Axioms of Research Continuation: A State-Restoration Theory for
Transcript-Suﬀicient Research Systems
Peter Bell
2026

Axioms of Research Continuation
A State-Restoration Theory for Transcript-Suﬀicient Research Systems
Peter Bell
ORCID: 0009-0004-0415-7774
New Reflexive Laboratory / Reflexive Laboratory Research Program
Draft status: v03 DOI-bearing local build / noncanonical candidate
VERSION_DOI: 10.5281/zenodo.20501949
CONCEPT_DOI: 10.5281/zenodo.20500050

Abstract
Research records may preserve enough information to replay a process without preserving enough state to resume it.
This distinction is familiar in systems contexts, where logs and checkpoints solve different problems, but it remains
underformalized in discussions of AI-assisted research infrastructure. This paper develops a state-restoration account of research continuation for transcript-suﬀicient research systems. It defines an operative research state,
an externalized record, a record projection, a replay operator, a restore operator, and a continuation-equivalence
relation based on admissible next operations. The central claim is modest: a record can support replay while
failing to support resumption when projection collapses states that are not continuation-equivalent. The paper
then specifies which state distinctions matter for research rather than generic computation: evidence-mediated
working state, re-entry and non-erasure, artifact identity, proposal/admissibility/update separation, enacted authority, and validation-gated advancement. These are organized as a tiered axiom system: a mathematical spine,
a transcript-state spine, and a governed-artifact spine. Graph representations, provenance standards, researchobject packages, process mining, checkpoint/restore mechanisms, and software configuration management are
treated as important analogues or implementation substrates, not as rivals. The contribution is not broad mathematical novelty. It is a formal semantics for research-state restoration: a way to state what a record must preserve
if research is to be continued rather than merely replayed.
Keywords: research continuation; state restoration; record projection; transcript suﬀiciency; artifact authority;
validation; canonicality; Reflexive Laboratory.

1

1. Introduction
Research continuation is a state-restoration problem. A project may leave behind transcripts, logs, files, packages,
drafts, generated artifacts, and publication records. Those records can be useful for reconstructing what happened.
They do not necessarily preserve enough state to determine what can responsibly happen next.
This distinction is not new in general systems terms. Logs, traces, checkpoints, workflow provenance, versioncontrol histories, and release packages already distinguish different kinds of recoverability. A log may support
replay. A checkpoint may support restoration. A package may preserve files and metadata. A workflow record
may support rerun. Checkpoint/restore systems such as CRIU explicitly aim to freeze a running process or
container, save its state, and later restore execution from that saved state; recent scientific-computing work
similarly describes checkpointing as stopping a long-running job and continuing it later on another resource
(CRIU Project 2025; Andrijauskas et al. 2024).
The problem addressed here is narrower than the general checkpointing problem. It asks what state restoration
requires when the object to be restored is not only a computation, but a research project with evidence, artifacts,
authority, validation, and publication status. Research continuation depends not only on process state, but
also on scholarly state: which claims are evidence-supported, which artifacts are current, which proposals have
been admitted, which objects are authoritative, which outputs have been checked, and which publication objects
function as interfaces rather than complete research objects.
The published article Replay Is Not Resumption: AutoResearch and the Architecture of Research Continuation
identified this problem conceptually. It argued that replayable traces, managed memory, provenance, packages,
generated artifacts, and process logs are valuable but insuﬀicient for accountable continuation unless composed
into governed research state (Bell 2026, Replay Is Not Resumption). The present paper supplies the formal kernel
for that claim. It does not attempt a full unified theory of the Reflexive Laboratory. It asks what record projection
must preserve if research is to be resumed rather than merely replayed.
The formal spine comes from the record-suﬀiciency model developed in P10. That paper distinguishes execution
suﬀiciency from state suﬀiciency by modeling operative state, externalized record, replay, and restoration. It shows
that a record may support replay while failing to preserve enough operative state for continuation (Bell 2026,
Execution Suﬀiciency and State Suﬀiciency). The present paper keeps that spine, but narrows its contribution. It
does not claim that the underlying projection result is mathematically surprising. It uses that elementary result
to define a research-specific restoration target.
That restoration target is continuation equivalence. A record does not need to restore every hidden detail
of a research process. It must restore enough state to preserve the same admissible next operations. If two
states license different next actions, then a record that collapses them is not suﬀicient for continuation. If two
states differ only in irrelevant details but allow the same responsible next moves, they may be equivalent for
continuation.
This shift matters because research state contains components not captured by ordinary action replay. In
transcript-suﬀicient research, current state depends on the retained corpus, evidence selected from that corpus,
working-state derivations, artifact identity, governance rules, validation records, authority states, open tasks, and
re-entry links. P7 supplies the transcript-to-state chain from corpus to indexed units, evidence, and working
state (Bell 2026, From Chat Corpus to Working State). P8 supplies proposal/admissibility/update separation
for bounded automation (Bell 2026, Bounded Autoresearch). P11 and P20 supply graph and artifact-registry
structure (Bell 2026, Research Lab as a Graph; Bell 2026, Corpus as Graph). P15 supplies enacted authority and
canonicality distinctions (Bell 2026, Canonicality Is Not Presence). P22 and P34 supply artifact-integrity and
sanity-check layers (Bell 2026, Process Transparency Is Not Artifact Integrity; Bell 2026, Sanity-Check Operators). P16 and P17 support publication-as-projection and process-legible object formation (Bell 2026, Research
as a Media Pipeline; Bell 2026, Constructive Closure and Reflexive Realization).
The paper’s contribution is therefore compositional and semantic. It identifies the research-state distinctions
that must survive record projection if a project is to be restored for continuation. It does not replace workflow provenance, W3C PROV, RO-Crate, process mining, checkpoint/restore systems, or software configuration
management. Those are close analogues and possible implementation substrates. The claim is that research continuation adds state semantics concerning evidence, artifact identity, authority, validation, publication projection,
and governed update.

