ZERO01
Remove a signal to measure its contribution.
AI MISSION FOUNDRY / AMF
A working AI system is more than a model
AMF is scientific infrastructure for engineering intelligence under mission constraints. It does not reduce the problem to training a model: it searches, materializes, composes, measures and qualifies a complete realization.
CANONICAL PRINCIPLE
Capability may propagate. Authority may not.
A capability that is inherited, transferred or recomposed never automatically inherits the authority of its source.
AMF terms, in plain English
Mission
A specific problem and the results required.
Realization
The working system: models, data, tools and rules.
Qualification
One exact version passed predefined tests for a specific task.
Evidence
Test results and files that support the reported claims.
Passing one mission does not mean every other task is proven. AMF should always state exactly what was evaluated.
01 / CONSTITUTIVE LOOP
Start with a real goal. Test before you deploy.
AMF explicitly separates scientific proposal, execution, measurement, adjudication and qualification. The Foundry may propose experiments; it cannot grant itself deployment authority.
01
Mission
Objectives, constraints, budgets, evidence criteria and termination conditions.
02
Search
Bounded hypotheses and experiments across models, representations, data, knowledge, composition and runtime.
03
Materialize
Reuse, transform, train, rematerialize or compose exact candidates.
04
Measure
Measure utility, robustness, cost, collateral effects, interactions and execution integrity.
05
Qualify
Apply protected gates to a frozen realization and record a bounded verdict.
06
Operate
Execute the QIR in the IQ Fabric without silently mutating its qualified identity.
02 / SEARCH SPACE
Different tools can solve different parts of a task.
AMF can explore multiple research loci without turning them into a rigid pipeline. Operations remain versioned, reversible where possible, and bound to exact identity.
01
CHARACTERIZE
What already exists?
02
INTERVENE
What changes under controlled perturbation?
03
REPRESENT
Can the function be represented differently?
04
LEARN / RECOVER
Can observed deficits be repaired?
05
REMATERIALIZE
Can the function exist in another substrate?
06
COMPOSE
What system interactions emerge?
07
DATA / KNOWLEDGE
Is the deficit informational rather than structural?
08
DEPLOY
Which runtime envelope is justified?
03 / INTERVENTIONAL SCIENCE
Ten ways to test how an AI component affects the result.
An intervention is defined by a target, primitive, scope, intensity and measurement contract. It produces scientific information before it ever produces a permanent modification.
MASK02
Hide a head, neuron, path or block without erasing its source.
BYPASS03
Route around a component to test whether it is necessary.
REPLACE04
Substitute an activation, component or baseline.
SCALE05
Vary influence continuously instead of making a binary cut.
NOISE06
Stress a pathway and measure the robustness of the realization.
TRUNCATE07
Reduce rank, dimension or retained structure.
DEGRADE08
Lower precision or substrate quality under controlled conditions.
SWAP09
Test whether another implementation can preserve the bounded function.
FREEZE10
Constrain learning to identify where adaptation is necessary.
Historically documented interventions include HEAD_OUTPUT_ZERO, Q / K / V / QK / QKV paths and ATTN_SCALE. A measured drop indicates contribution under the tested conditions — not the universal location of a capability.
04 / EXACT OUTPUT
The finished system includes tools and rules, not just model weights.
A Qualified Intelligence Realization binds one exact computational identity to a mission contract, admissible evidence, a qualification verdict and an authority envelope. Changing routing, a verifier, a policy, a knowledge source or a component may create a new realization.
01Executable graph
02Exact component identities
03Evidence package
04Qualification record
05Deployment envelope