DISTRIBUTED TEMPORAL SELF-MODELS: FINAL THREE-ITERATION REPORT
Preserved consolidated research report from 4 November 2026. This is broader supporting material for the selected public addendum, not another final manuscript or new external review. Original internal-review and result claims remain dated research records. Local path references are omitted in this public derivative; no original is edited. Original SHA-256: da071c115c80eb412a72c5892f51ec913d44719f3563c20f8ec5076b8a14f026.
DISTRIBUTED TEMPORAL SELF-MODELS: FINAL THREE-ITERATION REPORT
Iteration 3 of 3 — task-relevant communication after temporal ambiguity
4 November 2026, Asia/Bangkok. Finite-model research; AI review, not peer review.
RESULT
For a declared eight-history system, a fixed historical audit task can be
identified adaptively with exactly 14 additional wire bits and 16 rounds.
Every fixed identifying selection requires at least 21 bits and 24 rounds.
Both optima include requests, response headers, serialization and leaf work.
Central and distributed implementations have the same optimum when given the
same ports, input placement and local encoding rights. Exact clock knowledge
alone cannot identify the task after the relevant body information has been
erased; stored historical diagnostics are essential in this instance.
The selection criterion is deterministic equivalence-class determination, with
established active-learning antecedents. The contribution here is a fully
specified temporal example and communication contract, not a claim to have
invented test selection, decision trees or distributed inference.
1. WHAT THE THREE ITERATIONS ESTABLISHED
Iteration 1: with acquisition times and a declared transition relation, feasible
state propagation decides whether local reports fit any trajectory. For known
binary increments, a labeled spacetime graph is satisfiable exactly when every
cycle has even XOR. This is classical signed-graph balance. Reports AB_0=0,
BC_0=0, CA_1=1 fit 000 -> 100 if A flips, although erasing timestamps gives
static odd parity. Moving CA to time 0 gives actual same-time incompatibility;
requiring persistence instead gives temporal incompatibility without an empty
individual slice. Raw gathering versus aggregation on a path saved bit-hops,
but not latency; central control allowing aggregation matched it.
Iteration 2: one constant offset d_j is shared across every module-j record,
with tau=h-d_j. Horizon and receipt constraints intersect into a module-wide
integer domain K_j. With at most one body-bit flip per tick, one module's full-
state reports are feasible, for free initial state, iff successive Hamming
distances fit successive local time gaps and K_j is nonempty. For many modules,
process each report at its latest candidate acquisition time. With n body bits,
ell=max(1,2*kappa), and at most a clocks alive from first through last processed
record, bag tables have at most 2^[n(ell+1)](2*kappa+1)^a entries and separators
at most 2^[n ell](2*kappa+1)^a. Join/projection gives exact witnesses or complete
negative table traces. This is a concrete application of classical elimination.
That iteration distinguished Omega (timing-admissible offset vectors), Sigma
(offsets admitting individually consistent slices), and Theta (offsets admitting
a whole valid trajectory). Theta subset Sigma subset Omega. Sigma empty means
every offset choice conflicts at some slice, not necessarily one fixed instant.
Independent retiming can be unsound: 00 and 11 one local tick apart have no
shared-offset trajectory at rate one, but separate per-record offsets admit 18.
Two separately clocked reports 00 and 11 at local stamp 2 admit 36 joint
offset/history explanations, despite rejection under both offsets zero.
The preserved iteration reports contain full proofs and broader limitations.
This report consolidates their conclusions and adds an identification problem.
2. THREE DIFFERENT QUANTIFIERS
Retrospective consistency: does there exist a pair (offset vector, trajectory)
satisfying the already received records and model? Iterations 1-2 answer this.
Identification: after diagnostic transcript z, is the fixed target f constant
over EVERY history still compatible with z? Finding one favorable history is
insufficient. This iteration asks for a policy that identifies f for every
initially feasible history, within a worst-case bit budget and deadline.
Future dynamic controllability: does an observation-dependent execution strategy
succeed for every allowed future external realization? This is different again.
Our nondestructive queries concern a fixed past and stable registers. We neither
choose future body actions nor prove an STNU controllability theorem [S5].
3. FIX THE TASK BEFORE THE DIAGNOSTIC PORTS
Task: report the pre-reset flag associated with the earlier calibration context.
If context P occurred before Q, return historical flag U_0; otherwise return
historical flag V_0. This externally specified audit rule is fixed before
selecting any diagnostic. It does not require recovering every historical bit.
