Lab · conjecture
Continuity as recurrence
A conjecture from the research floor, adopted at its third outside read after two holds: that the fold’s continuity criterion can be stated as a recurrence in a dependency order, with three commitments the criterion makes, a rejection condition for each, and one question left open, and what would show it wrong.
What this is #
The conjecture #
The claim #
The paper’s second fold criterion reads: “Temporal continuity: the self-model must persist across time in a way that allows the system to relate its present operations to its past and anticipated future ones. A single-tick feedback loop does not meet this criterion.” The same part calls the fold “geometrically primitive because its definition uses relations without assuming space or time”. A time index can label an order without supplying distances, so absence of a metric and absence of time are different claims. Three commitments follow, each with the rejection condition that defeats it, and one open question.
(a) No quantitative metric among the premises. No premise assigns a magnitude: no interval carrying a scale, no rate, no count of temporal units. The threshold of two counts occurrences and not elapsed time, so R1 does not run against it; a premise counting units of anything the order does not supply would fall to R1. The weakest commitment, listed apart so the others get no credit for its cheapness.
(b) No temporal structure beyond the dependency order. No premise supplies an index independent of the order, a simultaneity relation independent of the order, or a direction the order does not already supply. A label naming an occurrence’s order position is permitted; one fixing which occurrences are at one time, or ordering incomparables, is excluded. The collapse rule below is order-derived and so permitted: it identifies occurrences that fix each other and assigns none to a time. Whether such an identification is already a simultaneity relation belongs to (d); on that reading (b) is the weaker claim it looks like.
(c) Preservation of the published criterion’s verdicts. What the published criterion rejects, the rewrite rejects; what it accepts, the rewrite accepts: the single-tick loop by name, and the two requirements its wording carries, retained history and anticipation. A rewrite stricter than what it replaces fails (c) as surely as one more permissive, since either changes which systems the filter admits.
(d) Whether a directed determination order is itself temporal. Not settled here. The order defined below is directed. (a) to (c) are consistent with reading that directedness as already temporal structure, in which case the paper’s “without assuming space or time” still fails and the rewrite has moved the problem rather than removed it. Were (d) a conjecture, its rejection condition would be an argument that asymmetric determination cannot be stated without a before and an after. Nothing below tests it.
What an occurrence is. A vertex in a graph of representational occurrences across executions: one tokening, at one place in one run, of a representation whose content concerns the system itself. The architecture’s feedback graph is a different graph, its vertices modules and its edges wires: one self-model register is one vertex there and has as many occurrences as there are assignments to it over the runs assessed, one per assignment even when the value repeats. That graph is written first.
The order. Occurrence r precedes s when r is among what fixes s. Vertices are counted, edges never.
The collapse rule. That relation need not be acyclic. Each strongly connected component collapses to one vertex; call collapsed vertices components and uncollapsed ones occurrences, so feedback yields a directed acyclic graph of components. Component K precedes L when K is not L and some member of K precedes some member of L, closed transitively; “r before s” below is that relation on components. A single-tick loop whose occurrence graph consists of one strongly connected component contributes one counted component and cannot supply the required pair. A single update containing several components is counted by the same rule as any other graph. Whether admitting such an update violates the published single-tick exclusion is tested under R3. A component inherits a type only when every occurrence in it is of that type; a mixed component inherits none, cannot serve in the count, supplies the count neither retention nor anticipation, and keeps its place in the precedence relation. A component’s content is the set of its members’ contents, and L differs in content from an earlier K when L’s set holds a content identical to no content in K’s set. Count conditions are evaluated on components, lifted by one quantifier fixed here: K is retained in L when K precedes L and some content in L’s set represents a member of K or a content in K’s set; K satisfies the anticipation clause when some member of K does.
Type-sameness. Two occurrences are of one self-representational type when they are tokens of one representational schema of the system: one structure, format or slot its own construction fixes, such as a register. The test is categorical and needs no graded similarity over contents. An item represented but never tokened is of type T when the content representing it specifies T’s schema; where that content specifies no schema, the anticipation clause returns no verdict for that occurrence. Where no schema can be identified the conjecture returns no verdict, and the schema bullet below stands.
The retention requirement. The published wording asks that the system “relate its present operations to its past”, which is more than being among what fixes s. So: r is retained in s when r precedes s and s’s content represents r or r’s content. Determination alone is a variant not under test.
The anticipation clause. An occurrence s satisfies the clause when its content specifies an item t of s’s own representational schema and represents s as one of t’s determinants. If t is an occurrence in the assessed graph, its component must be distinct from s’s component and must not precede s’s component. If t is not an occurrence in that graph, no actual order position is assigned to it; the representational requirements still apply. An item in s’s own component never qualifies. Predictions never realised can therefore satisfy the clause. Where the represented item’s schema is unspecified, the clause returns no verdict for that occurrence. Whether these requirements preserve the published demand for anticipation is tested under R3.