2

The axiom system is tiered. Tier 1 contains the mathematical spine: record projection and continuation equivalence. Tier 2 contains the transcript-state spine: evidence-mediated working state, re-entry, and non-erasure.
Tier 3 contains the governed-artifact spine: artifact identity, proposal/admissibility/update separation, enacted
authority, and validation-gated advancement. Tiering reduces overclaiming. It makes clear which axioms support
the elementary projection result and which axioms specify the research-domain restoration target.
The paper proceeds in ten sections. Section 2 defines research continuation as state restoration. Section 3 introduces state, record projection, and continuation equivalence. Section 4 states the tiered axiom system. Section 5
presents the projection-collapse lemma and the replay/resumption corollary. Section 6 identifies the research-state
components required for governed continuation. Section 7 relates the model to existing infrastructure. Section 8
interprets the Reflexive Laboratory as a worked transcript-suﬀicient case. Section 9 states limits and nonclaims.
Section 10 concludes.
Plain-language summary: this paper does not say that logs are useless. It says that a log can be enough to show
what happened and still not be enough to continue the work. To continue research, a system must preserve what
is current, supported, proposed, accepted, checked, authoritative, and still open.

2. Research Continuation as State Restoration
Research continuation is the task of restoring enough state to continue a project under its own rules. It is not
the same as reconstructing a past sequence. It is not the same as preserving every document. It is not the same
as publishing a final artifact. Continuation requires a current operational target: what may be done next?
Let 𝑋𝑡 denote the operative research state at time 𝑡. Let AdmNext(𝑋𝑡 ) denote the set of admissible next
operations from that state. These operations may include drafting, revising, checking, integrating, validating,
registering, publishing, deprecating, or reopening a branch. Which operations are admissible depends on more
than the action history. It depends on evidence, artifact state, governance rules, validation status, and authority
relations.
The target relation is continuation equivalence.
Equation E1. Continuation equivalence
𝑋𝑡 ≡cont 𝑋𝑡′

⟺

AdmNext(𝑋𝑡 ) = AdmNext(𝑋𝑡′ ).

Two states are continuation-equivalent when they support the same admissible next operations. This relation is
weaker than equality. It is too strong to require a restored record to reproduce every hidden feature of the original
research situation. Some details may be irrelevant to continuation. But it is too weak to say that any replayable
trace is suﬀicient. If two states differ in which artifact is authoritative, which proposal is enacted, which figure
is validated, or which evidence supports the active claim, then they may license different next actions. They are
not continuation-equivalent.
This definition also clarifies the role of governance. Governance does not appear here as bureaucracy. It appears
as part of the state that determines AdmNext(𝑋𝑡 ). In P8, an automated operator may generate a proposal,
but that proposal does not update working state unless admitted through a review path (Bell 2026, Bounded
Autoresearch). In P15, a visible artifact may be present or under review without being canonical; canonicality
requires governance-visible transition (Bell 2026, Canonicality Is Not Presence). Both cases show that admissible
continuation depends on state distinctions beyond what an action log alone records.
State restoration therefore has a bounded target. The target is not perfect reconstruction. The target is restoration
up to continuation equivalence. A record is state-suﬀicient if it supports reconstruction of a state that is equivalent
for admissible continuation.
Plain-language example: if a project pauses after three candidate figures were produced, it is not enough to
know that all three figures exist. To continue, one must know whether any of them was validated, which one
the manuscript currently uses, whether one was rejected, and whether another remains under review. Those
distinctions change the next valid action.

3

3. State, Record Projection, and Continuation Equivalence
The operative research state is modeled as follows.
Equation E2. Operative research state
𝑋𝑡 = (𝐶𝑡 , 𝐸𝑡 , 𝑊𝑡 , 𝐴𝑡 , 𝐺𝑡 ,
𝑉𝑡 , 𝐾𝑡 , Ω𝑡 , 𝐿𝑡 ).
The tuple is intentionally compressed. An earlier planning draft included U_t and 𝑃𝑡 in the main tuple. In this
polished draft, 𝑈𝑡 is kept as part of the transcript-state derivation path rather than a top-level state component,
and publication objects are treated through a publication-projection proposition rather than a core state axiom.
This keeps the model focused on restoration.
The components are as follows. 𝐶𝑡 is the retained transcript corpus. 𝐸𝑡 is the evidence register selected from
the corpus. 𝑊𝑡 is working state derived from evidence. 𝐴𝑡 is the artifact set. 𝐺𝑡 is the governance layer. 𝑉𝑡 is
validation state. 𝐾𝑡 is authority state. Ω𝑡 is open continuation context, including unresolved tasks and active
branches. 𝐿𝑡 is the set of re-entry links from state or artifact entries back to source.
The externalized record is produced by projection.
Equation E3. Record projection
𝑅𝑡 = 𝜌(𝑋𝑡 ).
The projection 𝜌 determines what becomes durable record and what is lost, compressed, or left implicit. The
central failure mode arises when 𝜌 maps two continuation-distinct states to the same record.
Replay and restoration are different operators. Replay is represented by the replay operator.
Equation E4. Replay operator
𝑃replay (𝑅𝑡 ).
Restoration is represented by the restore operator.
Equation E5. Restore operator
𝑄restore (𝑅𝑡 ).
The replay operator returns a replayable path or execution class. The restore operator attempts to recover a
continuation-equivalent state. The two targets differ. A record can support the first without supporting the
second.
A record is resumable when the restore operator recovers a continuation-equivalent state.
Equation E6. Resumability target
𝑄restore (𝑅𝑡 ) ≡cont 𝑋𝑡 .
Execution suﬀiciency and state suﬀiciency can now be stated as follows.
Equation E7. Execution suﬀiciency
ExecSuff(𝑅𝑡 )

⟺

Equation E8. State suﬀiciency

4

𝑅𝑡 → 𝑃replay (𝑅𝑡 ).

