# Source register: use current primary evidence without overclaiming

**Cutoff:** 9 October 2026, Europe/Berlin. Sources were inspected during this research run. Publication date, research period and access date are different fields. A search engine's crawl time is not a publication date.

**Selection rule:** prefer official specifications and implementation docs, research papers with methods, and first-party empirical studies. Retain limitations and commercial interests. Search snippets were used for discovery; retained claims were checked against the relevant source text. No SEO comparison pages are used as evidence.

## 1. October developments that are relevant now

| Development | Date evidence | Relevance | Use in the presentation |
| --- | --- | --- | --- |
| OSCAR model escalation and verification [S20] | arXiv submission 1 October 2026 | Task-specific evidence for verified smaller-model workflows | Research appendix; not a general savings percentage |
| Enterprise inference evaluation framework [S22] | arXiv submission 5 October 2026 | Organises hardware, serving, agent and enterprise assessment | Framework inspiration with caveats below |
| Bedrock Sonnet 5.5 cache-read price note [S12] | Live pricing page says effective October 7; note does not print a year | Shows why price versions and cache mix matter | Qualified live-page observation; refresh before workshop |

No robust new October enterprise ROI experiment was established in this search. That gap is not filled with undated marketing claims. Use the stronger older studies and clearly identify their dates.

## 2. Recent context that is not an October publication

| Source | Verified timing | Appropriate role |
| --- | --- | --- |
| Herschel [S21] | Submitted 30 September 2026 | Recent production-profiling research signal |
| State of Tokenomics [S27] | 23 September 2026; Release Candidate 1.0 | Exploratory executive survey; inspect its sample before quoting percentages |
| MLPerf Inference v6.1 [S23] | 16 September 2026 | Benchmark evolution toward RAG and agentic workloads |
| Claude cost engineering article [S33] | Originally 8 September; updated to include a 22 September model release | Provider implementation lessons; update date not fully specified |
| McKinsey State of AI [S24] | 25 August 2026; fieldwork May–June | Leadership context with self-report limitations |
| FOCUS 1.4 [S04] | 10 June 2026 | Current inspected financial-data standard release |

## 3. Annotated primary-source register

Evidence codes: **D** official documentation/specification; **E** empirical research; **P** preprint; **S** survey; **V** provider-authored engineering evidence; **C** contextual first-party guidance. These codes describe evidence type, not a universal quality score.

### Cost, ownership and financial management

| ID / source | Date or version | Type / inspected scope | What it is useful for | Limitation |
| --- | --- | --- | --- | --- |
| S01 [FinOps for AI](https://www.finops.org/framework/technology-categories/ai/) | Live; publication date not established | D; framework sections | Roles and AI-specific financial challenges | Normative guidance, not measured outcomes |
| S02 [Unit Economics](https://www.finops.org/framework/capabilities/unit-economics/) | Live | D; definitions and metrics | Distinguishing resource and business units | Organisation must define its own unit |
| S03 [State of FinOps 2026](https://data.finops.org/) | 2026 edition; exact publication day not established | S; report and sample | Practitioner priorities | FinOps-community sample; not all enterprises |
| S04 [Introducing FOCUS 1.4](https://www.finops.org/insights/introducing-focus-1-4/) | 10 June 2026 | D; release and reconciliation sections | Invoice-related datasets and integrity | Export implementation varies by provider |
| S30 [Allocation](https://www.finops.org/framework/capabilities/allocation/) | Live | D; shared-cost strategies | Ownership metadata and allocation methods | Policy choice still required |
| S31 [Forecasting](https://www.finops.org/framework/capabilities/forecasting/) | Live | D; roles and variance | Collaborative demand planning | Does not predict a specific AI workload |

### Consumption monitoring and operational controls

| ID / source | Date or version | Type / inspected scope | What it is useful for | Limitation |
| --- | --- | --- | --- | --- |
| S05 [LiteLLM budgets and rate limits](https://docs.litellm.ai/docs/proxy/users) | Live; implementation-version references | D; enforcement requirements | Budget scopes and DB dependency | Test the deployed version and topology |
| S06 [LiteLLM request architecture](https://docs.litellm.ai/docs/proxy/architecture) | Live | D; request lifecycle | Checks and asynchronous accounting | Not proof of strict zero-overshoot control |
| S07 [Langfuse token/cost tracking](https://langfuse.com/docs/observability/features/token-and-cost-tracking) | Live | D; cost calculation and caveats | Supplied/inferred usage and price definitions | Estimates need contract reconciliation |
| S08 [OpenTelemetry inference token metrics](https://github.com/open-telemetry/semantic-conventions-genai/blob/main/docs/gen-ai/gen-ai-token-metrics.md) | Live main branch | D; metric definitions | Billable counters versus distributions | Pin a revision for implementation |
| S09 [AWS Budgets](https://docs.aws.amazon.com/cost-management/latest/userguide/budgets-managing-costs.html) | Live | D; notification limitations | Explaining alert latency | Alerts are not request admission control |

