Files
Claude c99f48d5e3 StaffML: use the A17 Pro's 51.2 GB/s in five more questions
Follow-up to #2180 for the questions listed in the closing comment. The
mlsysim registry (hardware/data/mobile/iPhone15Pro.yaml) gives the A17 Pro
51.2 GB/s (LPDDR5); every figure derived from the old value is recomputed.

- mobile-0109 (68 GB/s): weight load 0.05 -> 0.07 ms; the ~3.8 ms
  overhead-bound total is unchanged.
- mobile-0998 (68 GB/s): A17 Pro TPOT bound 7.4 -> 9.8 ms. Snapdragon still
  wins both concurrency and TPOT, and Tensor G3 (10.0 ms) is still slower.
- mobile-1001 (68 GB/s): the scenario said iPhone 16 Pro but the title and
  analysis are A17 Pro, so it now says iPhone 15 Pro. The 7.4 ms baseline is
  faster than 500 MB / 51.2 GB/s allows, so TPOT is rescaled 7.4 -> 9.8 ms
  and 14 -> 18.6 ms; the implied sustained bandwidth is 26.9 GB/s, keeping
  the 52.5% ratio and the "roughly halved" conclusion.
- mobile-1021 (68 GB/s): the A17 Pro is an LPDDR5 design, so it now
  illustrates the LPDDR5 case that misses 30 ms. The LPDDR5 TPOT is
  1.5 GB / (51.2 - 2.6) GB/s = 30.9 ms; the old 51.2/50*30 = 30.7 ms had the
  ratio inverted. Tensor G3 (registry 60 GB/s, not 51.2) is dropped from that
  sentence.
- mobile-1060 (~100 GB/s): A17 Pro decode 66.7 -> 34.1 tok/s. Tensor G3
  uses the registry's 60 GB/s instead of 68 (45.3 -> 40.0 tok/s). The
  ranking becomes Snapdragon > Tensor G3 > A17 Pro; the lesson (decode
  follows bandwidth, not TOPS) is kept: Tensor G3 has the lowest TOPS but
  beats the A17 Pro.

Only text fields change. Refs #2180.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SxFzWDUc3UeWQ1QBGjw6UM
2026-11-06 08:27:11 +00:00
..
…
…
…

StaffML Vault

The authoritative home of the StaffML question corpus: the per-question YAML source, the taxonomy, release artifacts, schema, and the review ledger that got us here.

Corpus license: CC-BY-NC-4.0 — non-commercial use only with attribution. Commercial licensing requires written permission. See ../CONTRIBUTING.md for details.

What lives here

staffml/vault/
├── ARCHITECTURE.md          ← v2.2 design doc (definitive)
├── REVIEWS.md               ← 3-round adversarial review ledger
├── TESTING.md               ← test plan; see vault-cli for impl
├── README.md                ← this file
├── release-policy.yaml      ← SINGLE filter predicate for "what is published"
├── corpus-equivalence-hash.txt ← CI regression guard (release_hash pin)
├── schema/
│   ├── question_schema.yaml ← LinkML — sole schema source of truth
│   └── EVOLUTION.md         ← SemVer rules for schema versioning
├── questions/               ← THE SOURCE OF TRUTH — 9,657 YAML files
│   ├── LICENSE              ← CC-BY-NC-4.0
│   ├── cloud/{l1..l6}/<zone>/*.yaml
│   ├── edge/
│   ├── mobile/
│   ├── tinyml/
│   └── global/
├── exemplars/               ← curated, human-reviewed exemplar pool
├── drafts/                  ← unreviewed drafts awaiting `vault promote`
├── releases/                ← citable release artifacts
│   ├── 0.9.0/
│   │   ├── vault.db         ← compiled SQLite (committed, citable)
│   │   ├── release.json     ← release_hash, counts, git_sha, timestamp
│   │   ├── d1-migration.sql ← forward migration
│   │   └── d1-rollback.sql  ← inverse with prior-row bodies embedded
│   └── latest.txt           ← names the latest release; replaced atomically
├── id-registry.yaml         ← APPEND-ONLY log; CI rejects line deletions
├── schema/taxonomy_data.yaml ← topic graph; DAG-enforced
├── chains.yaml              ← chain definitions (optional)
├── zones.yaml               ← 8 ikigai zones
└── exemplar-gaps.yaml       ← local/CI-generated coverage audit output (gitignored)

Intermediate files: _pipeline/ (gitignored)

Tooling writes intermediate outputs (proposed chain files, coverage reports, run traces) to _pipeline/, which is gitignored as a unit. Only durable corpus artifacts (chains.json, id-registry.yaml, questions/) belong in git. When adding a tool, route default outputs through the PIPELINE_DIR constant (vault/_pipeline/) and never commit anything under it.