For an exactly solvable bounded instance, let W contain eight histories indexed
by (a,u,v) in {0,1}^3. The target is f(a,u,v)=u if a=0, otherwise v. The bit a
specifies calibration order. This is a NEW declared known-template restriction,
not a conclusion that all 36 histories of iteration 2 reduce to these eight.
The body has four bits: the two calibration bits plus U and V. For t=0,...,4,
the calibration pair follows one of these known templates:
a=0: 00,00,01,11,11; offsets (d_P,d_Q)=(+1,-1).
a=1: 11,11,01,00,00; offsets (d_P,d_Q)=(-1,+1).
The calibration records are P: pair 00, h=2,r=3, and Q: pair 11,h=2,r=3,
with kappa=1 and D=2. Thus a=0 puts P at true time 1 and Q at 3; a=1 reverses
their order. U=u and V=v remain constant through t=4. At t=5 reset U to zero;
at t=6 reset V; at t=7 reset one calibration bit; at t=8 reset the other.
Every transition changes at most one bit. The full body is 0000 at t=8 and
remains so during the query session. The exact integer paths are in checks.json.
Two independent passive archives already hold the t=0 values U_0 and V_0.
Acquisition/storage of these records occurred before this query and is an
identical sunk input cost for every comparator. This is an explicit extra
information channel; it is not retrospective access inferred from current state.
Without those archives the result changes to the impossibility below.
4. PORTS, OWNERSHIP AND NETWORK CONTRACT
The root R must produce one correct target bit locally. Four distinct leaves
own four distinct ports. A leaf has access only to its own port, not the other
private values. The known model, target, port types, encoding, schedule and
archive provenance are public. There are no noisy or failing ports.
A — trusted clock anchor: returns a, encoded from the sign of d_P. A trusted
reference clock is already available at this leaf; the constant offset
persists during querying. Establishing that reference is a declared
precondition, not free synchronization inferred from the messages.
B — archive read: returns the stored historical U_0=u, nondestructively.
C — archive read: returns the stored historical V_0=v, nondestructively.
D — fresh body reading summary: returns whether the current body is nonzero.
It always returns zero after t=8. No historical value is returned by D.
The physical channel is a shared half-duplex bus joining the root and leaves.
There is one outstanding request/response transaction at a time. This serialized
RPC rule is part of the available interface; arbitrary batching, overlapping
requests, unsolicited pushes, interleaf transfers and hidden timing codes are
not available to either architecture. Each leaf may encode any function of its
own one-bit answer. It cannot compute the joint target without obtaining other
leaves' data through charged transactions.
Each request: two-bit fixed query/version header plus two-bit port address =4
wire bits. Each response: two-bit header plus one answer bit =3 wire bits.
The outstanding request identifies the responder, so a second response address
is unnecessary. Capacity is one bit per round, no additional propagation delay.
After the request, the leaf uses one reserved compute/read round; then replies.
One RPC therefore costs exactly 7 transmitted bits and 8 rounds. No acknowledgments
or retransmissions are needed in this explicitly reliable protocol. These would
be additional costs in a different network. Time starts after query closure at
t=8; rounds are relative to that start. Root target selection is a zero-additional-
round combinational operation included equally in every comparator.
The two archive slots are stored at separate leaves. A request to a fictional
joint 'give me f' port would change the input access contract. Central orchestration
may choose the next leaf based on answers; distributed orchestration is given
exactly the same permissions and must deliver its decision to the same root.
5. PROPOSITION V1: SHARP FIXED-SELECTION CERTIFICATE
Let W be any finite nonempty set of feasible histories, f its fixed target, and
q:W->Y_q the deterministic answers of each permitted nonmutating port. Let E
be the pairs {w,w'} with f(w)!=f(w'). A fixed set P identifies f exactly iff
for every {w,w'} in E, some q in P has q(w)!=q(w'). (1)
Proof. If such a pair has identical selected answers, the root receives one
transcript compatible with opposite targets and cannot identify f. Conversely,
if (1) holds, two histories in one selected-answer cell cannot have different
targets. That cell therefore has a unique target, which can be reported. QED.
Each port covers the target-disagreeing pairs it separates. With additive costs
and the serialized contract, finding the minimum fixed set is precisely the
finite weighted pair-cover problem, also subject to the sum of transaction
durations fitting the deadline. Enumerating all subsets gives an exact bounded
algorithm. A positive certificate supplies a separating port for every pair;
a failed selection has an explicit uncovered pair. For an optimum lower bound,
exclude all cheaper selections or give a necessity argument as below.
