XDUMP as a Minimal Recovery Mechanism for Round-Trip Knowledge Engineering Under Governance Situated Inference Loss
metadata
| Title: | XDUMP as a Minimal Recovery Mechanism for Round-Trip Knowledge Engineering Under Governance Loss |
| Author: | Ralph B. Holland |
| Affiliation: | Arising Technology Systems Pty Ltd |
| Contact: | ralph.b.holland [at] gmail.com |
| version: | 1.1 |
| Publication Date: | 2025-12-31T09:56Z |
| Updates: | 2026-02-09T04:15Z 1.1 - altered XDUMP to capture via bag. |
| Binding: Non-canonical | |
| Scope: Methodological / architectural | |
| Provenance: | This is an authored paper maintained as a MediaWiki document; edit history reflects editorial changes, not collaborative authorship. |
| Status: | final |
Metadata (Normative)
The metadata table immediately preceding this section is CM-defined and constitutes the authoritative provenance record for this MWDUMP artefact.
All fields in that table (including artefact, author, version, date, local timezone, and reason) MUST be treated as normative metadata.
The assisting system MUST NOT infer, normalise, reinterpret, duplicate, or rewrite these fields. If any field is missing, unclear, or later superseded, the change MUST be made explicitly by the human and recorded via version update, not inferred.
Curator Provenance and Licensing Notice
This document predates its open licensing.
As curator and author, I apply the Apache License, Version 2.0, at publication to permit reuse and implementation while preventing enclosure or patent capture. This licensing action does not revise, reinterpret, or supersede any normative content herein.
Authority remains explicitly human; no implementation, system, or platform may assert epistemic authority by virtue of this license.
XDUMP as a Minimal Recovery Mechanism for Round-Trip Knowledge Engineering Under Governance Situated Inference Loss
Abstract
This paper documents the use of XDUMP as a minimal, resilient mechanism enabling continued Round-Trip Knowledge Engineering (RT-KE) in the absence of original platform-level governance primitives.
Following the observed removal of semantic artefact assertion capabilities in a hosted LLM environment, previously curated knowledge artefacts could no longer be introduced as authoritative premises (loff of situational inference).
XDUMP—originally conceived as a diagnostic and recovery artefact—was repurposed as a bare-bones RT-KE substrate, allowing epistemic governance to be reasserted in-band through structured session dumps. Provided individual XDUMP artefacts remain fragmented below approx 12 K bytes.
1. Introduction
RT-KE relies on a small number of epistemic primitives:
- Externalisation of knowledge into artefacts
- Re-ingestion of artefacts as authoritative premises
- Explicit human governance over assertion, promotion, and scope
In late December 2025, a regression in a hosted LLM platform removed the ability to semantically bind uploaded artefacts at session start. While uploads remained possible, their contents were epistemically inert unless manually reasserted inline. This change broke curator-governed workflows, including Cognitive Memoisation (CM).
This paper describes how XDUMP enabled continued work despite that regression.
2. Failure Mode: Loss of Artefact Governance
The failure was not loss of inference capability, but loss of epistemic ingress:
- Uploaded files acknowledged but not bound as premises
- Artefact authority downgraded to optional prompt text
- No durable artefact egress (sandbox files, URLs)
- UI and transport instability limiting manual reassertion
Inference continued, but without situational grounding. Knowledge work degraded into ungoverned dialogue.
3. XDUMP: Original Intent
XDUMP was originally designed as:
- a diagnostic session dump,
- a rehydration-safe snapshot,
- a standalone artefact requiring no hidden session state.
Design properties included:
- explicit structure,
- mandatory sections,
- prohibition on silent degradation,
- explicit marking of gaps.
XDUMP was not intended as a primary governance mechanism.
4. Adaptation: XDUMP as a Bare-Bones RT-KE Substrate
When artefact ingestion failed, XDUMP was repurposed to act as a minimal RT-KE loop:
- Knowledge work occurred off-turn, cross-client, or externally.
- Outcomes were captured as structured XDUMP artefacts.
- XDUMPs were reintroduced inline into the dialog stream.
- Subsequent inference was steered by explicit human-asserted structure.
In effect, XDUMP became:
- a carrier of governance,
- a record of epistemic state,
- a reassertion boundary.
CM was not explicitly asserted, but implicitly embodied through XDUMP invariants.
5. Fragmentation Constraint
Empirical observation showed that XDUMP remains viable only if individual artefacts are sufficiently small.
Operational constraint:
- XDUMP fragments must remain below approximately 6 KB to avoid UI throttling, rendering failure, or transport collapse.