StateSuff(𝑅𝑡 )

⟺

𝑄restore (𝑅𝑡 ) ≡cont 𝑋𝑡 .

The important shift is that state suﬀiciency is not defined as perfect memory. It is defined as restoration suﬀicient
for governed continuation.
Table 1. Notation Table
Symbol

Meaning

Role

𝑋𝑡
𝑅𝑡
𝜌
𝑃replay
𝑄restore
AdmNext
≡cont
𝐶𝑡
𝑈𝑡
𝐸𝑡
𝑊𝑡
𝐴𝑡
𝐺𝑡
𝑉𝑡
𝐾𝑡
Ω𝑡

operative research state at time t
externalized record at time t
record projection
replay operator
restore operator
admissible-next-operation function
continuation-equivalence relation
retained transcript corpus
indexed transcript units
evidence register
working state
artifact set
governance state
validation state
authority state
open continuation context

𝐿𝑡
Δ
𝛼
𝜈
𝜂

re-entry links
governed update operator
authority function
validation function
enactment relation

restoration target
projected record
possible site of state collapse
reconstructs or reruns a path
attempts state restoration
defines continuation target
restoration equivalence
source archive
derivation layer
admissible support
operational state
governable objects
rules and transitions
checking status
role-relative authority
active branches and unresolved
tasks
paths back to source
accepted state change
assigns artifact authority
assigns validation state
records authority transition

Plain-language summary: 𝑋𝑡 is the current condition of the research project. 𝑅𝑡 is what the record preserves.
Replay asks what can be rerun from the record. Resumption asks whether the project state can be restored well
enough to continue.

4. A Tiered Axiom System
The axiom system is organized into three tiers. The tiers are not ornamental. They prevent the manuscript from
pretending that all axioms have the same status. Tier 1 contains the elementary mathematical spine. Tier 2
specifies transcript-state requirements. Tier 3 specifies governed-artifact requirements.
Table 2. Tiered Axiom Set
Tier

Axiom

Formal core

1

A1 Record
Projection
A2 Continuation
Equivalence

𝑅𝑡 = 𝜌(𝑋𝑡 )

A3
Evidence-Mediated
Working State

𝐶𝑡 → 𝑈𝑡 → 𝐸𝑡 →
𝑊𝑡

1

2

AdmNext equality

5

Plain-language
meaning
record is a
projection of state
resumption
preserves admissible
next moves
working state is
evidence-derived

Scope
general
general

transcript-suﬀicient

Tier

Axiom

Formal core

2

A4 Re-entry and
Non-erasure
A5 Artifact Identity

𝑥 → 𝐶𝑡

3
3

3
3

Plain-language
meaning

identity fields

A6 ProposalAdmissibilityUpdate
A7 Enacted
Authority
A8
Validation-Gated
Advancement

proposal not update

enacted transition
validation gate

Scope

entries preserve
source path
artifacts must be
identifiable
suggestions do not
change state alone

transcript-suﬀicient

authority changes
visibly
high-reliance
artifacts require
checks

governed-artifact

artifact-governed
automation-bearing

trustworthy
continuation

Tier 1: mathematical spine
A1. Record Projection Axiom
There exists an operative research state X_t and an externalized record R_t such that Equation E3 holds. The
record is a projection of the research state, not necessarily the full state. This axiom creates the formal site where
state distinctions can be preserved or lost. Without it, projection collapse cannot be stated.
A2. Continuation Equivalence Axiom
Two states are equivalent for continuation when they license the same admissible next operations, as stated in
Equation E1. The restored state must support the same responsible next moves. This defines the restoration
target without requiring exact equality of all hidden details.

Tier 2: transcript-state spine
A3. Evidence-Mediated Working State Axiom
Working state is derived from admissible evidence selected from retained corpus.
Equation E9. Transcript-to-state derivation
𝐶𝑡 → 𝑈𝑡 → 𝐸𝑡 → 𝑊𝑡 .
For each load-bearing state entry 𝑤 ∈ 𝑊𝑡 , there exists supporting evidence.
Equation E10. Evidence support
∀𝑤 ∈ 𝑊𝑡 ,

∃𝑒 ∈ 𝐸𝑡

with

supports(𝑒, 𝑤).

The current working state must be evidence-mediated, not free summary. This axiom separates memory from
operational working state. P7 supplies the source model for transcript-to-state derivation (Bell 2026, From Chat
Corpus to Working State).
A4. Re-entry and Non-erasure Axiom
For each load-bearing state entry or artifact, there exists a re-entry path back to source.
Equation E11. Re-entry
∀𝑥 ∈ 𝑊𝑡 ∪ 𝐴𝑡 ,

∃ℓ ∈ 𝐿𝑡

6

such that

ℓ ∶ 𝑥 → 𝐶𝑡 .

Governed updates preserve prior trace or explicitly mark deprecation, supersession, or correction. Important
state and artifact entries must preserve a path back to source. This makes transcript suﬀiciency reconstructable
rather than decorative.

Tier 3: governed-artifact spine
A5. Artifact Identity Axiom
Every state-bearing artifact has identity, type, lineage, and status.
Equation E12. Artifact identity
∀𝑎 ∈ 𝐴𝑡 ,

∃ (id(𝑎), type(𝑎), lineage(𝑎), status(𝑎)).