### Model use, serving and deployment

| ID / source | Date or version | Type / inspected scope | What it is useful for | Limitation |
| --- | --- | --- | --- | --- |
| S10 [Claude prompt caching](https://platform.claude.com/docs/en/build-with-claude/prompt-caching) | Live | D; mechanics, TTL and pricing | Matching-prefix reuse and separate charges | Model/platform-dependent; prices change |
| S11 [Bedrock batch inference](https://docs.aws.amazon.com/bedrock/latest/userguide/batch-inference.html) | Live | D; processing and restrictions | Async independent-record processing | Different feature surface from interactive calls |
| S12 [Bedrock pricing](https://aws.amazon.com/bedrock/pricing/) | Live; October 7 note | D; tiers, batch, price note | Commercial structures and current rate evidence | Region/model-specific; refresh actual quote |
| S13 [vLLM optimisation/tuning](https://docs.vllm.ai/en/latest/configuration/optimization/) | Live V1 guide | D; preemption and scheduling | Workload-dependent serving trade-offs | No universal best configuration |
| S14 [vLLM prefix caching](https://docs.vllm.ai/en/latest/features/automatic_prefix_caching/) | Live | D; mechanism and limits | Reusing prefill state | Does not accelerate generation directly |
| S15 [vLLM quantisation](https://docs.vllm.ai/en/latest/features/quantization/) | Live | D; formats/support matrix | Precision/memory and hardware support | Compatibility is not end-to-end speedup |
| S16 [Triton dynamic batching](https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/user_guide/batcher.html) | Live | D; queueing and batching | Throughput/waiting-time control | Use model-specific measurement |
| S17 [Sarathi-Serve](https://www.usenix.org/conference/osdi24/presentation/agrawal) | OSDI, July 2024 | E; conference abstract and methods summary | Prefill/decode scheduling mechanism | Historical runtime/hardware baselines |
| S18 [DistServe](https://www.usenix.org/conference/osdi24/presentation/zhong-yinmin) | OSDI, July 2024 | E; conference abstract and methods summary | Phase disaggregation and goodput | Benefit depends on scale/interconnect |
| S19 [RouteLLM](https://arxiv.org/html/2406.18665v4) | Submitted June 2024; v4 February 2025 | E; full-text method/limitations | Quality/cost routing with preference data | Two-model setting and distribution shift |
| S23 [MLPerf Inference v6.1](https://mlcommons.org/2026/09/mlperf-inference-v6-1-results/) | 16 September 2026 | E; benchmark organisation release | RAG/agentic benchmark evolution | System results do not establish ROI |
| S32 [Bedrock prompt caching](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html) | Live | D; eligibility and restrictions | Platform-specific cache capabilities | Check model/endpoint compatibility |
| S33 [Claude cost/performance engineering](https://claude.com/resources/articles/reducing-cost-and-improving-performance-with-claude-platform) | Original 8 September 2026; later update | V; implementation discussion | Cache stability and request design | Provider-authored results and mutable page |

### Latest research, business value and audience context

| ID / source | Date or version | Type / inspected scope | What it is useful for | Limitation |
| --- | --- | --- | --- | --- |
| S20 [OSCAR](https://arxiv.org/pdf/2610.00912) | Submitted 1 October 2026 | P; abstract, PDF setup and limitations | Verified iteration and escalation | Five selected problems; preparation cost matters |
| S21 [Herschel](https://arxiv.org/html/2609.40247v1) | Submitted 30 September 2026 | P/V; full-text production discussion | Adaptive production profiling | Single-provider author-reported setting |
| S22 [Enterprise inference evaluation](https://arxiv.org/html/2610.07094v1) | Submitted 5 October 2026 | P; framework and equations | Multi-layer evaluation taxonomy | Synthesis/proposal; reliability claim needs qualification |
| S24 [McKinsey State of AI](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai) | 25 August 2026 | S; results and methodology | Adoption versus reported financial impact | Self-report, GDP weighting, non-causal |
| S25 [Generative AI at Work, v2](https://arxiv.org/abs/2304.11771v2) | Revision 6 November 2024; journal 2025 | E; revised abstract; journal discovery cross-check | Observed support-worker productivity | One work setting and older tools |
| S26 [METR experiment-design update](https://metr.org/blog/2026-02-24-uplift-update/) | 24 February 2026 | E; methods and caveats | Why old productivity headlines need context | Later data explicitly considered unreliable |
| S27 [State of Tokenomics](https://www.tokeneconomics.com/state-of-tokenomics/) | 23 September 2026; RC 1.0 | S; date/context inspected | Further executive-financial research | Do not quote its percentages without full sample review |
| S28 [Knauf IT values](https://careerit.knauf.com/pages/our-values) | Live | C; company values | Pragmatic, reliable tone | Not evidence of internal AI maturity |
| S29 [Germany business customs](https://www.trade.gov/country-commercial-guides/germany-business-travel) | 1 August 2025 | C; business customs | Structured communication and timing | General guidance; follow actual participants |

