intelligence-radar/INTENT.md

INTENT — Intelligence Radar

Status: Adopted baseline 0.1 (2026-09-13)
Date: 2026-09-13

1. Intent

Intelligence Radar exists to improve our ability to reason about intelligence itself and to make practical decisions about where, when, and at what price machine intelligence can be applied.

The facility shall continuously answer two different classes of questions:

Epistemic questions

• What do humans mean when they use the word intelligence?
• Which models of intelligence exist across biology, cognitive science, cybernetics, neuroscience, psychology, computer science, AI, economics, organizational theory, collective intelligence, and related disciplines?
• Which claims are supported, contested, speculative, superseded, or refuted?
• Which observations or experiments would distinguish competing models?
• Where are we confusing measurement convenience with the underlying construct?

Operational questions

For a concrete task:

• What capabilities does the task require?
• Which five current benchmarks provide the most useful evidence about those capabilities, and with what relevance?
• Which available execution profile is the Frontier choice?
• Which profiles are Topnotch — at most 10% predicted degradation from Frontier?
• Which Topnotch profile is the cheapest expected solution — the Efficient choice?
• Which profiles are credibly Sufficient to satisfy the task?
• Which Sufficient profile is the cheapest — the Budget choice?
• How clear is the task?
• How confident are we in the benchmark mapping, capability estimate, availability, cost estimate, and final recommendation?
• What is the expected cost and runtime?
• Would decomposition into subtasks and collaboration among specialized models improve quality, confidence, cost, or runtime?

2. Scope

Intelligence Radar spans:

• biological intelligence;
• cognition and neuroscience;
• psychometrics and measurement;
• cybernetics, control, adaptation, and viability;
• embodied and situated intelligence;
• collective, organizational, swarm, and social intelligence;
• artificial intelligence and machine learning;
• foundation models and generative systems;
• agents and multi-agent systems;
• software and systems engineering for intelligent systems;
• AI accelerators, inference systems, networks, storage, and energy;
• capability evaluation and benchmark science;
• economics, price, availability, and accessibility of intelligence;
• safety, security, alignment, reliability, and misuse;
• economic and social effects;
• law, regulation, standards, and public policy.

The facility must not equate AI progress with progress in understanding intelligence.

3. Three planes

3.1 Intelligence Compendium

A durable, versioned epistemic graph describing concepts, theories, claims, evidence, objections, thinkers, experiments, systems, and historical development.

It answers:

What does humanity currently have reason to believe about intelligence?

3.2 Intelligence Radar

A high-frequency event and delta layer.

It answers:

What changed, and which previous beliefs, measurements, policies, capability estimates, or economic assumptions should change because of it?

3.3 Capability Advisor

A task-specific decision system.

It answers:

What intelligence should be used for this task, under these conditions, today?

4. Normative vocabulary

Model

A base trained model, e.g. a named model family/version.

Execution Profile

A model plus the conditions that materially affect measured capability:

• reasoning/effort level;
• prompting/scaffold;
• system instructions;
• agent harness;
• available tools;
• context strategy;
• retrieval;
• memory;
• sampling parameters;
• number of attempts or ensemble strategy.

Deployment

An execution profile bound to practical availability:

• provider;
• API/product;
• region;
• service tier;
• rate limits;
• context/output limits;
• quantization or serving stack;
• price;
• latency;
• data-handling constraints;
• availability date/status.

Benchmarks primarily measure execution profiles, not Platonic model capability.

Frontier

The eligible execution profile with the highest predicted task performance, regardless of cost.

Topnotch

Eligible profiles whose predicted task performance is at most 10% below Frontier under the active ranking policy.

Default policy:

task_score(profile) >= 0.90 × task_score(frontier)

Efficient

The Topnotch profile with the lowest expected total task cost.

Sufficient

A profile whose predicted probability of satisfying the task's acceptance criteria exceeds the task's declared or inferred sufficiency threshold.

Budget

The Sufficient profile with the lowest expected total task cost.

Available Frontier

The highest-performing profile the requester can actually use under jurisdiction, access, privacy, product, and deployment constraints.

5. Foundational principles

P1 — Intelligence is contextual

A capability exists relative to an environment, objective, resource budget, interface, and observation method.