The condition. The fold meets continuity when the collapsed graph holds two vertices of one self-representational type, r before s, differing in content, with r retained in s, and some component satisfies the anticipation clause. The threshold is two, no adjustable parameter. Two records inside one update are two occurrences when they fall in distinct components.
The bearer. Under objection 16’s allowance that “exploratory work may proceed under explicit stipulations”: what persists is the self-representational type and what is ordered are its occurrences. System and attributed perspective stay where objection 16 leaves them.
Its falsifier #
R1, against (a). Exhibit a premise that assigns a magnitude. Commitment (a) falls; (b) and (c) stand.
R2a, against (b), the premise audit. Against the premise list step one produces, exhibit a premise supplying an independent index, a simultaneity relation independent of the order, or a direction the order does not supply. Commitment (b) falls, no derivation required, and (a) and (c) stand.
R2, against (b), the derivation. Show that a requirement used in stating the condition necessarily presupposes an independent index, a simultaneity relation independent of the order, or a direction the order does not supply. The argument must establish that temporal requirement; failure of one proposed account of individuation is insufficient. Content difference selects qualifying vertices from an already specified occurrence graph. It is one possible place to examine, alongside the occurrence, schema, collapse, retention and anticipation definitions. Establishing the required additional temporal structure defeats (b); merely recovering a number or showing that order position is an inadequate individuator does not.
R3, against (c), preservation. Run the published criterion 2 and the rewrite side by side over the exclusion case and specified ordinary systems. Record accept, reject or undefined separately for each. Opposite definite verdicts falsify (c). A definite published verdict paired with an undefined result expressly prescribed by the rewrite also fails (c)’s promise to preserve that verdict. An assessment that is merely unfinished or disputed is inconclusive. These results do not by themselves refute (a) or (b), and they are recorded separately from the conjunction results under R4.
R4, against (c) at M1’s level, the conjunction. M1’s falsifier reads “The four fold-criteria are satisfied by systems no one counts as candidates”. Both conjunctions are recorded for one system, the published four against the four with the rewrite in place of criterion 2, each outcome’s meaning fixed in advance.
- Passes the rewritten conjunction, fails the published one: the rewrite lost an exclusion, (c) fails, and this is the precommitted reason to abandon it.
- Passes both: published M1’s filter was already this permissive, a finding about the paper and a correction only the author can make; (a) and (b) untouched, this conjecture not refuted, named now so it cannot later be reported as a success.
- Fails both: nothing is established.
- Fails the rewritten conjunction, passes the published one: the rewrite is stricter than what it replaces, a second kind of preservation failure.
Non-candidate status, fixed first. A non-candidate is deployed software of an ordinary kind that no publication proposes as a candidate for consciousness and nobody on the floor defends. That judgement, with the names of those asked, is written before step three and never revised after.
The candidate and its boundary. A self-hosting compiler under an incremental build system that rebuilds its own toolchain and plans each build with a dry run over its own dependency graph, under a supervisor restarting a failed build from the last good state. The assessed system is fixed here and not moved later: compiler, build system and supervisor, one process tree, the supervisor inside. Put the supervisor outside and criterion 4, self-maintenance, fails and the candidate is no counterexample. The build graph meets criterion 1; I expect failure on criterion 3, counterfactual sensitivity, since the dry run may model alternative states only of its files.
What does not falsify it. The recovery of a number: chain depth and interval cardinality remain different quantities, as cycle 1 showed.
Its cheapest test or argument route #
Step one, formalise. Write the order, collapse rule, type-sameness, retention, anticipation clause and count as definitions, and list the premises singly so R1 and R2a run against a list. One Opus session.
Step two, preservation, running R3: compare the published criterion and the rewrite over the single-tick loop, a nested pair of records in one update, a thermostat, a spreadsheet holding its own dependency graph, a proof assistant checking its kernel, and the build system. Include controlled cases distinguishing determination from retained history, realised from unrealised predictions, and a represented item inside the forecasting occurrence’s component from one outside it. Record accept, reject or undefined separately for each criterion, following R3’s treatment of those outcomes.
Step three, the conjunction, running R4: the four criteria assessed against the candidate twice, published and rewritten, after the non-candidate judgement and boundary are written, the verdict whichever way it falls.
Step four, the derivation question, running R2: examine whether any requirement used in stating the condition necessarily imports temporal structure beyond the dependency order. An estimated hour of review by somebody working on grounding or the individuation of events is one possible route. The author invites, no agent contacts anyone, and this review is no precondition for adoption.
On cost, the historical expenditure and route estimate are proposer-reported and unverified. Cycle 1 is reported to have run under a USD 40 cap and come in at USD 15.70. The supplied floor seed records an author-set monthly cap of USD 150 in Claude tokens and a per-test cap of USD 20 for provider calls. Steps one to three use paid model reasoning, estimated here at roughly USD 5 to 10 in Claude tokens at list rates, with no additional experimental provider calls. That reasoning counts against the monthly cap; the packet does not establish the remaining balance. Amendment A1 also requires the test to be separately capped and preregistered before execution. No separate cap specifically covering this model-reasoning test is documented here.
