # these receipts are the 0.66 baseline; discovery baseline 8e1e9ad40.
      status: MEASURED
      receipt: evidence/dogfood/0.65.2/mini.json
      commit: 8e1e9ad40
      n: 5
      interleaved: true
      subject: 'published 0.65.2, default features; block REFUSED, c1 from parity_attempt'
      bands:
        c1: {agg: 4.53, dec: 8.85, prefill: 9.07}
        c4: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}
        c8: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}
        c16: {validity: 'INVALID-CORRECTNESS(#2753/#2776) — no PP-26 witness; not reported, not seeded (P-4)'}

# EXPIRY ANCHORS (PERF-056, #2777; PP-LLAMA-001 §8). A cell may date its expiry
# from an EVENT rather than the calendar, via `expires_after: {anchor: ID,
# days: N}`. §12 does exactly this for every non-root row and the matrix had no
# way to say it, so the rule was flattened into a hardcoded date that
# contradicted the section it came from.
#
# An unresolved anchor is NOT a free pass and NOT a silent one: the cell REPORTS
# with the anchor, its owner and its status printed on every run, so "W1 is
# still waiting on row 18" is a line in the gate output rather than an
# assumption. `perf_gate.sh` FAILS a cell whose anchor is undeclared here, whose
# `days` is not a non-negative integer, that declares both `expires` and
# `expires_after`, or whose anchor claims `status: merged` while recording no
# `merged_on` -- an expiry hanging off an event nothing defines is the absent-
# `expires` hole wearing a different field name.
#
# `scripts/spec_conformance.sh` derives each anchor's date from the master's §12
# blocked_by DAG and writes evidence/parity/derived_expiries.json; these entries
# are the matrix half of that join.
expiry_anchors:
  PERF-001:
    owner: '@noah'
    github_issue: 2706
    # DECLARED STATE, checked against docs/roadmaps/roadmap.yaml on 2026-08-29:
    # `status: in_progress`, downgraded from completed the same day by PERF-041
    # because the credited "3.32x aggregate at N=8" was throughput of GARBAGE
    # tokens -- batched CUDA decode emitted a constant token to the max_tokens
    # cap for every m>1 and never emitted a stop token (#2753), while the m=1
    # fast path in the same run returned coherent English.
    status: in_progress
    # Set by the PR that lands batching. Null means the clock has not
    # started; it does not mean the cell is exempt.
    merged_on: null
    decided_by: '@noah'
    cites: PP-LLAMA-001 §12
  PP-LLAMA-001-row-15:
    owner: perf-gate
    status: open
    merged_on: null
    decided_by: spec-owner
    cites: PP-LLAMA-001 §12 row 15
    reason: gx10 shakedown; until it lands, no gx10 W1 cell can be measured conformantly.
  PP-LLAMA-001-row-18:
    owner: perf-gate
    status: open
    merged_on: null
    decided_by: spec-owner
    cites: PP-LLAMA-001 §12 row 18
    reason: the reference measurement; until it lands, no host has a seed for Arm A.

# CELL EXCEPTIONS. A per-(host, band, arm) exception, distinct from a cell
# status: a band may be NOT_APPLICABLE for the comparator arm and still fully
# gated by Arm A, which needs no comparator.
#   NOT_APPLICABLE — permanent, excluded from the denominator, needs decided_by
#   UNMEASURED     — temporary, COUNTED against the denominator, needs an owner
#                    and an expiry, and goes RED once expired
# `comparator` must name a comparator some host declares; the entry that used to
# sit here named `vllm`, which no host in this file has ever declared, so it
# could not match any receipt and had never been read.
cell_exceptions:
- host: gx10
  band: 16
  comparator: llamacpp
  arm: L3
  status: NOT_APPLICABLE
  permanent: false
  decided_by: spec-owner
  decided_on: '2026-09-02'
  reason: >-
    gx10 admitted 32 slots against llama.cpp's server default; the c=16 pair is
    re-keyed by PP-24's derived ladder rather than compared across unequal
    admission. Removed when row 15 lands.