### Provider eligibility, exact usage and financial reconciliation

The following official sources were additionally inspected on 9 October 2026. They establish documented rules/interfaces, not a measured provider ranking or a completed compliance assessment. The detailed synthesis is in 09_provider_eligibility_tokens_and_tco.md [local reference omitted].

| ID / source | Date or version | Type / inspected scope | What it is useful for | Limitation |
| --- | --- | --- | --- | --- |
| S34 [EDPB international data transfers](https://www.edpb.europa.eu/sme/be-compliant/international-data-transfers_en) | Live guidance; accessed 9 October | D; transfer criteria, routes and ordinary obligations | Assessing deployment eligibility and cross-border processing | General guidance, not a workload-specific legal conclusion |
| S35 [OpenAI data controls](https://developers.openai.com/api/docs/guides/your-data) | Live; accessed 9 October | D; residency, processing, exclusions and requirements | Distinguishing EU storage from EU inference and feature eligibility | Mutable service/contract-specific capabilities |
| S36 [OpenAI ResponseUsage SDK types](https://github.com/openai/openai-python/blob/main/src/openai/types/responses/response_usage.py) | Main branch; accessed 9 October | D; input/output totals and cache/reasoning details | Exact field mapping for the Responses interface | Pin revision; types alone do not define every applicable tariff |
| S37 [Bedrock TokenUsage](https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_TokenUsage.html) | Live; accessed 9 October | D; response counters and optional cache details | Exact field mapping for Converse | Verify selected model/API billing semantics |
| S38 [OpenAI Usage and Costs API example](https://developers.openai.com/cookbook/examples/completions_usage_api) | Originally 14 January 2025; live page inspected 9 October | D; organisational usage/cost aggregation | Financial reconciliation and available aggregate interfaces | Not per-business-task attribution; not an October publication |

S10 was also rechecked for Anthropic's separate input/cache buckets and cache-write TTL details. No comparative workload runs, hardware power measurements, provider contracts or hardware quotations were collected during this extension.

### Knauf Digital audience relevance

| ID / source | Date or version | Type / inspected scope | What it is useful for | Limitation |
| --- | --- | --- | --- | --- |
| S39 [Knauf Digital: Purpose & Impact](https://career.knaufdigital.com/pages/purpose-impact) | Live; accessed 9 October 2026 | C; published mandate and About section | Grounding the running example in commercial excellence, marketing, sales, service and customer-facing software | Does not establish actual AI deployments, buying priorities or participant expertise |
| S40 [Knauf Digital: People & Culture](https://career.knaufdigital.com/pages/people-culture) | Live; accessed 9 October 2026 | C; four values and behavioural descriptions | Accurate company-specific context: Menschlichkeit, Entrepreneurship, Partnership and Commitment | Context for tone, not evidence of AI governance maturity |

Use company context to choose a relevant example. A technical product-information assistant is our proposed illustration; published company context does not establish participant expertise, actual AI deployments or measured accuracy/savings.

## 4. Claims excluded or qualified after inspection

### Navigara extension: public approach versus verified economic evidence

The detailed original synthesis is in chapter 10 [local reference omitted], with 92 retrieved public URLs and seven inaccessible targets documented in chapter 11 [local reference omitted]. All 43 articles linked by the refreshed blog index were retrieved. Long sources received targeted section review; coverage is not a guarantee of a complete site mirror or recursive OSS-record scrape.

