Agentic analysis records
Type: types/note.md
This contract defines records used across the runtime, memory and epistemic members, and what any finding citing them may claim from its evidence. The member types that declare or cite records refer to it and add their own fields and section placement. The sources and evidence layers a record cites are declared in the boundary's Source register.
Identity and grammar
The record kinds are components (CMP-*), operative objects (OBJ-*),
routes (RTE-*), claims (CLM-*), evidenced absences (ABS-*) and
behavioral-authority paths (BAP-*). SRC-* sources belong only to the
Source register. The analyst that establishes a record declares it in its
member, with its permanent prefix: runtime has RT- (RT-OBJ-store), memory has
MEM- (MEM-OBJ-store), epistemic has EPI- (EPI-OBJ-store).
Each report type declares its prefix as record-prefix, and validating the
report checks every declaration against it. The prefix identifies the declaring analyst, not the member discussing the
record. Keep a supplied ID unchanged in citations, annotations and supersessions.
Names: one to three lowercase hyphenated words, each starting with a letter;
digits may follow. Components/objects use available source-native names.
Drop extensions; split case changes and separators; lowercase and join.
Attach digit-only fragments to the preceding word (GPT-4 → gpt4). If
longer, keep three distinguishing words. Routes, claims, absences and
behavioral-authority paths use two or three words from the label. Qualify
generic names or split parts by source path or role, never a number.
Without a source-native name, role, input or label may suggest a handle.
Names are handles, not conclusions (validated-knowledge).
Within ## Shared records, kind headings group declarations:
### Components, ### Operative objects, ### Routes, ### Claims,
### Evidenced absences, ### Behavioral-authority paths. A declaration
is one level-four heading that is exactly the record ID, #### RT-OBJ-store,
followed by Label: Short label as the first line of its body. The heading
gives the record a stable anchor, the lowercase ID (#rt-obj-store).
Prose, lists and tables do not declare records. Member types specify which
empty kind headings remain. IDs are unique across the artifact.
A record citation is a Markdown link whose text is the full ID and whose
destination is the declaring member and anchor:
[RT-OBJ-store](runtime.md#rt-obj-store), or [RT-OBJ-store](#rt-obj-store)
within the declaring member. The layout's cites names the members a role
may cite; the prompt binds it to cites. A bare ID is literal text, not a citation: write one to mention a
record without asserting anything about it. Every citation must resolve to
exactly one declaration in its destination; aliases and intervals are not
inferred. Source quotations, fenced excerpts and inline code are excluded
from citation checks. SRC-* sources stay bare IDs resolved against the
Source register; numbered SRC-* ranges are refused, so list each source ID.
An annotation, #### On [RT-OBJ-store](runtime.md#rt-obj-store), supplies
another analyst's fields on a record declared elsewhere. Its heading links
exactly one record. It does not repeat generic identity or redefine the
referent, and never annotates a record the member declares. Its location and
permitted fields come from the annotating member's type.
A record's value changes only in the report that declares it: the declaring
analyst corrects it in a correction round. The reconciliation member states
identity between declared records and never replaces a value. A conflict it
cannot settle retains both findings. A
supersession uses Amendment: [MEM-RTE-selection-route](memory.md#mem-rte-selection-route) is superseded by [RT-RTE-policy-check](runtime.md#rt-rte-policy-check),
with identity evidence; both IDs stay declared. A split supersedes the combined
record only by parts already declared in analyst members, for example
Amendment: [RT-OBJ-output](runtime.md#rt-obj-output) is superseded by [EPI-OBJ-store](epistemic.md#epi-obj-store) and [EPI-OBJ-input](epistemic.md#epi-obj-input), with identity
evidence and affected findings. Reconciliation never allocates IDs. Supersede
a combined record only when its findings are wrong once the parts are
separated; a valid container can remain alongside its parts. No ID changes referent, and
no step renames IDs or rewrites another analyst's member. Provisional
labels are local tags. Keep IDs fixed through label or finding changes. A
corrected report keeps every ID its predecessor declared, because other
reports cite them; a record whose finding no longer holds keeps its
declaration and states the corrected finding.
A corrected member comes with its author's answers: a list with exactly one
entry per blocker addressed to that member, in the refusal's order, each
starting - corrected: or - declined: and a reason, and no other line
starting -. A corrected answer requires a changed member, tested against
the previous output the attempt was handed. When every answer is declined
the member may stay byte-identical if the answers change; that completes the
attempt and accepts nothing. Repeating both the member and the answers fails
the attempt and counts toward max attempts.