A research system cannot govern or validate artifacts it cannot identify. P11 and P20 provide the artifact graph
and directed artifact-representation basis for this axiom.
A6. Proposal-Admissibility-Update Separation Axiom
Let 𝑜 be a candidate-generating operator. Generated proposals do not automatically change state.
Equation E13. Proposal is not update
𝑜(𝑊𝑡 , 𝐸𝑡 ) = 𝑝

⇏

Δ(𝑋𝑡 , 𝑝) = 𝑋𝑡+1 .

⇒

Admit(𝑝, 𝑋𝑡 , 𝐺𝑡 ).

For load-bearing updates, admission is required.
Equation E14. Admission condition
Δ(𝑋𝑡 , 𝑝) = 𝑋𝑡+1

A proposal is not an accepted update. This prevents generated material from becoming state merely by being
produced. P8 supplies the bounded-operator model behind this axiom.
A7. Enacted Authority Axiom
Artifact authority changes only through governance-visible enactment.
Equation E15. Enacted authority
𝛼(𝑎, 𝑟, 𝑡 + 1) ≠ 𝛼(𝑎, 𝑟, 𝑡)
⇒ ∃𝑔 ∈ 𝐺𝑡 such that 𝜂(𝑔, 𝑎, 𝛼(𝑎, 𝑟, 𝑡), 𝛼(𝑎, 𝑟, 𝑡 + 1)).
Authority is enacted, not inferred from visibility, recency, polish, or inclusion. This axiom preserves the distinction
between presence and canonicality. P15 is the source basis.
A8. Validation-Gated Advancement Axiom
Let 𝐷𝑡 be the diagnostic-axis set at time 𝑡, and let 𝑠high denote a high-reliance target state such as released,
canonical, or publication-facing. High-reliance advancement requires validation state.
Equation E16. Validation-gated advancement
Advance(𝑎, 𝑠high )
⇒ 𝜈(𝑎, 𝐷𝑡 ) ∈ Allowed(𝑠high ).

7

Artifacts should not advance into high-reliance state without recorded checking. This axiom separates generation
from validation. P22 gives the artifact-integrity problem; P34 supplies sanity-check operators as a candidate
operational layer.
Plain-language summary: A1 and A2 say what restoration means. A3 and A4 say how transcript-derived state
remains evidence-based and re-enterable. A5 through A8 say how artifacts become governable, reviewable, authoritative, and checkable.

Figure 1. Record Projection and State Collapse
replay target preserved

X_t
operative research state

X'_t
continuation-distinct state

R_t
shared record

rho

P_replay(R_t)
replayable path

restore target underdetermined

Q_restore(R_t)
ambiguous restoration

rho

X_t not equivalent_cont X'_t

Figure 1: Record Projection and State Collapse. A record may preserve enough information to replay a path while
collapsing states that differ for continuation. State-suﬀicient records must discriminate continuation-relevant distinctions.

5. Projection Collapse and the Replay/Resumption Corollary
The central result is elementary but useful. It follows from non-injective projection over continuation-relevant
distinctions.

Lemma 1. Record Projection Collapse
Assume that two research states project to the same record.
Equation E17. Projection collapse
𝜌(𝑋𝑡 ) = 𝜌(𝑋𝑡′ ) = 𝑅𝑡 .
Assume also that they are not continuation-equivalent.
Equation E18. Continuation distinction
𝑋𝑡 ≢cont 𝑋𝑡′ .
Then the shared record is not state-suﬀicient.
Equation E19. Record projection collapse lemma
𝜌(𝑋𝑡 ) = 𝜌(𝑋𝑡′ ) ∧ 𝑋𝑡 ≢cont 𝑋𝑡′

⇒

¬StateSuff(𝑅𝑡 ).

Proof sketch: if the same record is compatible with two states that license different admissible next operations,
then the record alone cannot determine which continuation state should be restored. State suﬀiciency requires
restoration up to continuation equivalence. Therefore the record is not state-suﬀicient.
This lemma is not deep mathematics. It is the projection-collapse mechanism that makes the research-continuation
problem precise.

8

Corollary 1. Replayability Does Not Imply Resumability
Assume the conditions of Lemma 1 and assume the replay operator is defined. Then replayability does not imply
resumability.
Equation E20. Replayability does not imply resumability
Replayable(𝑅𝑡 ) ⇏ Resumable(𝑅𝑡 ).
Proof sketch: replayability requires that the record determine a replayable path or execution class. Resumability
requires that the restore operator recover a continuation-equivalent state. Since the same record may collapse
continuation-distinct states, replayability may hold while resumability fails.
This is best presented as a corollary or design theorem, not as a surprising mathematical discovery. Its value is
methodological. It identifies why action logs, transcripts, memory entries, and provenance traces may still fail as
continuation records.

Design Criterion 1. Continuation-Discriminating Projection
Let 𝒳 be a target class of research states. If a record projection is state-suﬀicient for 𝒳, then it must preserve
continuation-relevant distinctions.
Equation E21. Continuation-discriminating projection criterion
StateSuff𝒳 (𝑅𝑡 ) ⇒ ∀𝑋, 𝑌 ∈ 𝒳,

𝜌(𝑋) = 𝜌(𝑌 )
⇒ 𝑋 ≡cont 𝑌 .

