Loss is declared.
The experiment names its transformation relation, corpus envelope, bit budget, and unsupported cases.
Independent experiment · perceptual similarity
The
Kolmogrov
question
Can a bounded object preserve the neighborhoods we care about without pretending it contains the unbounded truth it points toward?
Current question
Can a bounded family of separately addressable measurements preserve declared deformation neighborhoods better than simpler equal-bit sketches—while exact records remain the only authority on what a file is?
01
The apparent paradox
The theorem is not the enemy. A mistaken job description is.
The original Zeta note began from a genuine tension: an arbitrary unique string cannot be packed, without loss, into one permanently fixed address; nevertheless, minds and models work through bounded internal states. What sort of compression are they performing?
Kolmogorov’s incompressibility argument gives the hard boundary. Fix a description language. Among all strings of length n, most cannot be described in materially fewer than n bits. More simply, no function from arbitrarily long strings into B bits can be both injective and reversible: there are only 2B destinations. Once the source set grows beyond that number, at least two sources share a destination.
A transformer does not refute this. A bounded model can respond usefully to unboundedly many inputs because it does not losslessly contain each input. It discards distinctions, exploits regularities in a distribution, and receives the present input again at evaluation. Its state is an instrument, not a universal archive.
That correction changes the experiment. Kolmogrov does not seek a “hash of truth.” It seeks a bounded measurement that is useful for proposing neighbors while the original object remains available for verification. The theorem survives untouched. The system gives up the premise that the fixed-width object must do everything.
02
The compromise
Move exact identity outside the sketch.
Every file keeps an exact external anchor: its bytes, stable record identity, and authoritative metadata. Kolmogrov may place two files near one another, fail to place them near one another, or collide them in a channel. None of those events makes the files equal. The sketch proposes; the record decides.
1Measure querybounded evidence
→
2Visit neighborhoodscandidate postings
→
3Fuse lateretain disagreements
→
4Verify exact recordauthority remains outside
The change of contractcandidate retrieval
addresses = measure(query) # what / where / when / relation
candidates = {}
for channel, keys in addresses:
candidates[channel] = postings.near(keys)
ranked = late_fuse(candidates) # disagreement is evidence
return exact_verify(ranked) # the sketch never becomes the recordChannels remain visible.
A content collision can be rescued by position, history, containment, or an exact comparison.
Overflow is an answer.
When length, candidate-pool size, or noise exceeds the measured envelope, the result must say unsupported—not silently return certainty.
03
The breakdown theory
What. Where. When. Then: with what.
The originating note compared the design to a sequence of spectral measurements: one measurement says how much, another says when, another says at which frequency. For strings and structured records, the proposed analogue was content, position, and combination.
The later research makes the idea more precise. The primitive object is not one vector but an exhaustive occurrence measure. It records a piece of evidence together with its location, scale, ordering, containment, and source. Edits act on that measure as transport plus destruction: some occurrences survive, some move, some are created, and some vanish.
WHAT
Evidence
- symbols and fragments
- counts and shapes
- literal and structural atoms
- what survives a crop
WHERE
Placement
- absolute and relative position
- object and neighborhood scale
- hierarchy and containment
- which context owns the evidence
WHEN
History
- order of occurrence
- which edit path preserves it
- where a mutation moves it
- when evidence is destroyed
WITH WHAT
Combination
- co-occurrence and adjacency
- ordered relations
- cross-axis interactions
- shared history coherence
“When” is not merely clock time. It is the history coordinate: under which operation, at which stage, and along which ordered path does an observation remain recognizable?
Combination is not a fourth bag of features. It is the coupling that prevents a content match from one history and a position match from another from being mistaken for one coherent witness.
Neighborhood identity
A neighbor is not a duplicate. Sometimes it is not even one object.
Exact identity is singular: this record, these bytes, this anchor. Perceptual identity is relational. A file belongs to several neighborhoods at once—literal content, ordered fragments, containment, scale, and deformation history. Another file can share one neighborhood and not the others.
Deletion exposes the point. If several longer strings yield the same shorter string after one deletion, the evidence no longer identifies a unique parent. Forcing one answer creates fictional precision. The right result is an ambiguity class, to be narrowed by independent evidence or returned honestly as a set.
This is the project’s deeper definition of neighborhood identity: not a hidden claim that two objects are the same, but an auditable account of the relations under which they remain candidates for one another.
WHAT
WHERE
WHEN
WITH WHAT
EXACTrecord anchor
Record identity
Exact bytes and immutable IDs. Never inferred from proximity.
Neighborhood identity
Task-relative memberships induced by declared transformations and measurements.
Interpretation identity
Candidate senses, roles, context, provenance, and uncertainty. A separate semantic system—not a promise made by this hash.
04
Perceptual tradeoff
Smooth enough to recognize; sharp enough to matter.
A useful perceptual map wants nearby edits to produce nearby evidence: one changed letter, one insertion, a moved fragment, a longer prefix. But uniform smoothness is not the objective. A comma can reverse who is being eaten. A short query contained inside a long record should be close in one direction without forcing the reverse relation. Common fragments should count differently from rare, diagnostic ones.
