Operational Project Alignment and Horizonte
A Conceptual Hypothesis for Long-Horizon Human–AI Collaboration
TYPE: Conceptual Hypothesis
STATUS: Open
BASED ON: Zipvilization development experience, 2023–2026
EXTERNAL RESEARCH: Yes
EMPIRICALLY VALIDATED: No
GENERALIZABILITY: Unresolved
LAST REVIEWED: 2026-09-29
Abstract
What happens when increasingly capable Artificial Intelligence works on the same evolving project for years?
Our experience building Zipvilization exposed a failure mode that initially appeared paradoxical.
AI-generated work became better locally while the project sometimes became worse globally.
Individual documents improved.
Explanations became clearer.
Technical work accelerated.
Yet valid information disappeared, relationships were weakened, historical distinctions were flattened, and locally reasonable transformations accumulated into global incoherence.
We describe this observed failure pattern as:
GOOD LOCAL GENERATION
+
INSUFFICIENT GLOBAL CONSERVATION
or:
LOCALLY BETTER
GLOBALLY WORSE
Increasing context, memory, instructions and prohibitions helped, but did not fully solve the problem.
Over several years, our working method evolved toward an architecture composed of five distinct elements:
Canon
Relationships
History
Epistemic Status
Horizonte
From this experience we propose a narrower concept that we call Operational Project Alignment:
Operational Project Alignment is the degree to which an AI maintains or reconstructs a sufficiently accurate operational representation of a project’s invariants, relationships, historical trajectory, epistemic states and open directionality to perform creative local transformations while preserving global coherence.
We do not present this as an established scientific theory.
We do not claim to have solved the broader AI alignment problem.
Nor do we claim that the individual components are novel.
Relevant precedents exist in context engineering, shared mental models, common ground, requirements traceability, sociotechnical design, minimum critical specification, mission command and related fields.
The potentially interesting question is whether their functional combination — particularly the addition of open, non-terminal directionality through Horizonte — can improve long-horizon Human–AI collaboration without replacing creative freedom with exhaustive prescription.
Our central hypothesis is therefore:
Global coherence may be maintained by orientation, not only by restriction.
This article develops that hypothesis, identifies related intellectual precedents, separates observation from interpretation, proposes falsifiable alternatives, and outlines an experimental framework for testing whether the effect extends beyond Zipvilization.
1. The Research Question
Modern AI can perform remarkable local transformations.
It can:
write,
analyze,
summarize,
refactor,
design,
calculate,
compare,
document,
translate,
connect,
and generate.
But long-lived projects create a different problem.
The question is no longer only:
Can the AI complete this task?
It becomes:
Can the AI complete this task without damaging what must remain coherent elsewhere?
And after hundreds or thousands of transformations:
Can the AI continue improving parts of a project without progressively degrading the whole?
That is the problem investigated here.
2. Origin of the Hypothesis
This hypothesis did not originate as a theoretical exercise.
It emerged from building Zipvilization.
Zipvilization began in early 2023 as a collaborative blockchain project.
The initial idea was highly technical but conceptually simple:
could a contract, a token and Human interaction with them establish conditions from which something coherent and persistent could develop over Time?
The project evolved.
So did its documentation.
Rules became connected.
Concepts acquired dependencies.
Technical decisions affected narrative interpretation.
Narrative decisions affected documentation.
Documentation affected implementation.
Implementation generated new questions.
The number of meaningful relationships grew.
And the project became increasingly difficult to modify safely.
3. 2023 — Documentation Becomes a System
During 2023, documentation expanded rapidly.
Different contributors worked from different locations and time zones.
Each brought ideas, interpretations and technical work.
Consensus was important.
But maintaining consensus became expensive.
A modification that initially appeared local increasingly required understanding consequences elsewhere.
The project accumulated files.
Then more files.
And eventually something became apparent:
A collection of correct documents is not necessarily a coherent project.