When other members are supplied, a new declaration cites its closest
supplied records and states distinct identity, possible-duplicate evidence,
or no counterpart after comparison. An existing referent receives an
annotation rather than another declaration; a different prefix or label
does not establish a distinct referent. Material parts with different
checks or consumers are declared and assessed separately. A declaration
whose referent is a material part of exactly one supplied record writes
Part of: [RT-OBJ-store](runtime.md#rt-obj-store) on its own unindented line
within that declaration. The field carries exactly one record citation and
cannot name the declaring record itself.
Local and artifact checks enforce its syntax; existing artifact resolution checks its
target. Explain a parent of a different record kind beside the relation.
Semantic verification checks containment; matching kinds alone does not prove it.
A candidate that can replace a record's referent is not thereby a part of it.
The part owns its fields and status; the container keeps its identity. Containment alone requires neither possible-duplicate evidence nor supersession. A record spanning several supplied records without being part of exactly one keeps the distinct-identity comparison, naming each overlap.
When a required part is undeclared, reconciliation retains an
Unresolved conflict: citing the combined record and naming the missing part, evidence and
prevented conclusion. The report verifier addresses it to the analyst who
should declare the part. Name the undeclared part in prose,
without inventing an unresolved ID.
Status fields
Conclusion status is exactly one of absent, inapplicable, uninspected,
claimed, afforded, wired, observed, or causally supported.
Different layers have separate labelled fields, such as implementation
conclusion status: wired and operation conclusion status: observed;
wired; observed is not one value. Guarantee strength is a separate field:
invariant, protocol, policy, best effort, deployment guarantee,
or no claimed guarantee. Epistemic architectural status and observed
candidate state retain their own vocabularies.
Every route declaration has at least one conclusion-status field on its own line. For example:
- implementation conclusion status: wired
- operation conclusion status: uninspected
Write one controlled value only; keep its evidence and explanation outside the value. A field cannot come from another record, annotation or source excerpt. Member acceptance checks presence, duplicate layer labels and controlled values; semantic verification judges the choice.
Record fields
| Kind | Required content |
|---|---|
| Component / operative object | Source-native identity, representational form, storage substrate and evidence |
| Route | Endpoints, progression, owner, context/state/action effects, applicable status fields and evidence |
| Claim | Claimed operation and its source |
| Evidenced absence | absent status, searched roots or files, query and revision, evidence, and conclusion supported or prevented |
| Behavioral-authority path | Consumer, channel, force and horizon |
Representational form classifies encoding and consumption, independently of storage. Natural-language content receives consequences through model or human interpretation. Symbolic content has localized units and fixed consumer rules determining its permitted behavior. Distributed-parametric content spreads numerical state across weights or dense representations. Split mixed objects by operative part or consumption path when their evidence, invalidation or rollback differs. Prompt names model input supply, not a fourth form.
Each distributed-parametric component used by inspected routes separately states whether parameters change during operation, whether its exact version is pinned, and whether a mutable endpoint can resolve to another model. Provider internals remain uninspected when inaccessible. Fixed weights do not establish absence of learning through other retained material.
Each route records immediate return, later read-back, delegated visibility, selection predicate, invalidation or expiry, activation or effect and evidence limits. An inapplicable or uninspected field gives a reason. Read-back is material accumulated or changed through use that reaches a later consumer invocation; static shipped material and ordinary current-run state are excluded. A summary that compacts or replaces current-run state is retained memory when a later invocation receives it, within or across runs. Activation requires evidence that delivered material changed behavior.
Route field syntax
Within each route declaration, write each of these exact labels once as an unindented bullet, with a non-empty answer on the same line:
- Immediate return: ...
- Later read-back: ...
- Delegated visibility: ...
- Selection predicate: ...
- Invalidation or expiry: ...
- Activation or effect: ...
- Evidence limits: ...
Replace every ... with the finding. When a field is inapplicable or
uninspected, use exactly inapplicable — reason or uninspected — reason,
replacing reason with the explanation. Supporting prose, tables and
quotations may follow. Another record or an annotation cannot supply these
fields for the declaration.
Local member validation and artifact validation reject missing, empty or duplicate fields and uncertainty values without reasons. They check presence and form; semantic verification judges the answers and their evidence. The separate status-field check above enforces conclusion-status form. Component fixity and conditional route fields retain their requirements without new presence checks.