This is the main design criterion. A state-suﬀicient record projection need not preserve all distinctions. It must
preserve distinctions that change admissible next action.
This criterion is more useful than the headline corollary. It tells system designers what to ask: does the record
distinguish states that require different next moves? If not, the record is not suﬀicient for continuation.
Table 3. Theorem / Proposition Register
ID

Type

Statement

Status

Limit

T1

Lemma

core formal lemma

elementary

T2

Corollary / design
theorem

headline result

replay remains
useful

T3

Design criterion

main practical
criterion

target-relative

P1

Proposition

derived

P2

Proposition

P3

Proposition

P4

Proposition

projection collapse
prevents state
suﬀiciency
replayability does
not imply
resumability
state suﬀiciency
requires
continuationdiscriminating
projection
memory does not
imply working state
provenance does not
imply authority
presence does not
imply canonicality
generation does not
imply validation

memory can be
extended
provenance may
inform authority
artifact-dense
systems
proof-carrying
exceptions

9

derived
derived
derived

ID

Type

Statement

Status

Limit

P5

Proposition

proposal does not
imply enactment

derived

P6

Proposition

projection claim

P7

Design proposition

publication does not
exhaust research
object
governed state
supports bounded
continuation

governed
auto-admission
possible
simple papers may
approximate object

synthesis

no truth/optimality
guarantee

Plain-language summary: the core lemma is simple. If two different project states require different next actions
but leave the same record, the record cannot tell you how to resume. That is the formal meaning of “replay is
not resumption.”

6. Research-State Components: Evidence, Artifacts, Authority, Validation, and Re-entry
The projection criterion says that records must preserve continuation-relevant distinctions. The next question is
which distinctions matter for research systems.

6.1 Evidence-mediated working state
In transcript-suﬀicient research, the current working state should be derived from evidence rather than from
ungrounded memory or summary. P7 supplies the formal chain in Equation E9. This chain matters because
transcript corpora are too large to function directly as working state and too valuable to erase through uncontrolled
summarization. Evidence mediation creates a layer between archive abundance and operational state.
Let 𝑀𝑡 denote a managed memory layer at time t. A memory layer may preserve or organize information, but
working state requires operational selection and evidence support. Therefore memory does not imply working
state.
Proposition equation P1. Memory does not imply working state
𝑀𝑡 = 𝑀𝑡′ ⇏ 𝑊𝑡 = 𝑊𝑡′ .

6.2 Artifact identity
Research continuation depends on artifacts. A later user must know which draft, figure, table, registry, package,
transcript, or publication object is being referenced. P11 supplies registry-state logic, and P20 supplies directed
artifact graph representation.
Artifact presence is not enough. The system needs identity, type, lineage, and status. Without those, artifacts
cannot be governed, validated, or restored into a continuation-equivalent state.

6.3 Authority and canonicality
Authority determines whether an artifact may be relied on for a role. P15’s core claim is that canonicality is
not presence. An artifact can exist, be registered, be reviewed, or be included in a package without becoming
authoritative. Authority changes through enactment.
Authority can be represented as follows, where a is an artifact, r is a role, and k is an authority state.
Equation E22. Authority function

10

𝛼(𝑎, 𝑟, 𝑡) = 𝑘.
Canonicality can be treated as a special authority state.
Equation E23. Canonical authority state
𝛼(𝑎, 𝑟, 𝑡) = canonical.
The separate symbol 𝜅 is unnecessary in the main text. It can be defined in an appendix if needed.
Let prov(𝑎, 𝑏) denote a generic provenance relation from artifact a to artifact b. Provenance does not imply
authority.
Proposition equation P2. Provenance does not imply authority
prov(𝑎, 𝑏) ⇏ 𝛼(𝑎, 𝑟, 𝑡) = canonical.
Presence also does not imply canonicality.
Proposition equation P3. Presence does not imply canonicality
𝑎 ∈ 𝐴𝑡 ⇏ 𝛼(𝑎, 𝑟, 𝑡) = canonical.

6.4 Validation state
Validation state records whether an artifact has passed specified checks. It does not guarantee truth. P22
establishes the non-equivalence between process transparency and artifact integrity. P34 turns this into an
operator layer: sanity-check procedures for claims, methods, equations, citations, figures, tables, reproducibility
conditions, and corpus relations.
Validation may be represented as follows.
Equation E24. Validation function
𝜈(𝑎, 𝐷𝑡 ) ∈ 𝑉𝑡 .
Generation does not imply validation.
Proposition equation P4. Generation does not imply validation
𝑎 = 𝑜(𝑋𝑡 )

⇏

𝜈(𝑎, 𝐷𝑡 ) = pass.

A system may produce a fluent manuscript, plausible table, or attractive figure without checking whether the
artifact is structurally reliable.

6.5 Proposal, admissibility, and update
P8 provides the key distinction. Operators may propose, but proposals do not update state unless admitted.
Proposition equation P5. Proposal does not imply enactment
𝑝 ∈ Proposal𝑡

⇏

Δ(𝑋𝑡 , 𝑝) = 𝑋𝑡+1 .

Proposal does not imply enactment. This is essential in automation-bearing systems because generated outputs
can otherwise mutate state without review.

11

6.6 Publication as interface
Publication projection is defined as follows.
Proposition equation P6. Publication projection
𝑃𝑡 = 𝜋pub (𝑋𝑡 ).
In general, the publication object does not exhaust the research object. P5 treats the paper as a reader-facing
projection of a deeper stack of transcripts, registers, rendered objects, containers, and continuity records. P16
models research as a media pipeline in which publication render, formal object, and trace are distinct layers.

6.7 Re-entry
Re-entry links preserve the path back from state or artifact to corpus. Without re-entry, a restored state may be
usable but not auditable. P7 treats re-entry as a structural requirement, and P17’s constructive-closure criterion
also depends on governed, reconstructable transformation paths.
Plain-language summary: research-state restoration is not generic checkpointing. It requires preserving the
distinctions that determine scholarly continuation: evidence, artifacts, authority, validation, proposal status,
publication interface, and source re-entry.

Figure 2. Tiered Axiom System
Tier 1: mathematical spine
A1
Record Projection

A2
Continuation Equivalence

Tier 2: transcript-state spine
Theorem family:
projection collapse;
replay not resumption;
continuation-discriminating records

A3
Evidence-Mediated
Working State

A4
Re-entry and
Non-erasure

Tier 3: governed-artifact spine
A5
Artifact Identity

A6
Proposal / Admit /
Update Separation

A7
Enacted Authority

A8
Validation-Gated
Advancement

formal spine

Research-state restoration semantics

Figure 2: Tiered Axiom System. The axiom set separates the mathematical spine from transcript-state and governedartifact requirements. Tiering prevents the elementary projection theorem from being overclaimed as a complete theory of
research infrastructure.