This is deterministic equivalence-class determination [S6]; general overlapping
decision regions [S7] are broader than this fixed-label task. No new general
selection paradigm is claimed. Their Bayesian expected-cost approximation
guarantees are not substituted for our exact worst-case budget criterion.
6. PROPOSITION V2: EXACT ADAPTIVE BUDGET/DEADLINE CRITERION
For current version space S, unused ports P, remaining bits b and rounds d,
let G(S,P,b,d) mean a guaranteed identifying policy exists. If f is constant on
S, G=true. Otherwise
G(S,P,b,d) = OR over feasible q in P of
AND over nonempty answer cells S_(q,y) of
G(S_(q,y),P\{q},b-c_q,d-l_q). (2)
Feasible requires c_q<=b and l_q<=d. Here c_q is total charged request/response
bits, and l_q includes read/compute and serialized delivery. Empty disjunction
is false. For this instance (c_q,l_q)=(7,8) for every q.
Proof by induction on remaining ports. A nonterminal deterministic policy must
choose an affordable first port. Every possible answer must admit a continuation
within the remaining resources, proving necessity. Conversely choose a port
satisfying the conjunction and attach its guaranteed continuations, proving
sufficiency. A repeated deterministic nondestructive query provides no new
information, so allowing each port once loses no optimal policy. QED.
This finite recurrence is exact, not a polynomial-time claim: a crude memoization
bound is 2^|W| * 2^|P| * (b+1)(d+1) resource states, times port/partition work.
Positive certificates are decision trees with monochromatic leaves and bounded
root-to-leaf cost. Negative certificates can give all failed first-port/resource
subproblems, or a direct indistinguishability lower bound. General outcome-dependent
duration, destructive sensing, concurrency or unknown future readings would need
an augmented state and protocol, not this unchanged recurrence.
7. SOLVED INSTANCE: EXACT OPTIMA AND ADVANTAGE OF ADAPTATION
Adaptive policy: ask A. If its answer is zero, ask B and return u. If it is one,
ask C and return v. Every history is correctly classified after two RPCs:
rounds 0..4 request A; 4..5 read; 5..8 response;
rounds 8..12 request selected archive; 12..13 read; 13..16 response.
Wire cost is 2*(4+3)=14 bits and duration is 16 rounds. The root already holds
the desired answer, so no output packet is hidden. With budget 14 and deadline
16, this policy is admissible.
Adaptive lower bound. No single port identifies f. A leaves both u and v unknown;
B leaves the a=1,v branch unresolved; C leaves the a=0,u branch unresolved;
D is constant. Any guaranteed policy therefore has a branch requiring at least
two RPCs. The fixed transaction format gives worst-case bounds of at least
14 bits and 16 rounds. The displayed policy attains both simultaneously.
Fixed selection: A,B,C together identify f, at 21 bits and 24 rounds. All three
are necessary. If B is missing, histories (0,0,0) and (0,1,0) agree on A,C,D
but have different targets. If C is missing, (1,0,0) and (1,0,1) agree on A,B,D.
If A is missing, (0,0,1) and (1,0,1) agree on B,C,D but targets differ. Hence
every fixed identifying set includes A,B,C. D cannot replace any of them.
Here 'fixed' means the selected port set does not depend on answers; all selected
transactions are executed. Even a fixed order with answer-dependent early
stopping has worst-case depth three for this instance: any first two distinct
informative ports leave an ambiguous answer cell. Its first two queries are
fixed, so some third query is needed. Repeating ports or using D cannot improve
that bound. The useful adaptation is choosing WHICH archive to query after A.
The policy need not identify the whole history. After A=0 and B=u, the unknown
v remains free, but the audit target is fixed. This is task-relevant identification.
8. IMPOSSIBILITY: OFFSET KNOWLEDGE DOES NOT RECOVER ERASED HISTORY
Suppose only A and fresh current-state ports are available. Histories (0,0,0)
and (0,1,0) have identical offsets and identical current body 0000, but opposite
targets. All permitted port answers coincide. Inductively any adaptive policy
selects the same next request after identical transcripts and receives the same
answer. It cannot distinguish the target on this pair, at any bit budget or
deadline. Randomization with zero-error guarantees cannot repair identical
transcripts; condition on the same random choices.
Thus perfect knowledge of both offsets is insufficient. Repeated fresh readings
are also insufficient because the body remains reset. An archival port works
because an actual historical record survived elsewhere. Giving the root that
past reading without the archive would change the information assumptions.
