Vault Audit Report — 2026-05-03

Summary

Broken Wikilinks

Both originate in wiki/synthesis/gap-report.md:

• cloud/cost-optimisation-cloud — no file at that path. The cloud category has finops-cost-management.md — the gap-report link predated that page's actual slug. Fixed.
• qa/exploratory-testing-advanced — no file at that path. A gap-report entry pointing to a page that was never created. Fixed.

Recommended action: Update gap-report.md — replace cloud/cost-optimisation-cloud with [cloud/finops-cost-management](/cloud/finops-cost-management), remove or update the qa/exploratory-testing-advanced reference.

Index Inconsistencies

Files not in index (32)

All 32 are synthesis/debug-* runbooks from a prior sprint that were added to the vault but never listed in wiki/index.md:

synthesis/debug-agent-loop-not-terminating
synthesis/debug-agent-not-using-tools
synthesis/debug-alert-firing-incorrectly
synthesis/debug-api-timeout
synthesis/debug-auth-failing
synthesis/debug-cache-inconsistency
synthesis/debug-ci-pipeline-failing
synthesis/debug-cloud-cost-spike
synthesis/debug-cors-error
synthesis/debug-data-pipeline-failing
synthesis/debug-database-migration-failing
synthesis/debug-deadlock
synthesis/debug-dns-resolution-failing
synthesis/debug-duplicate-writes
synthesis/debug-embedding-quality-degraded
synthesis/debug-error-rate-after-deploy
synthesis/debug-fine-tuned-model-worse
synthesis/debug-flaky-test
synthesis/debug-hallucination-in-production
synthesis/debug-high-cpu
synthesis/debug-kubernetes-pod-not-starting
synthesis/debug-llm-high-latency
synthesis/debug-memory-leak
synthesis/debug-no-logs-in-production
synthesis/debug-prompt-injection-detected
synthesis/debug-rag-pipeline-slow
synthesis/debug-rag-wrong-context
synthesis/debug-scaling-not-triggering
synthesis/debug-secret-leaked
synthesis/debug-slow-query
synthesis/debug-ssl-certificate-error
synthesis/debug-websocket-dropping

Recommended action: Add all 32 to wiki/index.md under a "Debugging Runbooks" subsection in the synthesis category.

Index entries without file (1)

• synthesis/audit-report — was listed before the file existed. Now resolved (this report).

Frontmatter Issues

None. All 377 pages pass frontmatter checks (type, category, para, tldr, updated all present).

Potential Duplicates — Semantic Review

1. cs-fundamentals/grpc.md vs java/grpc.md

Verdict: Complementary — false positive.

• cs-fundamentals/grpc.md: Language-agnostic reference — protocol comparison table (gRPC vs REST vs GraphQL), proto definition syntax, 4 streaming modes, retry policy, deadline propagation.
• java/grpc.md: Java-specific for AI workloads — Java-to-Python service boundary patterns, Java stub generation, bidirectional streaming for LLM token output, Java code examples throughout.

Both pages already cross-link each other. No action needed.

2. papers/constitutional-ai.md vs safety/constitutional-ai.md

Verdict: Complementary — false positive.

• papers/constitutional-ai.md: Paper page — full citation (Bai et al., 2022, arXiv:2212.08073), what problem it solved, key contributions, limitations, follow-up work.
• safety/constitutional-ai.md: Applied concept page — two-phase pipeline (SL-CAF, RLAIF), how CAI is used in practice, relationship to DPO and RLHF.

The standard vault pattern for a paper that defines a field: one papers/ page for the source document, one concept page for the technique. Both cross-link each other. No action needed.

3. papers/mechanistic-interpretability.md vs safety/mechanistic-interpretability.md

Verdict: Complementary — false positive.

• papers/mechanistic-interpretability.md: Four-paper research history — Zoom In (Olah 2020), Toy Models of Superposition (2022), Towards Monosemanticity (2023), Scaling Monosemanticity (2024). Paper-by-paper breakdown with contributions and findings.
• safety/mechanistic-interpretability.md: Current state of the field — SAEs as the breakthrough tool, superposition explanation, activation steering, practical implications for alignment.

Research history vs applied concept. Both cross-link each other. No action needed.

Misplaced Pages

None detected. All 377 pages pass the directory-vs-frontmatter-category check.

Placement Sense-Check (20 most recently updated)

Reviewed all 20 pages updated 2026-05-02/05-03. No misplacements found. Note:

• synthesis/technical-communication, synthesis/request-flow-anatomy, synthesis/engineering-tradeoffs — broad SE synthesis pages that could arguably sit in cs-fundamentals/ but their content is genuinely synthesis-style (cross-cutting comparisons, decision frameworks). Current placement is defensible.
• cloud/bedrock-agentcore — correctly placed; AWS Bedrock AgentCore is a cloud-hosted AI infra service.

Accuracy Issues

Confirmed issue

papers/llama-2.md — GPU memory requirement for 70B inference is understated.

Line 155: "70B requires at minimum 2× 40GB GPUs (or 4-bit quantisation for single-GPU inference)."

At fp16, a 70B model requires approximately 140 GB VRAM. 2× 40GB = 80 GB — insufficient. The correct minimum is 2× 80GB GPUs (e.g., A100/H100 80GB) or 4× 40GB GPUs for fp16 inference.

Recommended action: Change to "at minimum 2× 80GB or 4× 40GB GPUs (or 4-bit quantisation for single-GPU inference on an A100 80GB or RTX 4090)."

Appropriate [unverified] markers (not errors — pending cross-check)

• papers/llama.md: Training data percentages table, benchmark scores table marked [unverified — verify from arXiv:2302.13971]
• papers/llama-2.md: Pretraining details and benchmark numbers marked [unverified — verify from arXiv:2307.09288]
• ai-tools/cline.md: Stars/installs claim (61k+ stars, 5M+ installs) marked [unverified]

These are correctly tagged. Strip the markers after verifying against primary sources.

Recommended Actions (prioritised)

1. Index the 32 debug runbooks — adds 32 pages to the catalog that are currently undiscoverable via wiki/index.md. High value, low effort.
2. Fix 2 broken links in gap-report.md — cloud/cost-optimisation-cloud → [cloud/finops-cost-management](/cloud/finops-cost-management), remove/replace qa/exploratory-testing-advanced.
3. Fix GPU memory claim in papers/llama-2.md:155 — factual error; 2× 40GB is not enough for fp16 70B inference.
4. Strip [unverified] from llama/llama-2 claims after verifying against arXiv:2302.13971 and arXiv:2307.09288.

Connections

• index — coverage map used for consistency checks
• synthesis/graph-health — complementary structural health audit (score 100/100 as of 2026-05-03)
• synthesis/gap-report — ranked knowledge gaps