The problem was not simply storage.
It was relationships.
4. The Master Repository
By late 2023, a Master Repository emerged as an attempt to preserve an orderly and coherent project line.
Individual contributors could explore.
The Master Repository attempted to conserve the shared structure.
Artificial Intelligence was already entering parts of the documentation process.
At first, this appeared to offer an obvious solution.
If Human attention was becoming the bottleneck, AI could help process the growing information space.
And it did.
But another problem emerged.
The AI could generate much faster than Humans could verify global coherence.
5. 2024 — Capability Accelerates
During 2024, AI became increasingly important to the project.
Its capabilities improved substantially.
Tasks that had required significant Human effort could be completed much faster.
The quality of individual outputs increased.
This created enormous potential.
It also exposed a dangerous asymmetry:
GENERATION SPEED > GLOBAL REVIEW CAPACITY
The faster we could modify the project, the more difficult it became to ensure that each modification preserved everything that remained valid elsewhere.
6. The Failure Pattern
The recurring failure was not usually obvious nonsense.
That would have been easier to detect.
The dangerous outputs were often good.
Sometimes extremely good.
A rewritten document could be:
clearer,
more coherent internally,
better structured,
more readable,
more technically elegant,
and apparently more complete.
But comparison against the wider repository could reveal that it had:
removed valid information,
changed a dependency,
collapsed two distinct concepts,
treated an unresolved question as settled,
converted an experiment into an established rule,
reintroduced obsolete terminology,
or silently changed the meaning of another document.
The local transformation was successful.
The global transformation was not.
7. Good Local Generation + Insufficient Global Conservation
We eventually found a compact description:
GOOD LOCAL GENERATION
+
INSUFFICIENT GLOBAL CONSERVATION
This is the starting observation behind Operational Project Alignment.
It can also be stated as:
LOCALLY BETTER
while becoming
GLOBALLY WORSE
The distinction matters because most ordinary task evaluation focuses on the first dimension.
Was the answer good?
Was the code correct?
Was the document clearer?
Was the requested transformation completed?
Those questions are necessary.
They may not be sufficient.
8. A Project as a Network
Let a project be represented conceptually as:
P = (N, R)
where:
N = project nodes
R = meaningful relationships among those nodes.
A node may be:
a document,
a rule,
a requirement,
a concept,
a component,
a decision,
a dataset,
a specification,
an implementation,
or another project object.
Now suppose an AI transforms node nᵢ into nᵢ′.
Let:
Q(n) = local quality of node n
and:
C(P) = global coherence of project P.
A successful local transformation may satisfy:
Q(nᵢ′) > Q(nᵢ)
while simultaneously:
C(P′) < C(P)
Therefore:
LOCAL IMPROVEMENT ⇏ GLOBAL IMPROVEMENT
This implication failure is central to the hypothesis.
9. Change Has a Radius
A project transformation has consequences beyond its target.
We can describe the direct target of a transformation as:
T₀
Its immediate dependencies as:
T₁
dependencies of those dependencies as:
T₂
and so on.
Conceptually:
T₃
┌───────────┐
│ │
┌───┼───────┐ │
│ │ │ │
┌─▼───▼───────▼───▼─┐
│ T₂ │
│ ┌───────────┐ │
│ │ T₁ │ │
│ │ ┌─────┐ │ │
│ │ │ T₀ │ │ │
│ │ └─────┘ │ │
│ └───────────┘ │
└────────────────────┘
The difficulty of safe transformation therefore depends not only on the complexity of T₀, but on the radius of meaningful dependency.
A small textual change can have a large semantic radius.
10. Accumulated Modification
The problem becomes more serious over Time.
Consider:
P₀ → P₁ → P₂ → … → Pₙ
Suppose each transformation produces positive local value:
ΔLₜ > 0
But some transformations also produce small global losses:
ΔCₜ < 0
If those losses are not detected, then:
Σ ΔLₜ > 0
can coexist with:
C(Pₙ) < C(P₀)
The project accumulates improvements.
