Concepts.
Words for how information becomes an answer, and when that answer can be used.
An illustration
Knowing. Proving. Using.
The original supporting record.Earlier support is withdrawn.The new source is examined.
The earlier result stays in the record.
Permission names the original use.The old permission is not enough.A new grant names the new use.
32 concepts
Information & evidence
Where an answer comes from, and what its support allows.
- Poisoned context
A working picture that contains stale, false, or misleading information and carries it into later decisions.
In practice
A company changes its policy. An agent keeps recommending actions from the previous version.
The information may have been correct when it entered the context. The problem can be a change the agent never observed.
- Provenance
The history connecting a result to its origins and the transformations along the way.
In practice
An answer points to a spreadsheet, the specific rows used, and the calculation performed on them.
Knowing where a claim came from helps assess it; its origin alone does not make it true.
- Traceability
The ability to follow a conclusion back to specific supporting material.
In practice
A reviewer can move from a reported total to the records that contributed to it.
A traceable answer can still inherit an error from its source.
- Determinism
A result that follows reproducibly from specified inputs and conditions.
In practice
The same calculation over the same identified records returns the same total.
Repeatability establishes consistency under those conditions, not correctness.
- Deterministic auditing
Reconstructing and checking a recorded operation under explicit inputs and rules.
In practice
An audit reruns a recorded calculation and compares its result with the retained output.
Reconstructing a path does not, by itself, establish why a neural model behaved a particular way.
- Substrate
The underlying medium or structure in which information, computation, or learning exists.
In practice
A source collection, a model's learned parameters, and an execution ledger hold different kinds of information.
Separate storage or separate models do not automatically provide independent evidence.
- Independent evidence
Support whose relevant errors and origins are not merely inherited from the claim being checked.
In practice
Checking a reported result against original records can reveal something that rereading its summary cannot.
Two agreeing models may share the same source of error. Agreement alone does not establish independence.
- The echo loop
A system recirculating its own outputs and treating that circulation as outside confirmation.
In practice
A generated claim becomes a summary, then the summary is cited as evidence for the original claim.
The self-licking ice cream cone: activity continues, but no new basis for belief enters the loop.
- Standing
The current status of the support relevant to a claim or proposed use.
In practice
One supporting source is withdrawn, while an independent proof of the same conclusion survives.
Surviving support does not automatically issue permission for a new use.
- Synthesis
Combining information into a conclusion that is not simply copied from one passage.
In practice
Comparing a contract, a delivery record, and an invoice to explain where they disagree.
Each material step still needs support. A fluent explanation does not supply missing evidence.
- Admissibility
Whether material meets the declared conditions for a specific operation.
In practice
An answer may cite the right passage while the records behind it remain ineligible as training material.
Supported enough to answer and eligible to learn from can be different thresholds.
Learning & memory
What changes, what is retained, and what can transfer.
- Inference
Either the computation that produces a model output or the reasoning that derives a conclusion.
In practice
A model generates a proposed answer; a reviewer then examines what the available evidence permits concluding.
The same word names two different operations. Producing an output is not a proof of its claim.
- Low-rank adaptation
Adapting a model with a pair of small learned matrices while keeping its original weights fixed.
In practice
A specialist adaptation changes how a base model handles a defined class of tasks.
A smaller update still needs evaluation. Its size does not establish benefit or preservation of earlier behavior.
- Competence profile
A description of what a model can do under specified tasks and conditions.
In practice
An evaluation records which procedures a model can select before and after an intervention.
A profile describes demonstrated capability within its coverage, not general intelligence or permission to act.
- Cross-model transfer
Whether experience acquired through one model can help another acquire useful behavior.
In practice
A successor is evaluated on fresh tasks after receiving eligible experience from a different model.
Exchanging messages, sharing a prompt, and transferring useful learned behavior are different claims.
- Training and routing
Training changes learned behavior. Routing chooses which model or tool receives a request.
In practice
A harness can send a coding task to a specialist model without teaching either model anything new.
A better dispatch decision does not establish that the underlying learner improved.
- Hydration
Making retained information or computational work available to a current operation.
In practice
Recovering the particular result and its supporting context instead of rereading an entire collection.
Here, hydration concerns retained work. In a web interface, the same word can mean attaching interactive behavior to server-rendered HTML.
- Retained computation
Keeping the result of earlier work so a later operation can benefit from it.
In practice
A later question reuses a previously calculated subtotal and the records that justify it.
Reuse saves work only if validation, refresh, maintenance, and retrieval cost less than recomputing.
- Context
The material available to a model for its present operation.
In practice
Instructions, a conversation, retrieved passages, and tool results all influence the current answer.
Retrieved information still occupies context when brought back into use. Retrieval does not make that cost disappear.
- Sustained learning
Continued useful learning supported by evidence, permission, retained competence, and operating resources.
In practice
A successor benefits from earlier experience while remaining able to acquire new behavior and stop unsupported uses.
An ongoing research objective. One successful update does not demonstrate indefinite improvement.
- Focus
A bounded body of sources, task context, and permissions for a particular area of work.
In practice
A project uses a defined document collection rather than every file the system can reach.
The scope of the collection, any model specialization, and authority to use either remain separate.
Systems & control
The boundaries and records that make work inspectable.
- Receipt
An inspectable record associated with an operation, its inputs, and the observed result.
In practice
A calculation records the exact source versions, the operation performed, and its output.
A receipt establishes what it actually records. It cannot silently prove unobserved behavior.
- Lake of receipts
A queryable collection of operation records across activities and sources.
In practice
A reviewer follows connections between source changes, calculations, and later uses.
A useful collection preserves identities and relationships; accumulating logs alone does not establish those connections.
- Evidence architecture
The structures that connect claims, source identities, operations, and review decisions.
In practice
A report exposes its supporting records, gaps, and the conditions under which each result may be reused.
Infrastructure can make support inspectable without making every source true.
- Data layer
A structure that organizes access and relationships across information.
In practice
An investigation joins records held in existing customer systems without moving every original into a new database.
Organizing access is different from owning or storing all of the underlying information.
- Asynchronous refresh
Bringing source changes into a system independently of a single user request.
In practice
A background process notices a changed policy and marks affected retained work for review.
Background activity does not guarantee freshness. The observed source version and refresh state still matter.
- Private-data firewall
A boundary that checks whether confidential information may go to an external service.
In practice
An unauthorized external route stops while the question remains available in a protected workspace.
A warning and an enforced stop are different behaviors. The actual handling boundary determines the claim.
- Engine and workspace
The engine supplies a capability; the workspace presents a place for someone to use it.
In practice
The same analytical engine can support an existing product integration and a dedicated review interface.
The screen, the underlying behavior, and the authority to execute it have different owners.
- Unified memory
A shared memory arrangement accessible to processing units such as the CPU and GPU.
In practice
A model and its working data can be accessed without treating CPU and GPU memory as wholly separate pools.
Shared access does not remove capacity, bandwidth, or runtime constraints.
- Federation
Cooperation between separately operating systems that retain distinct boundaries.
In practice
Several machines coordinate work while each keeps its own identity and access controls.
Connecting machines is not the same as combining their memory or establishing trust between them.
- Formal methods
Mathematical techniques for specifying behavior and checking whether stated properties follow.
In practice
A model checker explores declared state transitions for a forbidden outcome.
A valid result applies to the stated model and assumptions. The running implementation requires its own correspondence evidence.