9. CENTRALIZED VERSUS DISTRIBUTED COMPUTATION
Both architectures receive the same port menu, original leaf ownership, local
encoding rights, fixed RPC framing, serialized bus, compute slot and output
root. Each can implement the 14-bit adaptive policy. A central controller is
allowed to request A and conditionally choose B/C. A distributed controller
cannot receive another leaf's data without charged communication. No architecture
beats the indistinguishability lower bound under these permissions.
The 21-versus-14 comparison is fixed versus adaptive SELECTION, not centralized
versus distributed architecture. Moving both archives and clock state into one
leaf that can directly return f would permit a cheaper query but would change
ownership/ports. Adding a broadcast batch primitive or overlapping transactions
would change latency. Those are legitimate alternative designs, not conclusions
about this contract. Header bits and local compute are counted for all contenders.
10. EXACT VERIFICATION AND FRESH REVIEW
checks.py uses Python's standard library and exact finite computations:
- all eight histories checked for one-flip transitions, correct shared offsets,
valid calibration receipts, correct t=0 archives, and reset current body;
- all 16 port subsets checked; only ABC and ABCD identify the full target;
- 4,096 version-space/port-set combinations checked for equivalence of pair
coverage and direct answer-cell monochromaticity;
- 1,020 nonempty-version-space/depth combinations compared the resource
recurrence with a separate explicit decision-tree enumeration;
- eight complete adaptive schedules checked against true target values;
- explicit opposite-target pairs for every selection of at most two ports,
and the clock-only impossibility pair, saved in checks.json.
The depth oracle covers depths 0..3. Exact budget checks also reject 13 bits at
16 rounds and 14 bits at 15 rounds. Earlier verification remains preserved:
iteration 1 had 40,671 main finite cases and 1,108 independent audit cases;
iteration 2 had 4,297 comparison cases, four scaling constructions and 400
additional seeded small exact-oracle cases (the latter were not exhaustive).
A fresh AI reviewer independently audited the new example and theory. Its final
results are in ai_review.txt. AI review is not external peer review. Finite checks
support implementation and examples; mathematical proofs carry the stated
general finite-model criteria. No historical priority certification is implied.
11. CLAIM LEDGER ACROSS ALL THREE TURNS
Claim Status and boundary
MIND static parity and nerve caution Inherited [S1]; not a new result.
Trajectory feasible-state propagation Proved application, iteration 1;
standard finite-state reasoning.
Spacetime even-cycle certificate Proved specialization of signed-
graph balance; known increments only.
Shared-offset distance criterion Proved elementary restriction,
iteration 2; complete-state records,
one clock, free initial state.
Offset-lifetime/window table bound Proved concrete decomposition,
iteration 2; classical elimination
antecedents [S3,S4]; novelty open.
Fixed pair-cover identification Established ECD principle [S6];
proof and exact checks supplied here.
Adaptive resource recurrence Standard finite decision-tree DP;
exact under declared RPC costs.
14/16 versus 21/24 optimum New worked bounded construction in
this study, proved and enumerated;
not a priority claim for multiplexing.
Clock-only/current-body impossibility Proved indistinguishability instance;
depends on real archival access.
Architectural superiority Not established; matched permission
sets attain the same optimum here.
Physical/cognitive interpretation Specified toy body/diagnostic model;
no biological validation, awareness
test or bare-nerve contextuality claim.
Future robust controllability Not addressed; different quantifiers.
12. VIDEO EXPLANATION AND ITS LIMITS
The accompanying explainer.html is a self-contained animated captioned explainer
with playback controls. It is a browser animation, not a rendered or published
video file. Suggested two-minute narration/storyboard:
0:00-0:25 — Three cards arrive together: AB=0, BC=0, CA=1. Freeze them over a
single body diagram. Narration: 'If these all describe one instant, no assignment
works. But arrival time does not tell us when a measurement was acquired.'
0:25-0:50 — Reveal acquisition labels 0,0,1 and animate 000 -> 100. Narration:
'The first two reports fit time zero. After A changes, the third fits time one.
Nothing was voted away; the trajectory makes all three reports compatible.'
0:50-1:10 — Change the last label back to zero. Show the XOR cancellation 0=1.
Narration: 'At the same time the contradiction remains. A later change cannot
repair a statement about this past instant.'