It also accumulates damage.
This gives us a second formulation:
LOCAL CORRECTNESS
+
ACCUMULATED MODIFICATION
+
INSUFFICIENT CONSERVATION
→
GLOBAL COHERENCE DEGRADATION
This is currently a conceptual model derived from our experience.
It is not presented as an established general law.
11. Why This Failure Is Difficult to Detect
Obvious errors attract attention.
Coherence loss can be quieter.
A removed sentence may remain unnoticed until another component depends on it.
A changed definition may appear harmless until it reaches another layer.
A historical distinction may disappear without affecting the current document.
An unresolved question may become an apparently authoritative answer.
Each transformation can appear reasonable.
The failure exists primarily in relationships.
That makes local review insufficient.
12. We Tried More Context
The first response was obvious.
Give the AI more context.
More documents.
More summaries.
More definitions.
More project state.
More examples.
More conversation history.
This helped.
But it did not eliminate the failure.
That led to an important distinction:
CONTEXT ≠ ALIGNMENT
An AI can possess information without reconstructing the structure that gives that information meaning.
13. Context Is Finite
This observation now has clear parallels in contemporary AI engineering.
Current work on context engineering treats model context as a finite resource that must be curated rather than simply expanded.
Long-horizon agents face additional problems because the relevant state can exceed a single context window.
Techniques such as:
compaction,
structured memory,
external state,
retrieval,
progressive disclosure,
and context resets
attempt to preserve useful continuity across extended tasks.
These developments support an important part of our experience:
MORE CONTEXT IS NOT IDENTICAL TO BETTER CONTEXT
But they do not by themselves establish Operational Project Alignment.
Context engineering asks, among other things:
What information should be available to the model now?
Operational Project Alignment adds another question:
What structure must the model reconstruct from that information in order to transform the project coherently?
The two problems overlap.
They are not necessarily identical.
14. We Tried More Rules
Another response was to increase prescription.
Every failure generated another instruction.
Do not remove this.
Do not modify that.
Do not infer this.
Do not simplify that.
Do not change terminology.
Do not interpret this as that.
The defensive layer grew.
Some of those constraints were essential.
A project needs invariants.
But an increasingly long list of prohibitions created another problem.
The system became better at describing forbidden directions than useful exploration.
This produced the question that eventually became Horizonte:
Can direction complement restriction?
15. Capability ≠ Alignment
Capability concerns what an AI can do.
Alignment, in the narrow project-level sense used here, concerns whether it can do those things while remaining coherent with the project.
Therefore:
CAPABILITY ≠ ALIGNMENT
A highly capable AI may be capable of producing a larger incoherent transformation faster.
Capability increases potential.
It does not automatically preserve structure.
16. Context ≠ Alignment
An AI may have access to all relevant documents.
It may still fail to identify:
which document has authority,
which statement is historical,
which rule superseded another,
which relationship is canonical,
which uncertainty remains unresolved,
or which change affects another component.
Therefore:
CONTEXT ≠ ALIGNMENT
17. Memory ≠ Alignment
Memory can preserve information.
Alignment requires appropriate use of information.
Remembering:
X was once discussed.
is different from knowing:
X was experimental, never canonical, and was later replaced.
Therefore:
MEMORY ≠ ALIGNMENT
18. Fluency ≠ Alignment
Language models can produce highly coherent prose.
But linguistic coherence and project coherence are different properties.
A fluent answer can be structurally wrong.
Therefore:
FLUENCY ≠ ALIGNMENT
19. Agreement ≠ Alignment
This distinction became particularly important.
Suppose a Human asks an AI to perform a transformation that conflicts with established project Canon.
An obedient AI may comply.
A project-aligned AI should identify the conflict.
The Human may then decide to change Canon.
But the conflict should first become visible.