7. Relation to Existing Infrastructure
The axiom system should not be presented as a replacement for existing infrastructure. Its best use is as a
semantics layer that clarifies what existing infrastructures must preserve when the target is research continuation.
Workflow provenance already records important execution histories, inputs, outputs, and dependencies. Workflow
Run RO-Crate, for example, is described by its working group as an extension of RO-Crate and Schema.org for
capturing workflow execution provenance at different granularities and bundling associated products such as
inputs, outputs, and code; the associated 2024 PLOS ONE article is Recording provenance of workflow runs with
RO-Crate (Leo et al. 2024). The present paper treats workflow-run provenance as a strong substrate for replay
and partial restoration, while adding that scholarly continuation also requires authority state, validation state,
proposal status, evidence support, and publication-interface status.
W3C PROV is also a close substrate. PROV defines provenance as information about entities, activities, and
people involved in producing a data item or thing, and its family of documents supports interoperable provenance
exchange through models and serializations (Groth and Moreau 2013). The present paper does not replace PROV.
It says that provenance alone does not decide which artifact is authoritative, which recommendation was enacted,
which object passed validation, or which next action is admissible.
12

Figure 3. Layer Ladder as State Refinement
Raw source /
retained corpus
index / compile

Managed memory

record actions

Replayable trace

restore state

Evidence-mediated
working state
assign identity

Identified governed
artifact
check / validate

Validation-gated
research object
enact / publish

Enacted authority /
publication interface

Figure 3: Layer Ladder as State Refinement. Each layer adds structure not guaranteed by the previous layer. The ladder
expresses non-equivalence claims rather than a mandatory universal workflow.

13

RO-Crate is a research-object packaging substrate. RO-Crate Metadata Specification 1.2 is the newest release,
published on 2025-06-04, and its specification includes sections for data entities, contextual entities, provenance of
entities, profiles, workflows and scripts, and RO-Crate JSON-LD (Sefton et al. 2025). A governed research-state
package could plausibly be encoded using RO-Crate-like structures. But the package format does not itself define
the restoration semantics proposed here.
Process mining provides tools for event logs, conformance checking, process discovery, and operational support.
The IEEE Task Force on Process Mining describes the Process Mining Manifesto as intended to promote processmining maturity and improve process design, control, and support (IEEE Task Force on Process Mining 2012). A
transcript-suﬀicient research system could emit process-minable events such as capture, segment, select evidence,
propose, review, validate, enact, publish, deprecate, and restore. Process mining can help analyze behavior. It
does not by itself define artifact-authority semantics.
Checkpoint/restore systems are the closest mathematical analogue. The paper’s central result is not novel relative
to checkpointing. A log is not a checkpoint; replay is not restoration. CRIU’s documentation states that it can
freeze a running application or container, checkpoint its state to disk, and later use the saved data to restore the
application and run it as it was at the time of the freeze (CRIU Project 2025). The contribution here is to specify
what checkpoint-like restoration means for research state rather than process state alone.
Software configuration management is the strongest practical analogue. Git’s oﬀicial documentation describes
commits as objects pointing to snapshots of staged content and parent commits, and it explains that branch
switching can revert files in a working directory to the snapshot a branch points to (Chacon and Straub 2014).
Pull requests, review, tests, merge, tags, releases, and issue histories already model many distinctions similar
to proposal, validation, enactment, and release. The present paper should therefore treat software configuration
management as a close analogue, while noting that research systems add evidence, scholarly authority, publication
interface, transcript suﬀiciency, and epistemic validation constraints.
Table 4. Comparator Infrastructure Matrix
State-restoration
gap

Infrastructure

Preserves well

Overlap

Workflow
provenance

executions, inputs,
outputs,
dependencies
entities, activities,
agents, derivations

replay/provenance

scholarly authority
and validation state

substrate

lineage

provenance
substrate

package, metadata,
workflow-run
provenance
event logs,
conformance,
process discovery
saved state and
process restoration
snapshots, branches,
history,
review/release
analogues
corpus, evidence,
artifacts,
governance,
validation,
authority

research object
record

admissible next
action and
canonicality
semantics of
restoration not
intrinsic
artifact authority
not intrinsic

W3C PROV

RO-Crate /
Workflow Run
RO-Crate
Process mining

Checkpoint/restore
Software
configuration
management
Reflexive
Laboratory

event-layer analysis

closest formal
computational state,
analogue
not scholarly state
proposal/review/releaseevidence and
analogues
scholarly authority
semantics need
domain modeling
worked
not universal
interpretation

Use in paper

package substrate

instrumentation

analogue
practical analogue

case

Plain-language summary: existing infrastructure is not the enemy of this paper. It is the implementation landscape. The paper’s contribution is to say what state semantics those systems need when the object is research
14

continuation.