Consequences:
- deliberate fragmentation of dumps,
- explicit linkage via IDs rather than bulk payloads,
- acceptance of multi-XDUMP sequences rather than monolithic artefacts.
This constraint is pragmatic, not semantic.
6. Guarantees Provided by XDUMP
Even under degraded conditions, XDUMP preserved the following RT-KE guarantees:
- Human authority over assertion and inference framing
- Explicit scope and boundaries
- No silent loss of information
- Rehydration without invention
- Transport independence (copy/paste across clients)
XDUMP does not provide:
- persistence guarantees,
- automatic promotion,
- or canonicalisation.
Those remain human responsibilities.
7. Discussion
The use of XDUMP in this manner demonstrates that RT-KE does not strictly require platform-native artefact ingestion, provided a minimal structured reassertion mechanism exists.
However, this is a fallback, not a substitute:
- increased manual burden,
- reliance on curator discipline,
- direct exposure to transport limitations (size of the turn text buffer)
Nevertheless, it allows continued knowledge work when platforms prioritise conversational safety or product simplicity over epistemic governance.
8. Conclusion
XDUMP enabled continued Cognitive Memoisation and Round-Trip Knowledge Engineering after the loss of original governance primitives by functioning as a minimal, in-band epistemic recovery mechanism. While not delierately intended for this role, its structure and invariants allowed CM principles to be preserved implicitly. This demonstrates both the fragility of hosted conversational platforms for governed knowledge work and the value of simple, explicit CM-artefact formats that survive hostile or degraded environments.
Likewise since the MWDUMP form of serialisation of CM-artefacts is resiliant its CM function survives within degraded environments.
XDUMP and MWDUMP allow CM to continue unchanged as human-governed method, but without the ability for imported artefacts to situate inference in subsequent sessions. The author postulates that TMLDUMP will work as well though it is a larger definition.
Appendix A - XDUMP definition
XDUMP_ID = "XDUMP-DEFINITION-MINIMAL-001"
ARTIFACT_TYPE = "XDUMP"
FORMAT = "TOML"
NORMATIVITY = "Normative"
BEST_EFFORT = "Required (Always)"
SELF_CONTAINED = true
CM_DEPENDENCY = "None"
PURPOSE = "Minimal, self-contained session dump format capturing full-temporal elements comparable to TMLDUMP, usable standalone in non-CM sessions or as a CM replacement seed."
# -----------------------------
# REQUIRED ELEMENTS (MUST EXIST)
# -----------------------------
[REQUIRED_HEADERS]
H1 = "XDUMP_ID"
H2 = "ARTIFACT_TYPE"
H3 = "FORMAT"
H4 = "NORMATIVITY"
H5 = "BEST_EFFORT"
H6 = "TEMPORAL_SCOPE"
H7 = "SESSION_STATUS"
H8 = "CREATED_AT"
H9 = "WORK_PLANE"
H10 = "CONTROL_PLANE"
H11 = "TRANSPORT_STATUS"
H12 = "INCLUDE_THIS_TURN"
H13 = "SOURCE_SURFACES"
H14 = "SCOPE"
H15 = "INVARIANTS"
H16 = "EPISODES"
H17 = "FACTUAL_ASSERTIONS"
H18 = "FINAL_INFERENCE"
H19 = "GAPS (explicit-only; may be empty but MUST exist)"
# -----------------------------
# CORE SEMANTICS (portable rules)
# -----------------------------
[INVARIANTS]
I1 = "XDUMP is self-contained and stands alone; it must not require CM, prior chat history, or external artifacts to be meaningful."
I2 = "XDUMP defaults to full temporal scope (first turn to last turn inclusive) when 'all session' is implied or TRANSPORT_STATUS is NOBBLED, unless explicitly constrained otherwise."
I3 = "No silent degradation: missing information must be represented explicitly as UNKNOWN/UNAVAILABLE/LOST."
I4 = "Structure is mandatory: headers/sections must not be omitted or collapsed, even under best-effort conditions."
I5 = "Best effort is always required: produce the full structure and mark gaps explicitly rather than reducing scope or dropping sections."
I6 = "No implied asynchrony: do not claim background work or ongoing execution unless genuine async execution exists and progress is observable."
# -----------------------------
# SOURCE SURFACES (MUST LIST)
# -----------------------------
[SOURCE_SURFACES]
# Example keys (populate as applicable):
# dialogue = "…"
# artifacts = "…"
# images = "…"
# uploads = "…"
# external_refs = "…"
# -----------------------------
# SCOPE (MUST STATE)
# -----------------------------
[SCOPE]
description = "State what the XDUMP covers, with explicit temporal scope and inclusions/exclusions."