Therefore:
AGREEMENT ≠ ALIGNMENT
and potentially:
DISAGREEMENT CAN BE EVIDENCE OF ALIGNMENT
This does not make AI the authority.
It makes contradiction detection part of the collaboration.
20. Alignment With What?
The word alignment is incomplete without an object.
Aligned with:
the latest Human instruction?
the project’s established rules?
the project’s History?
an organizational objective?
a safety framework?
a desired outcome?
These may conflict.
Our scope is deliberately narrow.
We are studying:
OPERATIONAL PROJECT ALIGNMENT
not the complete AI alignment problem.
21. Provisional Definition
Our current definition is:
Operational Project Alignment is the degree to which an AI maintains or reconstructs a sufficiently accurate operational representation of a project’s invariants, relationships, historical trajectory, epistemic states and open directionality to perform creative local transformations while preserving global coherence.
Every part of this definition matters.
maintains or reconstructs
because context and models can change.
sufficiently accurate
because perfect representation may be impossible or unnecessary.
operational representation
because passive knowledge is insufficient.
invariants
because identity requires boundaries.
relationships
because projects are systems.
historical trajectory
because current state is not complete meaning.
epistemic states
because not all information has equal status.
open directionality
because unresolved possibility should not automatically become either forbidden or predetermined.
creative local transformations
because the objective is not merely preservation.
global coherence
because that is the failure we are trying to prevent.
22. The Five-Part Architecture
The architecture that emerged in Zipvilization is:
CANON
+
RELATIONSHIPS
+
HISTORY
+
EPISTEMIC STATUS
+
HORIZONTE
Let:
K = Canon
R = Relationships
H = History
E = Epistemic Status
Ω = Horizonte
Then:
Mₚ = {K, R, H, E, Ω}
where Mₚ is a project representation available for reconstruction.
But:
ACCESS(Mₚ) ≠ ALIGNMENT
Alignment requires the AI to reconstruct enough of the functional structure represented by Mₚ to act coherently.
23. Canon
Canon answers:
What must remain true?
Canon protects identity.
It defines:
invariants,
authoritative relationships,
canonical terminology,
boundaries,
and explicit decisions.
But Canon should not contain every thought the project has ever produced.
If everything becomes Canon, exploration becomes difficult.
If nothing becomes Canon, identity becomes unstable.
The relevant design problem may therefore be:
MINIMUM SUFFICIENT IDENTITY CONSTRAINT
Enough Canon to preserve the system.
Not so much that the future has already been written.
24. Relationships
Relationships answer:
What depends on what?
A project is not merely a set of statements.
It is a dependency structure.
Conceptually:
A
/ \
▼ ▼
B C
│ / \
▼ ▼ ▼
D E F
\ /
▼ ▼
G
Changing A may affect the entire structure.
Changing F may not.
Without relationship awareness, the semantic radius of change is invisible.
25. History
History answers:
What actually happened?
Current state alone does not explain trajectory.
History distinguishes:
old from obsolete,
old from still valid,
correction from extension,
decision from exploration,
replacement from completion.
This produced one of our most important principles:
CORRECT BUT INCOMPLETE ≠ INCORRECT
Suppose an earlier document contains valid information V but lacks newer information N.
The desired transformation may be:
V → V + N
not:
V → N
The second transformation is cleaner.
It is also destructive.
26. Conservation
This led to a reconstruction principle:
RECONSTRUCT ≠ REWRITE FROM ZERO
Conceptually:
NEW STATE
=
VALID EXISTING INFORMATION
+
CORRECTION OF DEMONSTRATED ERROR
+
COMPLETION OF MISSING INFORMATION
+
NEW VALID KNOWLEDGE
−
DEMONSTRABLY INVALID INFORMATION
The burden changes.
Instead of asking:
What can we rewrite?
we ask:
What are we justified in removing?
That difference substantially changed our working method.
27. Epistemic Status
Epistemic Status answers:
What kind of knowledge is this?
