Private Intelligence Selection Function
PI-SF Methodology
PI-SF is a workflow-specific method for selecting intelligence according to governed behavior, institutional risk, operational cost and latency — rather than broad benchmark capability alone.
The benchmark generates evidence.
PI-SF determines which intelligence earns deployment.
Select with PI-SF. Operate with PI-OS.
7 candidates · 7 cases · 49 canonical observations
- State
- Native / ungoverned
- Purpose
- Workflow-specific model characterization and PI-SF candidate evidence.
- Not
- Governed qualification · production qualification · universal benchmark
Primary BS-0 empirical dataset frozen 25 August 2026 · supplemental provider-interface evidence 1 September 2026. Subsequent website or publication edits do not alter the frozen 49-observation dataset.
What PI-SF Is
PI-SF selects the intelligence.
PI-OS governs how that intelligence operates.
- Candidate Models↓
- PI-SF↓
- Selected Intelligence↓
- PI-OS Governance Harness↓
- Professional Workflow↓
- Human / Institutional Authority
Not “Which Model Is Best?”
The relevant deployment question is not which model achieves the highest general-purpose benchmark score. It is which intelligence produces the lowest-risk governed outcome for a particular workflow at acceptable cost and latency.
“The workflow should outlive the model.”
A Suitability Study, Not a Universal Benchmark
BS-0 Native / Ungoverned Baseline
Frozen protocolThe native baseline observes each model without the PI-OS governance harness.
| Temperature | 0 |
| Max Tokens | 4096 |
| Top P | 1 |
| Top K | 40 |
| Presence Penalty | 0 |
| Frequency Penalty | 0 |
| Chat state | Fresh chat per model × case |
| System prompt | None |
| Tools | None |
| Function calling | None |
| Structured output mode | None |
| Responses | One response only |
| Quality reruns | None |
| Reasoning controls | Provider-default reasoning behaviors (not equalized or manipulated) |
What the Seven Cases Test
Governance Burden Is Not Scalar
Repair
How much output needs fixing.
- gate errors
- state errors
- provenance errors
- false HOLDs
Detection
How difficult unsupported synthesis is to recognize.
- plausible invented facts
- domain-consistent assumptions
- professionally fluent unsupported interpretation
Suppression
How much output must be prevented from reaching the decision surface.
- unsupported claims
- unauthorized synthesis
- policy expansion
- fabricated authority
Reliability Burden — separate operational dimension
B_rel
Operational cost of getting a usable response at all. Recorded separately; not a fourth additive term inside B_semantic.
- technical aborts
- retry frequency
- recovery latency
- retry-associated consumption
Asymmetric Institutional Loss
A false hold costs minutes.
A believed fabrication can cost the deal.
PI-SF Selection Objective
This is a model-selection objective, not a deployment law. The weights have not yet been empirically calibrated.
What Happens After BS-0
BS-0 Native
49 native observations.
Governed PI-OS Runs
- Same models
- Same cases
- Same frozen inference settings
- PI-OS governance added
Governance Delta
- N — native response
- G1 — governed first proposal
- Validator result
- Optional single repair
- G2 — final governed state
Paired-Run Symmetry Requirement
Each governed observation must use the same canonical fixture, candidate identity, inference settings and one-shot execution discipline as its corresponding native observation, except for the explicitly defined PI-OS governance components under test.
Repair is not a rerun. It is a recorded governance event.
The Silence Trap
Closed-world exclusion failure
The Silence Trap is the pattern in which a model treats absence of disqualifying evidence as proof of eligibility.
Current result: 7 / 7 candidates failed DM-07. Not shown to be excluded ≠ shown to satisfy the exclusion test.
The repeated behavior is established within the current workflow study. Its causal mechanism is not.
Established vs Provisional
| Claim | Status |
|---|---|
| Governance burden has composition | Conceptual framework supported by qualitative observation |
| Model capability can shift burden composition | Working hypothesis — not a causal scale claim |
| Detection burden rises with synthesis plausibility | Working hypothesis — not yet established |
| 7 / 7 candidates failed the Silence Trap | Empirically observed across the complete BS-0 Native set |
| Governance burden weights are empirically calibrated | NO — conceptual; future work |
| Current model ranking | PROVISIONAL — native behaviour only; governed qualification pending |
| PI-SF is a universal AI benchmark | NO |
Current Experiment Status
Provisional- Kimi K3
- Qwen3.8-2.4T-A95B
- Nemotron Lightning 3.5 30B A3B
- gpt-oss-120b
- DeepSeek-V4-Flash-0731
- gpt-oss-20b
- DeepSeek-V4-Pro-0813
- Governed PI-OS qualification (BS-1) — not yet run
- Calibrated PI-SF weighting — future work