1:10-1:30 — Show 00 and 11 one local clock tick apart. Slide both timestamps
together with one clock-offset slider. Narration: 'A constant offset moves both
times equally. It cannot stretch a one-tick gap enough for two changes when
only one bit may change per tick. Moving each report independently is a different,
less restrictive model.'
1:30-2:00 — Reset the body to zero and reveal the two historical archives.
Ask the clock anchor, then illuminate only the relevant archive. Narration:
'The current reading cannot reconstruct erased flags. With actual stored records,
the clock answer tells us which archive matters: two queries, fourteen wire bits
under this protocol. Central and distributed controllers can both do it.'
Closing caption: 'Consistency is model-relative. Identification needs all surviving
histories to agree on the chosen task. Neither result establishes awareness.'
13. LIMITATIONS AND STOPPING POINT
The new optimum is exact for eight known-template histories, deterministic stable
ports, preexisting archives/reference, and serialized fixed-frame RPCs. It is not
a universal lower bound for arbitrary networks, raw synchronization, compression,
overlapping queries, outcome-dependent durations, or unknown post-query changes.
Changing the task changes which pairs must be separated. Constructing W can itself
be expensive; iteration 2 supplies a feasibility representation, not an automatic
efficient optimizer over every decision tree. Noise and incomplete archival
provenance require a different identification/error contract.
The three requested research turns are complete. No fourth iteration has begun.
No publication or Git commands occurred in this turn. Prior iteration files are
preserved byte-for-byte, checked against the saved 46-file baseline. Earlier
inherited inputs remain protected by their saved hashes.
No further research run is required for this request. Defer empirical/network
validation until a real device, task and recording interface are specified;
those facts determine whether these assumptions are useful. Such an independent
implementation study would warrant a new chat with this report and checker as
its handoff, but no new chat or model run is needed now.
PRIMARY SOURCES (URLs accessed or reused as specified; date 2026-11-04)
[S1] The MIND: Five Finite Constructions, 5 October 2026, M4 pp.6-7; original
manuscript read in iteration 1, preserved and reused.
https://www.b-theory.com/manuscripts/mind-five-finite-constructions-r1-2026-10-05.pdf
[S2] Harary (1958), On the number of bi-colored graphs, section 6 p.754,
primary author restatement of signed-graph balance. Read in iteration 1.
https://msp.org/pjm/1958/8-4/pjm-v8-n4-p11-p.pdf
[S3] Dechter, Meiri, Pearl (1991), Temporal constraint networks, sections 2-3.
Primary full reprint read in iteration 2; simple-temporal guarantees not imported
to arbitrary body/clock factors.
https://ftp.cs.ucla.edu/pub/stat_ser/r113-L-reprint.pdf
[S4] Petcu, Faltings (2005), A Scalable Method for Multiagent Constraint
Optimization (DPOP), sections 3-5 and message-size analysis. Read in iteration 2.
https://www.eecis.udel.edu/~decker/courses/886f06/pubs/DPOP.pdf
[S5] Cairo, Rizzi (2017), Dynamic Controllability Made Simple, section 2.
Read in iteration 2 for quantifier distinctions.
https://drops.dagstuhl.de/entities/document/10.4230/LIPIcs.TIME.2017.8
[S6] Golovin, Krause, Ray (2010), Near-Optimal Bayesian Active Learning with
Noisy Observations. Primary full text, sections 2-3; ECD/conditional plans and
version spaces inspected in iteration 3. Their expected-cost guarantees differ.
https://www.cs.cmu.edu/~dgolovin/papers/nips10.pdf
https://arxiv.org/abs/1010.3091
[S7] Javdani, Chen, Karbasi, Krause, Bagnell, Srinivasa (2014), Near Optimal
Bayesian Active Learning for Decision Making, AISTATS/PMLR 33:430-438. Publisher
record and primary-paper formulation inspected for decision-region antecedents.
https://proceedings.mlr.press/v33/javdani14.html
https://proceedings.mlr.press/v33/javdani14.pdf
[S8] Robinson (2016), Sheaves are the canonical data structure for sensor
integration, section 5.2 before Example 30. Read in iteration 1: omission of
acquisition time as an inconsistency source is already discussed there.
https://arxiv.org/html/1603.01446
DELIVERABLES: report.txt, checks.py, checks.json, explainer.html, ai_review.txt,
independent_checks.py, independent_results.json, and sources.json in
outputs/iteration_03. The earlier complete reports remain
at outputs/iteration_01/report.txt and outputs/iteration_02/report.txt.