P2 — Do not collapse intelligence into one scalar

Aggregate indices are useful summaries, not ontologies.

P3 — Benchmarks are instruments

Every benchmark has a construct, validity envelope, error model, saturation state, contamination risk, ecological validity, and expiration risk.

P4 — Five benchmarks are selected per task

There is no permanent canonical five. The selector chooses five task-relevant measurements while penalizing redundancy.

If evidence coverage is poor, the system still returns its best five candidates but explicitly marks the mapping as weak.

P5 — Prefer independent evidence

Five correlated benchmarks do not constitute five independent observations.

P6 — Distinguish observed fact from policy

Raw benchmark results, prices, availability observations, and task outcomes are facts with provenance.

topnotch degradation <= 10%, benchmark weighting, sufficiency thresholds, and confidence aggregation are policies and must be versioned.

P7 — Price is time indexed

Promotional rates, negotiated rates, region, provider, caching, batch/flex tiers, token lengths, and tool charges can change the economics drastically.

P8 — Cost per token is not cost per solved task

Expected cost includes retries, reasoning tokens, tool calls, cache behavior, failures, and verification.

P9 — Outcome evidence outranks leaderboard intuition

Public benchmarks seed predictions. Verified real-world outcomes on structurally similar tasks should progressively calibrate or override generic benchmark priors.

P10 — Preserve disagreement

The Compendium records competing hypotheses and unresolved contradictions rather than prematurely synthesizing them away.

P11 — Primary sources first

Secondary reporting discovers events. Durable claims should resolve to primary research, official documentation, benchmark maintainers, regulator texts, standards, or first-party measurements whenever possible.

P12 — Decomposition is an optimization problem

A task may be better solved by a workplan of heterogeneous models than by one global Frontier model.

6. Core service contract

Input:

• task/question;
• supplied context;
• constraints;
• required output;
• acceptance criteria, if known;
• risk/reliability requirements;
• environment/tool constraints;
• optional budget/runtime limits.

Output:

1. Task interpretation
2. Task clarity score and missing/ambiguous information
3. Capability profile
4. Five selected benchmarks
   • relevance weight;
   • rationale;
   • measurement quality;
   • data freshness;
   • known weaknesses
5. Candidate execution profiles
6. Frontier
7. Topnotch set
8. Efficient
9. Sufficient set
10. Budget
11. Expected cost with p50/p90 or equivalent interval
12. Expected runtime with uncertainty
13. Recommendation confidence
14. Evidence gaps
15. Optional optimized workplan
16. Comparison against monolithic execution
17. Post-execution outcome record when results become available

7. Feedback loop

Every completed task should be eligible to create a calibration observation:

• canonical task profile;
• selected benchmarks and policy versions;
• predicted performance;
• chosen execution profile;
• actual model/provider configuration;
• input/output/cached/reasoning token use;
• tool calls;
• retries;
• elapsed and active runtime;
• acceptance-test results;
• verifier/judge outcomes;
• human assessment;
• failure mode;
• environmental context.

These observations form the Task → Model Performance Ledger.

The long-term goal is:

Replace generic benchmark inference with empirically calibrated predictions for classes of real work while retaining benchmarks as transferable prior evidence.

8. Success criteria

Intelligence Radar succeeds when it can:

• explain why a recommendation was made;
• reproduce a historical recommendation from its dated evidence and policy versions;
• detect when a benchmark ceases to predict real outcomes;
• distinguish model capability from scaffolding and deployment effects;
• identify cost/performance Pareto frontiers rather than merely rank models;
• provide actionable model choices for tasks with explicit uncertainty;
• reduce cost without materially reducing task success;
• improve its own predictions from observed outcomes;
• maintain a useful, non-dogmatic map of scientific and philosophical models of intelligence.

9. Non-goals

Intelligence Radar is not:

• a universal AGI score;
• a model fan leaderboard;
• a news-content farm;
• an automatic claim that benchmark performance equals real-world competence;
• an authority deciding what intelligence “really is”;
• a system that hides uncertainty behind precise-looking numbers.

10. One-line description

Intelligence Radar continuously maps what intelligence appears to be, what intelligent systems can demonstrably do, and the cheapest reliable way to apply those capabilities to a concrete task.