The compromise is therefore closer to rate–distortion than universal metric preservation. Spend a fixed support budget on the distortions the task can tolerate, and protect the distinctions it cannot. The allocation must be measured against a candidate population, not praised in the abstract.
| We want | Without paying | Declared test |
|---|---|---|
| Stability under ordinary edits | Blindness to protected punctuation or order | Mutation atlas and hard negatives |
| High recall | Candidate lists so broad they become an exact scan | Tail candidate load at equal bits |
| Multi-scale access | Duplicating every scale in storage | Progressive support and index bytes |
| Containment | False symmetry between part and whole | Directional query benchmarks |
| Collision rescue | Independent channels recombined into one opaque collision | Ablation and late-fusion audit |
| Fixed width | A lie about arbitrary-length capacity | Explicit overflow policy |
There is no single globally correct distance. There is a declared family of transformations, a protected set of distinctions, a population, a budget, and a cost for being wrong.
05
From Zeta to the current object
The experiments kept the question and replaced the answers.
The first substantial candidate was ConeDAG: content, soft position, and path features arranged so small edits usually deform the sketch rather than randomize it. On a 1,080-query adversarial benchmark, it reached .955 mean reciprocal rank and .919 Recall@1, ahead of equal-width word bags, character grams, and the earlier sequence sketch. It also lost to Zeta’s lexical/graph search on a tiny semantic set. That result set the enduring boundary: good deformation geometry is not semantic understanding.
File Manager research then replaced flattering vector geometry with a harsher oracle. First enumerate occurrence evidence exactly; prove how deletion, substitution, and transposition act on it; only then ask which projections preserve the required behavior. This exposed several failures: exact position-resolved edit closure reconstructs literal identity; one scalar refinement depth incorrectly couples unrelated axes; narrow independent views can lose witness provenance; and hash width alone says nothing about perceptual capacity unless length, population, decoder, noise, and tolerated excess are named.
- ZETA
ConeDAG
A smooth sequence sketch established that content, position, and path can outperform simpler equal-width controls on generated deformation retrieval.
- ORACLE
Occurrence geometry
An exhaustive, inspectable state records content, gap position, ordered combination, scale, direction, interaction family, and source anchor.
- PROOF
Action closure
Edits become exact transport and destruction laws. Compression is judged by whether those actions still commute approximately with the reduced state.
- CURRENT
Coupled history tuples
Each deletion history produces one coupled tuple. Coordinates may refine independently, but the coordinates of different histories may not borrow one another’s evidence.
Current candidate, not production contract
The whole history gets one coupled key.
For a filename atom stream of length n, the current radius-one candidate computes the complete descendant produced by deleting each possible position. Each descendant is projected into several nested coordinates, retained as one tuple. A rolling recurrence emits all n keys in linear work per coordinate.
Coupling matters. If coordinate one is allowed to match deletion history A while coordinate two matches history B, the system manufactures a witness that never existed. A coupled tuple requires all coordinates to refer to the same descendant history.
Conditional support law
w ≥ log2(M · nx · ny / τ)
- w
- tuple bits
- M
- records in the routed partition
- nx, ny
- history counts
- τ
- tolerated expected excess
This is a working bound under uniform independent coordinates—not a universal capacity theorem.
Established inside the declared model
- exact recall for the radius-one common-deletion relation;
- linear rolling construction;
- sparse postings rather than a materialized address universe;
- exact contraction from finer to coarser tuple cells.
Still unproved or unmeasured
- million-record and adversarial capacity;
- the final hash family and width;
- native key, dictionary, update, and cold-cache cost;
- whether optional normalized views preserve the declared relation.
Retained counterexample
The current narrow coordinate family makes ldfioia and kbmedfa a false full-profile candidate despite no common exact deletion descendant. Exact authority and explicit capacity limits are not ceremonial safeguards; this collision is why they exist.
06
What “compromise with complexity” means
Not a loophole. A smaller promise with a measurable bill.
The project hopes to live productively beside the incompressibility boundary by changing five things at once:
- Reconstruction becomes retrieval. The bounded state need only place useful candidates in reach.
- Universal strings become a declared domain. Filename profile, edit radius, length, partition size, and overflow behavior are versioned.
- One address becomes complementary evidence. What, where, when, containment, and exact identity can disagree without being erased by premature fusion.
- Unique answers become ambiguity classes. When an edit destroys the evidence that distinguished several parents, the system retains the set.
- Compression quality becomes task loss. The useful question is not “Did we preserve every bit?” but “At the same serialized support, what recall, false-candidate tail, influence, latency, and verification cost did we buy?”
This is the honest bargain. Kolmogorov complexity says the territory cannot fit losslessly inside the fixed map. Kolmogrov asks whether the map can still preserve the roads a particular traveler actually needs—and whether it can admit, precisely, where those roads disappear.
Promotion gate
It must beat the boring controls.
Beauty is not a gate. The candidate must outperform balanced random projections, lexical and n-gram controls, exact scan, and simpler certificate paths at equal serialized support. It must report recall, candidate tails, build and update bytes, cold and warm latency, probe amplification, invisible mutations, and every overflow.
Freeze the task
Observation profile, transformation relation, length and population envelope.
Protect recall
Prove or exhaustively check the declared relation before optimizing selectivity.
Attack the map
Adversarial collisions, omitted directions, Unicode seams, long inputs, empty inputs.
Count the whole bill
Keys, dictionaries, postings, migration, compaction, cache misses, and exact verification.
Source trail
The argument is recoverable from the work.
This account draws on the original Zeta design statement, the ConeDAG retrieval and ambiguity audits, and the later independent File Manager research program. The current candidate remains gated.