8. Worked Interpretation: Reflexive Laboratory
The Reflexive Laboratory is a worked transcript-suﬀicient case. It should not be treated as the only possible
implementation of the axiom system.
The mathematical spine is already present in P10. That paper defines operative state, record projection, replay,
restoration, execution suﬀiciency, and state suﬀiciency. It also states that a record may be replayable without
being resumable when the projection collapses state.
The transcript-state spine is present in P7. It defines a retained corpus, indexed transcript units, admissible
evidence, working state, re-entry links, and an update operator.
The proposal/update spine is present in P8. Bounded autoresearch operators can surface evidence, detect discrepancies, flag low-confidence gaps, generate state-update proposals, and check bounded consistency. But they
do not replace human admissibility judgment or silently update working state.
The artifact-identity spine is present in P11 and P20. P11 treats the research lab as a governed artifact graph
and uses the Master Global Artifact Index to record artifact identities, statuses, dependencies, and authority
relations. P20 treats the corpus as a directed artifact graph with papers, transcripts, and support artifacts linked
by typed relations.
The authority spine is present in P15. That paper distinguishes source presence, registration, proposal/review,
enactment, and canonicality. It argues that artifact authority depends on governance-visible transition rather
than mere inclusion in an archive or transcript.
The validation spine is present in P22 and P34. P22 distinguishes process transparency from artifact integrity;
P34 proposes sanity-check operators as an operational layer before stabilization or canonicalization.
The publication-projection spine is present in P5 and P16. P5 treats the paper as one interface to a deeper
research object, while P16 models research as a media pipeline in which source, intermediate knowledge, formal
object, publication render, and trace are distinct layers.
In this interpretation, the Reflexive Laboratory satisfies the axiom system as a candidate architecture. It preserves
record projection and state-suﬀiciency distinctions; derives working state from transcript evidence; preserves
re-entry; identifies artifacts; separates proposal from update; enacts authority; records validation; and treats
publication as interface.
This does not prove that the system is optimal. It does not prove that its outputs are true. It does not prove
that all research should be organized this way. It only shows that the Reflexive Laboratory supplies a concrete
case where the axiom system has an interpretation.

9. Limits and Nonclaims
The paper’s main theorem family is elementary. This must be stated directly. A non-injective projection cannot
be inverted into a unique state. A replayable log is not a checkpoint. A record that collapses continuation-distinct
states is not state-suﬀicient. The paper does not claim otherwise.
The contribution is not mathematical novelty in the abstract. It is a domain-specific formalization for research
systems. The paper defines continuation equivalence as the restoration target and identifies the research-state
components that should be preserved when continuation is the goal.
The axiom system is target-relative. A small project may not require all tiers. A simple computational workflow
may only need workflow provenance and checkpointing. A manuscript-only project may not need a full artifact
graph. A transcript-suﬀicient, AI-assisted, artifact-dense research program requires more.
Authority is not truth. A canonical artifact may later be corrected, deprecated, or superseded. Canonicality
records enacted authority for a role.

15

Figure 4. Research-State Restoration Semantics
AdmNext(X_t)
admissible next operations

evidence

working state

Implementation substrates:
PROV; RO-Crate;
workflow provenance;
process mining;
configuration management

artifacts
may carry record components

R_t
externalized record

Q_restore(R_t)
restore operator

equiv_cont

X_t
continuation-equivalent
research state
authority

validation

re-entry

open context

Figure 4: Research-State Restoration Semantics. The restoration target is not the full past process but a continuationequivalent research state. Existing infrastructure can carry parts of that state, but the record must preserve the distinctions
that determine admissible next operations.

Validation is not correctness. A validated artifact has passed specified checks. It may still be wrong. Validation
state records review status, not infallibility.
Publication is not the whole research object. It is a reader-facing projection. The deeper research object may
include transcripts, evidence registers, artifact graphs, validation records, source specifications, packages, and
governance traces.
The Reflexive Laboratory is not universal. It is a worked case. Other systems may implement the same semantics
through workflow engines, semantic graphs, research-object packages, version control, process logs, or specialized
databases.
The axiom system also remains incomplete. It does not yet include a full uncertainty model, a quantitative
validation model, a full deontic logic of governance, or a category-theoretic operator algebra. Those may be later
papers. They are not necessary for the present contribution.
Plain-language summary: the paper should not oversell itself. Its durable claim is small but useful: a record for
continuing research must preserve the distinctions that change what the next valid research move can be.

10. Conclusion
This paper has repositioned the axioms of research continuation as a state-restoration theory for research systems.
The core problem is simple: a record can preserve enough information to replay a process without preserving
enough state to resume the research.
The formal spine contains two basic ideas. First, a research record is a projection of operative state:
Equation E3. Record projection
𝑅𝑡 = 𝜌(𝑋𝑡 ).
Second, restoration should be judged by continuation equivalence:
Equation E1. Continuation equivalence
𝑋𝑡 ≡cont 𝑋𝑡′

⟺

AdmNext(𝑋𝑡 ) = AdmNext(𝑋𝑡′ ).
16

From these, the projection-collapse result follows. If two states are not continuation-equivalent but project to the
same record, the record is not state-suﬀicient. If the same record still supports replay, then replayability does not
imply resumability. The mathematically elementary result becomes useful when applied to research systems.
The domain-specific claim is that research states contain more than action history. They include evidencemediated working state, artifact identity, proposal/admissibility/update status, authority, validation, re-entry,
publication interface, and open continuation context. These distinctions determine what can responsibly happen
next.
The revised axiom system is tiered. The first tier supplies the mathematical spine. The second supplies transcriptstate requirements. The third supplies governed-artifact requirements. This tiering makes the contribution more
precise and reduces overclaiming.
Existing infrastructures remain central. Workflow provenance, W3C PROV, RO-Crate, process mining, checkpoint/restore, and software configuration management are not displaced by this model. They are close analogues
and implementation substrates. The axiom system specifies the research-state semantics that such infrastructures
must carry when the goal is accountable continuation.
The Reflexive Laboratory provides a worked case because its prior papers have already developed the required
pieces: transcript-to-state derivation, bounded automation, record suﬀiciency, artifact graph governance, canonicality, media projection, constructive closure, corpus graphing, and validation operators. The present paper
composes those pieces into a compact formal kernel.
The next step is not more formal expansion. The next step is review of the built artifact and any final release
decision. The paper should remain narrow: continuation-discriminating record projection and research-state
restoration semantics.