Source-native coverage and uncertainty
Memory and epistemic accounts retain an inventory of included material parts and mechanisms, including opaque alternatives. Record supported facts beside their actual part, route, transformation, input or consumer and evidence layer. A positive witness establishes existence, not complete enumeration. Name each unresolved included part, the missing fact, inspection/access limit and conclusion prevented. Available but uninspected evidence is not evidenced absence; inspected but inconclusive evidence is not an unsupported positive. Use the existing status vocabularies; do not invent a record conclusion status for profile coverage. Bounded absence requires a searched boundary and warrant. Inapplicability requires the relevant boundary and reason.
Source analysts describe these dimensions without assigning controlled profile values. Preserve enough source-native detail for later classification:
| Dimension | Unit and distinctions to retain |
|---|---|
| Storage | Each operative object/part, including opaque provider state; distinguish encoding from substrate. |
| Form | Each operative part and consumption path; separate mixed text, symbolic and numerical material. |
| Lineage | Each object/part and derivation path; keep unresolved initial or embedding provenance local. |
| Authority | Each retained part, actual consumer and effect; delivery is not compliance, and update and downstream consumers differ. |
| Admission control | Each write/admission mechanism: human decision supplying, editing, approving or replacing that content versus software/model admission; authorship and physical I/O are separate. Starting a workflow is not per-content approval; generic caller identity leaves human control unresolved. Reads alone do not establish writes. |
| Curation | Each implemented transformation with its evidence layer; requested behavior and route names do not establish changed meaning or a new claim. |
| Read-back direction | Request, selection and delivery operations in a chain; requested delivery and independent unsolicited supply differ. |
| Selection | Actual selector input and selected retained part; an identifier on a requested file is not evidence of targeted selection. |
| Trace-fed update | Each automatic trace-fed write, its durable behavior-shaping result and later consumer; retaining raw logs alone is insufficient. |
| Trace origin | Original input to each qualifying write, including mixed or opaque provenance; adapter labels do not establish origin. |
Several effects or transformations can coexist without resolving an opaque alternative. Group only units sharing the same scope, control and coverage. A trace-fed artifact update may give retained experience learning authority at the update consumer while its output gives a later agent knowledge; neither establishes improved capacity. Preserve these actual paths rather than a system-wide implication or score.
Faithfully scoped uncertainty is an acceptable result. Unsupported findings, concealed included parts, unjustified complete coverage, unsupported negatives and malformed references remain defects. The declaring analyst may narrow or correct an assertion, and reconciliation may retain a named conflict; no step erases supported positives to make another part's uncertainty disappear. A missing fact becomes a problem only when no faithful bounded account is possible or required inputs or frozen scope prevent completing the task. Keep learning, reflection, autonomy and self-improvement, and each theory-builder condition, as independent route/property conclusions. A supported contribution coexists with independently unestablished properties; never bundle them into a negative.
Evidence interpretation
Presence, delivery, activation and demonstrated benefit are distinct; so are claim, affordance, wiring, observation and causal support. Curation alone establishes no warrant. Every negative, thin, conflicting or uncertain finding names its inspected boundary and the conclusion it prevents. A casual search miss or uninspected branch supports a limitation, not an evidenced absence, and omission from an open-ended mechanism list is not evidence of absence.
When a conclusion depends on information reaching a consumer, establish what that consumer receives after intervening selection, transformation or omission. Do not infer completeness, correctness or successful execution from the producer's output label. Inspect these transformations where they could change the conclusion, not as an exhaustive audit of data handling.
Distinguish implemented capacity, exercised behavior and demonstrated benefit. Assess each property using the evidence required by its definition. Missing execution evidence limits claims about exercised behavior or benefit; it does not by itself make an inspectable architectural capacity indeterminable. Conversely, an implemented path does not establish that it was exercised or beneficial.
A contrast is necessary but not sufficient for causal identification: attribution is no finer than the treatment and comparison actually performed, and a component effect requires independent variation of that component.
Source-native operation precedes a Commonplace interpretation. The fit and any partial or unresolved mapping are stated; ontology never replaces the operational account.
Conditional route fields
These requirements apply to the named route classes. Inapplicability is not evidence of an absent epistemic phase.
| Applies to | Required content |
|---|---|
| Diagnosis, candidate comparison, admission or successor selection | Who proposes, decides and can veto; separate computational and human contributions; any answer oracle (expected answer or reference outcome, provider, use); operating mode and improvement triggers |
| Routes admitting product, knowledge/instruction, capability or production-machinery changes | Trigger, proposed change, admission, rejection ability, rollback/recovery; guidance that shaped the proposal, what it says and what persists |
| Routes involving formulated theories | The theory account and separately evidenced properties below |
Model judgment alone establishes no answer oracle. Operating modes distinguish open requests, bounded experiments/curricula and multiple modes; oracle use belongs to its mode. Routine logs, counters and unchanged checkpoints need no separate revision account unless they change later decisions or recovery. Writes governed by one admission mechanism may be grouped.