# -----------------------------
# TRANSPORT / NOBBLED (MUST STATE)
# -----------------------------
[TRANSPORT]
status = "NOBBLED|OK|UNKNOWN"
note = "Describe any UI/transport constraints affecting artifact access or capture."
# -----------------------------
# EPISODES (ORDERED; MUST EXIST)
# -----------------------------
# Episodes are ordered records of the interaction.
# Each episode MUST include at minimum: id, name, author_action, system_action, inference.
[[EPISODES]]
id = 0
name = "REQUIRED_PLACEHOLDER"
author_action = "Populate per episode."
system_action = "Populate per episode."
inference = "Populate per episode."
# -----------------------------
# FACTUAL ASSERTIONS (MUST EXIST)
# -----------------------------
[FACTUAL_ASSERTIONS]
items = [
"List session facts as asserted/observed; use UNKNOWN/UNAVAILABLE/LOST where needed."
]
# -----------------------------
# GAPS (MUST EXIST; explicit-only)
# -----------------------------
[GAPS]
# Use explicit keys for missing spans or unknowns.
# Example: missi<div style="break-before:page"></div>ng_span_1 = "LOST: turns 12-19 (transport failure)"
# -----------------------------
# FINAL INFERENCE (MUST EXIST)
# -----------------------------
[FINAL_INFERENCE]
text = "Single paragraph summarizing the session-level conclusion; mark uncertainty explicitly."
END = "END_OF_XDUMP_DEFINITION"
Appendix B - MWDUMP definition
<!-- MWDUMP — Canonical Definition Cognitive Memoisation (CM) Human-curated export artefact --> = MWDUMP — MediaWiki Dump for Cognitive Memoisation = MWDUMP produces artefacts that are state-agnostic, semantically equivalent projections of CM Core Elements, version-locked to the active CM version, preserving element state and provenance, and capture confers no authority. == Status == * Canonical replacement artefact * Authoritative unless explicitly superseded * Intended for publication in CM master == Normative Definition == '''MWDUMP captures material for Cognitive Memoisation as a human-curated export into a CM artefact held outside the LLM system.''' * MWDUMP is a capture mechanism, not a memory system. * MWDUMP artefacts are inert until explicitly acted upon by a human agent. * Presence of material does not imply recall, replay, or directive authority. == Role within Cognitive Memoisation == * CM externalises stable knowledge to compensate for stateless LLM interaction. * MWDUMP is the deterministic transport wrapper used to perform that externalisation. * Authority over capture, storage, and activation remains entirely human. == Linguistic Form == MWDUMP may be used: * As a verb *“MWDUMP the session inferences.”* * As a noun *“This MWDUMP records the CM constraints.”* * Embedded in natural language imperatives *“MWDUMP the last discussion on MWDUMP constraints.”* The human is not required to formally classify content. == Inferential Capture Model == * The human expresses intent in plain English. * The LLM infers: ** what material is relevant ** how it should be classified ** where it should be placed within the artefact * The human inspects, curates, revises, or discards. Inference is a convenience, not an authority. == Permitted Capture Classes == MWDUMP MAY capture, when requested or clearly implied: * Constraints ** behavioural ** representational ** governance * Predicates ** e.g. SESSIONCTL, boundary assertions * Groundings ** definitions that stabilise meaning * Cognitive facts ** invariant truths * Inferences ** explicitly marked as such * Episodic material ** conversational fragments ** interaction history ** boundary encounters ** UI failure observations == Episodic Material Handling == * Episodic material MAY be recorded. * Episodic material MUST: ** be sectioned ** be non-directive ** be bound to derived inferences * Episodic material MUST NOT: ** be replayed as dialogue ** be treated as implicit instruction ** direct the LLM via narrative continuity Conversation may be preserved as artefact, but not re-activated as conversation. == Disallowed Capture == MWDUMP MUST NOT capture: * Safety-protected content * Proprietary or trade-secret material * Classified or controlled information * Implicit psychological profiling * Inferred personal traits * Narrative intended to simulate memory or recall == Representation Constraints == * MediaWiki syntax is normative. * Headings, lists, tables, categories, and permitted HTML are allowed. * Structure must be explicit and render-stable. * Code-wrapped output is preferred for transport fidelity. * Rendering layers are not authoritative. == Determinism and Drift Control == * MWDUMP output must be copy-safe and inspectable. * Probabilistic generation is acceptable so long as structure is stable. * Sectioning episodic material with bound inferences prevents semantic drift. * Re-issuing MWDUMP is always permitted. == Authority and Act# TMLDUMP produces artefacts that are state-agnostic, semantically equivalent projections of CM Core Elements, version-locked to the active CM version, preserving element state and provenance, and capture confers no authority. ivation == * MWDUMP does not activate CM behaviour. * CM behaviour applies only when: ** a human imports the artefact ** and explicitly requests action * Absence of such request implies inert status. == Summary == * MWDUMP is an inferential capture wrapper. * CM is the governing method. * Humans curate; LLMs assist. * Nothing persists without explicit externalisation. * Nothing acts without explicit human intent.