17

Appendix Notes
Appendix A should contain the full v01/v02 axiom crosswalk. Its purpose is to preserve the reduction path from
the broader fifteen-axiom exploratory system to the tiered eight-axiom v03 system.
Appendix B should contain formal module details for the graph module, governance module, validation module,
and publication-projection module.
Appendix C should contain the expanded comparator matrix, including workflow provenance, W3C PROV, ROCrate / Workflow Run RO-Crate, process mining, checkpoint/restore, software configuration management, managed memory systems, AutoResearch loops, and the Reflexive Laboratory.
Appendix D should contain the math-rendering protocol, including equation IDs, LaTeX-safe source syntax,
rendered-page checks, symbol-definition checks, and visual PDF audit requirements.

18

References
Internal Reflexive Laboratory sources
Bell, Peter. 2026. The Reflexive Laboratory as a Scientific Computing Stack for Research: From TranscriptGoverned Inquiry to Containerized Research Objects. Zenodo. Version DOI: 10.5281/zenodo.18991940. Concept
DOI: 10.5281/zenodo.18902418.
Bell, Peter. 2026. From Chat Corpus to Working State: Machine-Readable Methods for Transcript-First Research.
Zenodo. Version DOI: 10.5281/zenodo.19020479. Concept DOI: 10.5281/zenodo.18902418.
Bell, Peter. 2026. Bounded Autoresearch: Operator-Constrained Automation in Transcript-Derived Research
Systems. Zenodo. Version DOI: 10.5281/zenodo.19026495. Concept DOI: 10.5281/zenodo.18902418.
Bell, Peter. 2026. Execution Suﬀiciency and State Suﬀiciency: Record Suﬀiciency in Reflexive Research Systems.
Zenodo. Version DOI: 10.5281/zenodo.19042930. Concept DOI: 10.5281/zenodo.18902418.
Bell, Peter. 2026. The Research Lab as a Graph: Snapshot Governance in a Transcript-Suﬀicient Research
System. Zenodo. Version DOI: 10.5281/zenodo.19045093. Concept DOI: 10.5281/zenodo.18902418.
Bell, Peter. 2026. Canonicality Is Not Presence: Artifact Authority in Transcript-Suﬀicient Research Systems.
Zenodo. Version DOI: 10.5281/zenodo.19104311. Concept DOI: 10.5281/zenodo.18902418.
Bell, Peter. 2026. Research as a Media Pipeline: Compression, Reconstruction, and the Role of Trace in
Mathematical Objects. Zenodo. Version DOI: 10.5281/zenodo.19119710. Concept DOI: 10.5281/zenodo.18902418.
Bell, Peter. 2026. Constructive Closure and Reflexive Realization: A Minimal Criterion for Process-Legible
Research Objects. Zenodo. Version DOI: 10.5281/zenodo.19142333. Concept DOI: 10.5281/zenodo.18902418.
Bell, Peter. 2026. The Corpus as Graph: Directed Artifact Representation in the Reflexive Laboratory. Zenodo.
Version DOI: 10.5281/zenodo.19153448. Concept DOI: 10.5281/zenodo.18902418.
Bell, Peter. 2026. Process Transparency Is Not Artifact Integrity: Why Automated Research Systems Need
Structural Validation. Zenodo. Version DOI: 10.5281/zenodo.19157570. Concept DOI: 10.5281/zenodo.18902418.
Bell, Peter. 2026. Sanity-Check Operators for Buggy Research in a Reflexive Laboratory. Zenodo. Version DOI:
10.5281/zenodo.19226488. Concept DOI: 10.5281/zenodo.18902418.
Bell, Peter. 2026. Replay Is Not Resumption: AutoResearch and the Architecture of Research Continuation:
Managed Memory, Governed State, and the Reflexive Laboratory. Zenodo. Version DOI: 10.5281/zenodo.20500050.
Concept DOI: 10.5281/zenodo.18902418.

External infrastructure sources
Andrijauskas, Fabio, Igor Sfiligoi, Diego Davila, Aashay Arora, Jonathan Guiang, Brian Bockelman, Greg Thain,
and Frank Wuerthwein. 2024. CRIU – Checkpoint Restore in Userspace for Computational Simulations and
Scientific Applications. arXiv:2402.05244. DOI: 10.48550/arXiv.2402.05244.
Chacon, Scott, and Ben Straub.
scm.com/book/en/v2.

2014.

Pro Git.

2nd ed. Apress / Git oﬀicial online book.

https://git-

CRIU Project. 2025. Checkpoint/Restore In Userspace Documentation. https://criu.org/Main_Page.
Groth, Paul, and Luc Moreau, eds. 2013. PROV-Overview: An Overview of the PROV Family of Documents.
W3C Working Group Note, 30 April 2013. https://www.w3.org/TR/prov-overview/.
IEEE Task Force on Process Mining. 2012. Process Mining Manifesto. In Business Process Management
Workshops, Lecture Notes in Business Information Processing 99, 169-194.
Leo, Simone, Michael R. Crusoe, Laura Rodriguez-Navas, Raul Sirvent, Alexander Kanitz, Paul De Geest, Rudolf
Wittner, Luca Pireddu, Daniel Garijo, Jose M. Fernandez, et al. 2024. Recording Provenance of Workflow Runs
with RO-Crate. PLOS ONE 19(9): e0309210. DOI: 10.1371/journal.pone.0309210.

19

Sefton, Peter, Eoghan O Carragain, Stian Soiland-Reyes, Oscar Corcho, Daniel Garijo, Raul Palma, Frederik
Coppens, Carole Goble, Jose Maria Fernandez, Kyle Chard, et al. 2025. RO-Crate Metadata Specification 1.2.
researchobject.org community. Version DOI: 10.5281/zenodo.13751027.

20