| ID / primary source | Date/version | Use | Evidence limit |
| --- | --- | --- | --- |
| S41 [Navigara methodology](https://research.navigara.com/methodology) | Live, accessed 9 October | Two-stage classification/scoring design | Vendor-defined proxy; aggregation needs clarification |
| S42 [Adaptive limits](https://navigara.com/product/token-limits) | Live | Floors, mandate exceptions, escalation and review mode | Management pattern, not proven marginal returns |
| S43 [AI-tool integration docs](https://docs.navigara.com/integrations/ai-tools) | Live | OTLP ingestion and identity linkage | Instructions, not a tested billing pipeline |
| S44 [Q2 report PDF](https://research.navigara.com/research/Navigara-Performance-Report-Q2-2026.pdf) | Revised 7 September 2026 | Observational evidence and rescoring limits | No quantified causal AI attribution |
| S45 [Bottleneck loop](https://navigara.com/blog/three-bottlenecks-one-loop) | 24 August 2026 | Whole-workflow reasoning | Vendor argument, not universal measured effect |
| S46 [Changelog](https://docs.navigara.com/changelog) | Latest listed: 7 October, v0.16.4 | Repository spend and issue linkage | Release notes, not empirical study |
| S47 [Claude Code monitoring](https://code.claude.com/docs/en/monitoring-usage) | Live | Approximate cost metrics and version-dependent streaming counters | Validate installed version against provider billing |
| S48 [DORA metrics](https://dora.dev/guides/dora-metrics/) | Updated 5 January 2026 | Current five-metric delivery framework | Context-specific, not a cash-ROI metric |
| S49 [Faros](https://www.faros.ai/) | Live | Primary-source market cross-check | Public product claims, not vendor ranking |
| S50 [Navigara token spend](https://navigara.com/product/token-spend) | Live | Routing beta and cost/output framing | Displayed examples not validated benchmarks |
| S51 [Performance docs](https://docs.navigara.com/concepts/performance) | Live | Scope, supporting roles and scoring limits | Raw cross-repository comparability limited |
| S52 [CEO questions](https://navigara.com/blog/ceo-questions-ai-roi) | 17 September 2026 | Tool-induced compute and review cost | Useful questions, not measured Knauf costs |
| S53 [Navigara security](https://navigara.com/security) | Live | Configured LLM endpoint and responsibility boundary | Deployment label does not establish eligibility |

Exclude the inconsistent spend/ETV examples and unsupported marginal-output forecast checked in chapter 10. Do not present vendor capacity equivalents as realised cash ROI, product mockups as customer savings, or competitor comparison pages as an independently verified market ranking. This extension adds no matched provider task runs, hardware quotes or measured self-hosting TCO.

### Reliability headline in the 5 October preprint

The paper claims that 2% per-step failure across 20 steps erases a 2× decoding advantage. **Author check:** `0.98^20 ≈ 0.6676`; the inverse success-probability factor is about 1.498. If each failed attempt incurs the same full work and there are no other penalties, a 2× speed advantage becomes approximately `2/1.498 = 1.335×`, not zero advantage.

Early aborts, checkpoints, tool overhead and correlations change the calculation. The useful general principle is to measure reliability and task economics jointly; the headline numerical assertion is not used in the slides. Do not turn this framework paper into a claim that one hardware vendor is generally superior. [S22]

### Other exclusions

- A FinOps member listing advertising generic 20–40% savings: commercial claim without methods sufficient for this talk.
- Third-party mirrors and reposts: use original publications instead.
- Raw search snippets, undated trend pages and crawl dates: discovery only, not freshness proof.
- Historical METR slowdown without its 2026 update: temporally misleading.
- Sarathi/DistServe speedups presented against current vLLM: baseline mismatch.
- Herschel's inefficient-trace percentage presented as monetary savings: wrong denominator.
- Provider discounts multiplied together without compatibility checks: unsupported economics.
- A guessed Blue Yonder/Knauf source URL that did not resolve: not used to claim a current Knauf deployment.
- A provider's nationality, EU storage label or self-hosting described as proof of complete GDPR compliance: assessment scope is insufficient.
- A universal GDPR price premium or local-versus-cloud winner inferred from rate cards: no matched task measurements and full lifecycle inputs.
- Hypothetical break-even/document examples presented as actual provider benchmarks: no empirical basis.

## 5. Reading priorities

**For the founders:** S24, S01, S02, S05, S07, S13, S20, S21. This covers executive context, accountability, measurement and recent technical opportunities.

**For technical preparation:** S08, S10, S14–S19, S32 and S33. Inspect current model/platform constraints before preparing implementation examples.

**Before 29 October:** re-check live price tables, gateway budget behaviour, model availability, runtime versions, cache eligibility and any new October empirical results. Update the source date and the slide claim together; a newer model name alone is not a substantive research update.

[S04]: https://www.finops.org/insights/introducing-focus-1-4/
[S12]: https://aws.amazon.com/bedrock/pricing/
[S20]: https://arxiv.org/abs/2610.00912
[S21]: https://arxiv.org/html/2609.40247v1
[S22]: https://arxiv.org/html/2610.07094v1
[S23]: https://mlcommons.org/2026/09/mlperf-inference-v6-1-results/
[S24]: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
[S27]: https://www.tokeneconomics.com/state-of-tokenomics/
[S33]: https://claude.com/resources/articles/reducing-cost-and-improving-performance-with-claude-platform