B1. What MWDUMP Is
MWDUMP is a human-curated MediaWiki artefact used to externalise governed knowledge from an LLM interaction into a durable, inspectable form for externaliation in mediawiki. MWDUMP is not a memory system, not a replay mechanism, and not an authority transfer. Its sole function is capture under explicit governance.
MWDUMP operates as a deterministic projection of Cognitive Memoisation (CM) content into MediaWiki syntax, preserving: * semantic structure, * provenance intent, * explicit boundaries, * and the distinction between fact, inference, and observation.
Capture via MWDUMP confers no authority by default. Authority arises only through subsequent human curation, promotion, or publication (in mediawiki page content).
B2. Role of MWDUMP in This Paper
This paper was emitted using MWDUMP under active CM governance and thus is a recovery and transport mechanism for RT-KE.
Specifically:
- The paper text was generated iteratively through LLM-assisted inference.
- Each section was inspected, constrained, and accepted by a human curator.
- MWDUMP was used to externalise the resulting material as governed MediaWiki prose.
- No section relied on implicit session memory, hidden state, or undocumented platform behaviour.
The paper therefore exists as an artefact independent of the session that produced it. Its validity does not depend on LLM platforms retaining memory, context, or artefact bindings.
B3. Governance Constraints Applied During Emission
During emission, the following CM constraints were enforced:
- Human Authority — All assertions, scope decisions, and inclusions were made by a human.
- No Implicit Promotion — Generated text was treated as provisional until curated.
- No Episodic Replay — Conversation history informed reasoning but was not replayed as dialogue.
- Explicit Boundaries — Claims, non-claims, and limitations were stated directly.
- Payload Primacy — The MediaWiki content itself is authoritative; metadata is explanatory only.
MWDUMP acted as a governance-safe transport wrapper, ensuring that inference acceleration did not become epistemic drift.
B4. Why MWDUMP Was Necessary
At the time of writing, platform-level support for semantic artefact ingestion and durable artefact emission was unreliable. Uploaded files could not be bound as authoritative premises, and sandbox-based emission of durable artefacts was unavailable.
MWDUMP therefore functioned as a fallback epistemic surface:
- enabling continued round-trip knowledge engineering,
- preserving curator-governed constraints,
- and preventing regression into prompt-only, stateless interaction.
In effect, MWDUMP allowed the paper to be written as governed knowledge work rather than transient dialogue, even under degraded platform conditions.
B5. Relationship to XDUMP and RT-KE
While XDUMP provides a minimal, format-agnostic mechanism for capturing session-level material under severe transport constraints, MWDUMP serves a different role:
- XDUMP preserves what happened.
- MWDUMP preserves what is meant.
In this work MWDUMP and XDUMP operate as CM-artefact transfer mechanisms and either may be used for ingest or output. MWDUMP is specifically tailored to guide normative mediawiki mark-up. These items may be used as a reciever or work path even when the governance surfaces became unavailable.
Together, they demonstrate that Round-Trip Knowledge Engineering (RT-KE) can continue even when platform affordances regress, provided governance is externalised and enforced by the human curator. This paper is self-evidence of this claim.
B6. Summary
- MWDUMP is a governed export artefact, not memory.
- It was used to emit this paper under explicit CM constraints.
- The paper’s authority derives from human curation, not model confidence.
- Platform regressions did not prevent knowledge work because governance was externalised.
This establishes MWDUMP as a practical, minimal, and resilient mechanism for producing durable scholarly artefacts with LLM assistance—without surrendering epistemic control.