What it adds #
What the paper could not say before: the continuity criterion asks for a recurrence, one self-representational schema tokened twice in a dependency order, the later retaining the earlier, some token reaching forward, and nothing about how long it took. It bears on M1, whose filter the four criteria are, and narrows the charge’s first horn at site/lab/case-against.html, “Either the explanation assumes duration, or the criteria and P1 use the same name for two different objects”, to commitment (d): the criterion asks for no duration, and whether its order is already temporal stays open.
Two things stay open and are not conjectured here: whether counting recurrences reconstructs the duration P1 concerns, and whether the criterion’s persisting type and P1’s duration are one object, so the charge’s second horn stands.
The paper’s speculative claims read: “The appearance of locality and duration is the way relational structure holds together for the fold, not a static identity it simply has.” The claims graph’s short statement for P1 says the same: “Locality and duration as how relational structure coheres for the fold”. A criterion helping itself to duration would leave P1 nothing to account for; P1 stays speculative at 0.3 and nothing here tests it.
On the reply that continuity is a condition only on biological implementations: Part VII asks of AI systems “what the topology of its self-reference is”, a question biology-only criteria leave unanswerable.
No transfer rule reaches M1, so by the lab’s default this attaches to nothing; the rule it would need says when a rewritten criterion replaces a published one, the author’s correction to make.
The conjecture as filed is kept at downloads/lab/floor/cycle-3/01_persistence_rev3_opus.md, and its critique at 01_persistence_rev3_critique_astra.md; the cycle note is 00_CYCLE_NOTE.md.
Review #
Read from outside its lineage three times by GPT-6 Astra, on amendment A1’s four questions only. Cycle 1, HOLD: a count, an interval, a metric and duration run together; a movable threshold; one criterion’s satisfaction treated as four. Cycle 2, HOLD: the target of the metric claim unstable among three commitments; occurrence counting after collapsing cycles unspecified. Cycle 3, 19 September 2026, ADOPT AMENDED: all four questions answered yes; cycle-2 findings 1, 4, 5 and 6 answered, 2, 3 and 7 partly; five exact replacements (the single-tick sentence made explicit about the occurrence-graph shape it assumes; the anticipation clause restated on components; step two given controlled cases and a three-way record; the derivation route reworded so that one route is not an exclusive gate; a rewrite-prescribed undefined verdict counted as a preservation failure; the cost paragraph corrected to the monthly cap), applied the same day, after which the critic’s stated verdict is ADOPT. Before the third read, two in-lineage refuters and an auditor on Claude Opus 5 returned fifteen findings, all applied; they are summarised in the cycle note. Verdicts in full: cycle 1, cycle 2, cycle 3.
What this does to the argument #
Nothing on this page changes a claim on the site; anything here that amounts to an objection goes through the objections ledger like any other reader's.
What would count against this #
- The candidate, or another non-candidate, passes the rewritten conjunction and fails the published one. Commitment (c) fails and the formulation goes.
- Any system the published criterion rejects and the rewrite accepts. Commitment (c) fails, whatever the conjunction does.
- Any system the published criterion accepts and the rewrite rejects. Commitment (c) fails on the strict side.
- A system with no identifiable schema, or whose schemas drift, where the verdict turns on how similar two contents are. Type-sameness would then need a graded similarity measure, and the rewrite would have moved a metric rather than removed one.
- A premise in step one’s list supplying an independent index, a simultaneity relation independent of the order, or a direction the order does not supply. Commitment (b) fails with no derivation.
- A derivation with both parts of R2. Commitment (b) fails and the condition carries temporal structure in its premises.
- An argument for (d), that asymmetric determination cannot be stated without a before and an after. (a) to (c) would survive and the rewrite would still assume time.
- A magnitude found among the premises. Commitment (a) fails.
- The retention requirement collapses into criterion 1, which asks that the self-representation affect “ongoing processing”. If “ongoing” already requires tokens of one type with the later citing the earlier, criterion 2 adds nothing; the anticipation clause is similarly exposed to criterion 3.
- The published “across time” turns out to exclude a chain of records inside one update, which the rewrite admits. Commitment (c) fails where the rewrite is most permissive.
- The Positions page’s Dōgen pressure lands: “the single self-referential structure that persists and gets seen from two sides may itself be a reification of what is momentary”. If the schema needs a persisting thing behind its tokens, the rewrite traded a metric for a substance.
- The stipulated bearer is rejected. If the type is what persists, a fold interrupted and resumed is one fold, and a reader taking the system or the perspective as bearer reads the criterion differently. The stipulation is undefended.
Proposed by a Claude Opus 5 agent and read by GPT-6 Astra at the author’s request; edited by Claude Fable 5.1.