Theory account
A theory route states the guidance it applies, derives from, criticizes, revises/replaces, and retains or reconstructs. It names what persists: formulated theories, criticisms, input/outcome records, parameters or other source-native material. A rule set can express a theory without historical rationale; classification depends on its content and exposed assumptions, scope conditions and parts, not its storage label. Missing rationale does not establish absent formulated criticism. Retained rationale names its later reader, if any.
Each theory-builder condition has
its own conclusion status and evidence under the label
theory-builder conditions 1–4:
| Condition | Evidence required and limits |
|---|---|
| 1. Localized content | A theory stated in natural or formal language; its minimum is one whole unit carrying what the theory says. Whole replacement meets this minimum. |
| 2. Consumption | Decisions depend on what the theory says. Storage, citation and delivery alone do not establish it. |
| 3. Content-directed criticism | Stated criticism names the challenged claim, result and placement of blame (theory, test, data or auxiliary assumption), and resulting revision or changed reliance. Derivation under an unchanged theory and scores selecting variants do not establish it. A prose contradiction is an interpretation unless codified. A surviving theory can change reliance or test selection without changing text; rejection and replacement count as revision. |
| 4. Iteration | Criticism's result is kept and shapes a next round, including within one run. Reconstruction from retained criticisms counts; unused critic reports or retained input/outcome records alone do not. A builder ends when another system freezes its product for deployment. |
No condition is inferred from another. The report distinguishes absent formulation or criticism from inaccessible model processing; opacity alone establishes neither presence nor absence. Additional properties are assessed separately:
| Property | Required finding |
|---|---|
| Addressability | Degree to which assumptions, scope conditions and parts are individually inspectable and revisable, and the assessment boundary; finer grades exceed condition 1's minimum |
| Persistence | What persists, later consumer and attained horizon: reasoning episode, rounds of one run, runs on one task, or problems/sessions |
| Learning | Under learning: improved capacity for future action, assessment boundary, improvement evidence and attribution evidence. Membership, revision, persistence or connected steps alone do not establish it; capacity need not have been exercised. Later capacity traces persistence to that time; later/recurrent use also traces the retained result and consumer. The memory axis trace_learning has its own write-route meaning. |
| Reflection | Selected aspects within the boundary, their self-representation and both causal directions: aspect changes can update the representation; representation-mediated operations can affect later behavior. Subject matter alone is insufficient; direct machinery modification is unnecessary. |
| Reflective theory builder | Its method texts meet conditions 1–4 and are criticized against records of its own operation |
| Autonomous theory builder | Role-by-role evidence that computation performs every internal operation; users supplying problems and judging products are outside the boundary. Autonomy does not establish reliability. |
Self-improvement attribution
When attributing or qualifying self-improvement, assess the named pathway independently of learning, reflection and autonomy. Self-improvement is operative, evidence-responsive change to the bounded system's own behavior-determining organization, not merely improvement of an external work product. An output can enter that organization when retained and consumed in later operation; its output label alone decides neither inclusion nor exclusion.
Declare the boundary, assessment horizon and improvement objective. The objective must be specifiable independently of the change it licenses. For an exercised pathway, establish all four causal links within that boundary and horizon:
- Evidence bearing on the objective causally shapes determination of the update.
- The result changes the system's own organization, rather than remaining evidence, a proposal or an external product.
- The changed organization enters a live behavioral-authority path: consumer, channel and force capable of reaching later behavior.
- Subsequent operation exercises that path and causally depends on the change.
Storage, retrieval, acceptance, installation and loading alone do not close these links. Name an unestablished link and the attribution it prevents. A standing but dormant pathway supports only a marked dispositional claim, not exercised self-improvement over the horizon. Membership establishes improvement-directed self-change, not successful improvement; success needs separate outcome evidence. Neither reflection, autonomy nor a separate evaluator or rejection gate is a membership condition. This test governs self-improvement claims; it adds no universal assessment obligation for other routes.
Learning, reflection, autonomy and self-improvement remain independent claims. Revision selection prefers explanatory-reach among revisions that fit the evidence; it does not trade fit away for reach. Explanatory-reach means that a criticizable account of why a pattern works continues to apply beyond its originating case because the mechanism persists. Vary a load-bearing premise and ask what change the explanation predicts; transfer or local success alone does not establish that account.