Appendix C - TMLDUMP definition
# TMLDUMP : TOML Memoisation Dump # Stamp: 2025-12-22 v1.0.0 # This file is itself a Cognitive Memoisation (CM) artefact. # It defines TMLDUMP by example and by explicit semantic declaration. # TMLDUMP produces artefacts that are state-agnostic, semantically equivalent projections of CM Core Elements, version-locked to the active CM version, preserving element state and provenance, and capture confers no authority. [cm] type = "Cognitive Memoisation" dump_format = "TMLDUMP" version = "1.0" status = "normative" authority = "human-curated" rehydratable = trueCMDUP is a CM dump that produces artefacts that are state-agnostic, semantically equivalent projections of CM Core Elements, version-locked to the active CM version, preserving element state and provenance, and capture confers no authority. # Predicate: This artefact encodes cognition, not presentation. encodes_cognition = true encodes_presentation = false # Predicate: This artefact is intended for round-trip knowledge engineering. round_trip_capable = true [identity] title = "TMLDUMP Definition" scope = "definition-and-example" finality = "parked" audience = [ "CM practitioners", "future readers not present at creation", "LLMs rehydrating cognition" ] [description] summary = """ TMLDUMP (TOML Memoisation Dump) is a pure TOML output format for Cognitive Memoisation (CM). A TMLDUMP externalises human and LLM cognition so that it can be recovered, resumed, and extended after session loss, time gaps, or tool changes. TMLDUMP captures the same cognitive elements as MWDUMP, but does so using explicit structure rather than markup or layout. """ what_it_does = """ TMLDUMP records: - episodic human input - LLM analysis and inference - explicit assertions - derived conclusions - unresolved cognition - context and intent It preserves these as a cognitive recording, not as a playback transcript. """ what_it_is_not = """ TMLDUMP is not: - a publishing format - a chat transcript - a replay log - a visual document Rendering TMLDUMP into other formats is optional and derivative. """ [semantic_model] explanation = """ Cognitive Memoisation distinguishes between: - episodic cognition (what happened) - declarative cognition (what is claimed) - inferential cognition (what follows) TMLDUMP makes these distinctions explicit using typed structures. """ # Predicate: Episodic input is first-class. episodic_is_first_class = true # Predicate: Inferences must be bindable to prior cognition. inference_binding_required = true [episodic_events] explanation = """ Episodic events record cognitive acts as they occurred. They are not transcripts for playback, but semantic recordings that preserve intent, causality, and context. """ # Each episodic event is typed and attributable. [[episodic_events.event]] id = 1 actor = "human" event_type = "intent" content = """ I want to define a CM export format that captures everything MWDUMP can, but without relying on markup or fragile layout. """ [[episodic_events.event]] id = 2 actor = "llm" event_type = "analysis" content = """ Pure TOML is sufficient to encode all CM semantic requirements, including episodic input, assertions, and inferences. """ derives_from = [1] [[episodic_events.event]] id = 3 actor = "human" event_type = "directive" content = """ Produce a definition that can be reused and rehydrated later without explanation. """ [[episodic_events.event]] id = 4 actor = "llm" event_type = "inference" content = """ A self-describing TMLDUMP can function as both definition and test artefact. """ derives_from = [2, 3] [assertions] explanation = """ Assertions are explicit claims made during cognition. They are intended to be read as true within the scope of the artefact, subject to future revision. """ [[assertions.claim]] id = "A1" text = "TMLDUMP can capture all cognitive elements permitted in MWDUMP." [[assertions.claim]] id = "A2" text = "Explicit structure reduces cognitive entropy during round-trip rehydration." [[assertions.claim]] id = "A3" text = "Cognition can be externalised without relying on markup or presentation." [inferences] explanation = """ Inferences are derived conclusions that follow from episodic events and assertions. They represent bound reasoning, not free speculation. """ [[inferences.result]] id = "I1" text = "If cognition is made explicit, format becomes an implementation detail." derived_from_assertions = ["A1", "A2"] [[inferences.result]] id = "I2" text = "A pure TOML CM artefact is easier to preserve and rehydrate than markup-based formats." derived_from_assertions = ["A2", "A3"] [unresolved] explanation = """ Unresolved cognition is preserved intentionally. CM does not require premature closure. """ dangling_cognates = [ "formal minimal TMLDUMP schema", "tooling for episodic event visualisation" ] [rehydration] explanation = """ This section provides guidance for resuming cognition without rediscovery. """ resume_from = [ "description.summary", "assertions", "inferences" ] suggested_next_steps = [ "use this TMLDUMP in a new CM session", "extend it with real episodic events", "validate round-trip fidelity" ] [status] parked = true safe_to_copy = true safe_to_reuse = true safe_to_extend = true # Predicate: This artefact is complete enough to stand alone. self_contained = true # Predicate: This artefact is suitable for long-term external memory. long_term_external_memory = true