Quick commands

# From repo root, with venv activated and `pip install -e staffml/vault-cli/[dev]`:
vault --help                      # subcommands, phase-aware
vault build                       # compile YAML → vault.db
vault check --strict              # fast + structural invariants (<60s)
vault check --tier slow           # nightly tier incl. LSH scenario dedup
vault stats                       # scorecard over latest vault.db
vault doctor                      # 8 diagnostic subchecks
vault verify 0.9.0                # academic-citability round-trip
vault api --port 8002             # local Worker-surface shim for dev

Full CLI reference in ../vault-cli/README.md.

Flow — YAML → paper, YAML → site

                      vault/questions/*.yaml
                              │
                              ▼
                     `vault build` ─────────────► vault.db
                              │                      │
                   `vault check --strict`            │
                              │               ┌──────┴───────┐
                              ▼               ▼              ▼
                       (CI green or fail)   paper         D1 + Worker
                                            macros.tex    edge API
                                            via `vault    via `vault
                                            export-paper` deploy`
                                                           │
                                                           ▼
                                                   staffml site
                                                   (reads via vault-api.ts)

Every output is a function of the YAML source, versioned by release_hash (SHA-256 Merkle over content_hashes + taxonomy + chains + zones + policy + canon_version leaves; ARCHITECTURE.md §3.5). Paper and site agree by construction — they read from the same vault.db.

How to add a question

vault new --track cloud --level l4 --zone diagnosis \
          --topic kv-cache-management \
          --title "KV Cache Bandwidth Bottleneck on H100"
# Opens $EDITOR with a scaffolded YAML.
# On save: schema validates; failures inject comment block at top and re-open.

vault check --strict                    # make sure it passes invariants
git add staffml/vault/questions/...  # explicit add, not git add -A
git commit -m "question: KV cache bandwidth"

On vault new:

  • A content-addressed ID is allocated (<track>-<level>-<zone>-<topic>-<6hex>-<seq>).
  • id-registry.yaml is appended with {id, created_at, created_by}.
  • authors: is auto-populated from git config user.email.
  • A git pull --rebase runs first to reduce collision rate.

See ../CONTRIBUTING.md for the full contribution workflow, including post-rebase collision recovery via vault renumber.

Release workflow (composed)

vault publish 1.0.0                              # build + snapshot + migrations
vault verify 1.0.0 --git-ref v1.0.0              # citation-grade round-trip
vault diff 0.9.0 1.0.0 --classify                # review changes pre-ship
vault ship 1.0.0 --env staging                   # D1 → Next.js → paper-last
# Soak on staging, then:
vault ship 1.0.0 --env production --canary-percent 10

Invariants (26 total, tiered)

  • Fast (pre-commit, <1s): schema validation, unique IDs, lowercase paths, enum-valid path components, filename format, URL scheme allowlist, no raw HTML in scenarios, YAML DoS hardening (256KB cap, depth 10, no aliases).
  • Structural (CI, <30s parallelized): topic-in-taxonomy, chain-ref exists, contiguous chain positions, provenance metadata consistency, registry append-only, taxonomy DAG, applicability matrix, release-policy single-source.
  • LSH scenario-dedup (CI, <10s, LSH-blocked): MinHash + 16-band LSH + Jaro-Winkler within-bucket candidate pairs at 0.95 threshold.
  • Nightly: deep-dive URL reachability, napkin-math dimensional analysis (Pint).
  • Weekly: secret-leak grep.

Full specification: ARCHITECTURE.md §5.

Three-round adversarial review

This architecture survived three full rounds of adversarial review with four reviewers each (Chip Huyen, Jeff Dean, Soumith Chintala, and an industry ML engineer "David"). Every convergent finding was resolved in spec + code with inline engineering decisions. See REVIEWS.md for the finding-by-finding ledger and the deferred-items list with rationale.

License

The corpus under this directory (vault/questions/, taxonomy, chains, zones, release artifacts) is licensed CC-BY-NC-4.0. See questions/LICENSE. The vault-cli Python tooling under ../vault-cli/ is a separate artifact and is not covered by this LICENSE.