Which routes deserve a company test?
Public evidence bounds the candidate set. It does not advance any API, local, seat, or assembled hybrid route.
Research publication · Edition dated 29 August 2026
The 2026 LLM System Decision Manual
A 30-page CEO decision framework that indexes 45 API, local, and seat candidates, shows how to evaluate hybrid routes, and defines the company evidence required before a workload-specific choice.
Free PDF. Evidence cutoff: 29 August 2026. Reverify current terms before use; the evidence horizon expires by 31 December 2026.
- Format
- 30 pages
- Candidate set
- 45 indexed candidates
- Evidence register
- 80 evidence records
The executive decision
Choose a system, not a leaderboard winner.
The useful comparison includes the model, access route, hardware or provider, workflow, controls, human review, and the quality bar the business actually needs.
What is worth owning?
Separate control, privacy, latency, continuity, and capability requirements from preferences that do not change the business outcome.
What must be signed?
Bind one workload, one data class, one exact configuration, a quality floor, operating limits, and accountable owners before selection.
Inside the manual
Seven comparison families, four operating-capacity postures.
Small, Medium, Big, and Huge govern candidate counts and resource ceilings; they are not company-size classifications. The system frame includes routing, retrieval, evaluation, observability, permissions, fallback, and human review.
- 01Dated tariff lens; private workload performance remains to be measured
- 02Innovation-result template with uncertainty and sensitivity
- 03Local-capacity packet for 21 artifacts; no hostability conclusion
- 04Local-resource and realized-benefit ledger
- 05Seat/API separation and frontier-service economics
- 06Eight implementation-burden measures
- 07Common-gate templates for coding, marketing, and data ingestion
Evidence boundary
Dated, inspectable, and deliberately conditional.
Public evidence defines the bounded candidate set and dated tariff inputs. It does not establish production performance, fit, savings, hostability, readiness, or approval.
This release is intentionally uninstantiated. Missing or incomparable evidence remains UNKNOWN; no system advances from public facts alone, and hybrid routes must be assembled from separately shortlisted components.
Read IMPAKT's editorial standardsMove from market noise to company evidence.
Frame the workload
Name the decision, quality threshold, data boundary, volume, latency, and failure cost.
Design the evidence
Freeze cases, gates, denominators, tariff inputs, routes, and the decision authority before testing.
Sign the shortlist
Advance only after company evidence closes quality, safety, service, economics, and operating constraints.
Formats and companions
Read the content. Work with the schemas.
Because the designed PDF is not tagged for assistive technologies, a linearized plain-text edition is available. Visual tables and writable forms are represented as text in reading order.
Open the plain-text editionExpected gate catalog
Common gates, expected sets, actual-result templates, and decision-packet templates.
SHA-256 b0ad4300cd3c8aeda4964c04672827cf6e71529483b4d820aff6bd01940a2c8e
Shortlist and frontier-fit schema
Capacity-scaled defaults, shortlist schema, blank ledgers, and fit-result templates.
SHA-256 4549b39659fbc9a46faa5747586ebd8ff06d713187f52bc959cca7477e46c00c
Decision workbook
Machine-readable posture, economics, and fit-execution workbook.
SHA-256 e7566cc74820d6df9c8d23689d8f0d9fca4735a56490e66e1a7675cd51a2731b
Publication integrity
Download PDF · 12.6 MB Verify the file you send.
SHA-256
9112a61f88dade08e67ac385e8744a51a660e500355a1e2bb3bcd4b705d7af9bThis public 30-page manual is separate from the 61-page LLM Operating Playbook delivered to newsletter subscribers.
Apply it to your company
Contact IMPAKT