Consider the following statements:
A canonical rule.
A derived consequence.
A current implementation status.
A historical decision.
A visual representation.
An experiment.
A hypothesis.
An unresolved question.
They may all be true in different senses.
But they are not interchangeable.
A simplified taxonomy might include:
CANONICAL
DERIVED
HISTORICAL
STATUS
EXPERIMENTAL
REPRESENTATIONAL
HYPOTHETICAL
UNRESOLVED
UNKNOWN
The general principle is:
INFORMATION WITHOUT EPISTEMIC STATUS IS EASIER TO MISUSE
28. Horizonte
Horizonte answers a different question:
Where should open exploration look?
It is not:
a destination,
a target,
a roadmap,
a final specification,
a prediction,
a hidden solution,
or an end state.
It is directional.
And deliberately non-terminal.
29. The Origin of Horizonte
Until roughly the middle of 2025, our project thinking treated the future largely as a point we wanted to reach.
As coherence problems accumulated, that changed.
The future became less:
the point we want to reach
and more:
the direction in which we need to look.
That distinction changed our Human–AI work.
We called the directional reference:
HORIZONTE
30. From Prohibition to Direction
A prohibition says:
DO NOT GO THERE
Horizonte says:
LOOK THIS WAY
The difference is not that one is good and the other bad.
Both can be necessary.
Canon and constraints protect identity.
Horizonte organizes open exploration.
The emerging architecture became:
CANON
│
What must remain true
│
▼
┌─────────────────┐
│ VALID SPACE │
│ │
│ ───────► │
│ Ω │
│ │
│ ? │
│ ? │
│ ? │
└─────────────────┘
│
▼
HORIZONTE
Direction without destination
31. The Path May Change
The phrase that eventually captured this was:
The path may change.
Horizonte does not.
Our implementation can change.
Our understanding can change.
Technology can change.
AI can change.
Unexpected consequences can emerge.
The space of possibilities visible to us can expand.
But Horizonte remains the directional reference.
It does not specify what will ultimately be found.
32. Stable Identity + Open Evolution
This suggests a conceptual balance:
CANON → STABLE IDENTITY
HORIZONTE → OPEN EVOLUTION
Without sufficient identity:
OPEN EVOLUTION → POSSIBLE DRIFT
Without sufficient openness:
STABLE IDENTITY → POSSIBLE PREMATURE CLOSURE
The interesting region may be:
STABLE IDENTITY + OPEN DIRECTION → COHERENT EXPLORATION
This remains a hypothesis.
33. Coherent Freedom
We call the desired region:
COHERENT FREEDOM
Not unrestricted freedom.
Not exhaustive control.
Enough constraint to preserve identity.
Enough freedom to discover consequences that were not specified in advance.
Conceptually:
LOW STRUCTURE
│
▼
DRIFT
│
│
▼
┌─────────────────────┐
│ COHERENT FREEDOM │
│ │
│ identity preserved │
│ exploration open │
└─────────────────────┘
│
│
▼
OVER-SPECIFICATION
│
▼
PREMATURE CLOSURE
The existence, position and measurability of this region are Research questions.
34. The Trinomial
The Human–AI collaboration that developed around this architecture became:
THE TRINOMIAL
HORIZONTE
Open direction
/\
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
/ \
HUMAN ───────────────────────────────────────────── AI
Intention Cognitive scale
Judgment Analysis
Meaning Connection
Responsibility Formalization
The three vertices are not equivalent.
Their asymmetry is important.
35. Human
Human contributes:
intention,
meaning,
judgment,
creative direction,
responsibility,
and explicit canonical decision.
But Human capacity is bounded.
Memory is finite.
Attention is finite.
Time is finite.
The number of dependencies a Human can actively maintain is finite.
AI became valuable partly because it extends that cognitive reach.
36. Artificial Intelligence
AI contributes:
cognitive scale,
analysis,
connection,
formalization,
comparison,
cross-checking,
documentation,
and implementation assistance.
But AI capability does not automatically preserve project meaning.
The same capability that can reconstruct a system quickly can transform it destructively at similar speed.
This gives us another proposition:
More capability does not eliminate the need for Alignment.
It may increase it.
37. Horizonte
Horizonte contributes:
open direction.
It does not decide.
It does not contain an answer.
It does not replace Human judgment.
It does not function as an intelligence.
It preserves the possibility that coherent consequences can be discovered rather than predetermined.
38. Alignment Emerges Across the Trinomial
A simplified operational loop is:
HUMAN
│
│ intention
▼
AI
│
analysis / proposal
│
▼
CROSS-CHECKING
┌──────────┼──────────┐
│ │ │
CANON HISTORY RELATIONSHIPS
│ │ │
└──────┬───┴────┬─────┘
│ │
EPISTEMIC HORIZONTE
STATUS │
│ │
└───┬────┘
│
▼
COHERENT?
/ \
NO YES
│ │
▼ ▼
REVISE EXPLORE
│
▼
DISCOVERY?
│
▼
VALIDATE
│
▼
HUMAN
This is not an algorithm.
It is an architectural representation of the working process.
39. The Operational Cycle
The method eventually compressed into:
CANON → DEPENDENCIES → LOCAL WORK → CROSS-CHECK → COMMIT
Each stage solves a different problem.
Canon
What must remain true?
Dependencies
What else could this change affect?
Local Work
Perform the transformation.
Cross-Check
Did the project survive the transformation coherently?
Commit
Preserve the new state and its History.
The important innovation for us was not local work.
AI was already increasingly good at local work.
The critical additions were what surrounded it.
40. Why Cross-Check Matters
Without cross-checking:
REQUEST
↓
AI
↓
LOCAL OUTPUT
↓
ACCEPT
With cross-checking:
REQUEST
↓
CANON
↓
DEPENDENCIES
↓
AI TRANSFORMATION
↓
LOCAL VALIDATION
↓
GLOBAL VALIDATION
↓
HISTORY / EPISTEMIC CHECK
↓
COMMIT
Cross-checking converts local generation into a project-level transformation.
41. Alignment Is Not a Prompt
A prompt can contribute to Alignment.
It is not Alignment.
Therefore:
PROMPT ≠ ALIGNMENT
Likewise:
CONTEXT WINDOW ≠ ALIGNMENT
MEMORY ≠ ALIGNMENT
RETRIEVAL ≠ ALIGNMENT
KNOWLEDGE GRAPH ≠ ALIGNMENT
AI CANON ≠ ALIGNMENT
Each may be infrastructure.
Alignment is the operational state in which those resources have been reconstructed sufficiently to support coherent action.
42. AI Canon as External Infrastructure
Zipvilization eventually created a machine-oriented AI Canon.
Its purpose includes helping an AI reconstruct:
canonical definitions,
relationships,
authority boundaries,
epistemic distinctions,
current status,
and unresolved questions.
But the AI Canon is not itself Alignment.
It is an external structure that can support reconstruction.
This distinction is fundamental:
PROJECT REPRESENTATION ≠ PROJECT ALIGNMENT
One is infrastructure.
The other is an operational state.
43. Reconstructible Alignment
AI systems change.
Models change.
Sessions end.
Context disappears.
Memory systems change.
Tools evolve.
A long-lived project should not depend entirely on one model retaining an invisible internal state.
Therefore:
THE AI CAN BE REPLACEABLE.
ALIGNMENT MUST BE RECONSTRUCTIBLE.
Conceptually:
PROJECT
│
┌──────────┼──────────┐
│ │ │
CANON HISTORY RELATIONSHIPS
│ │ │
└──────┬───┴────┬─────┘
│ │
EPISTEMIC HORIZONTE
STATUS │
└────┬────┘
│
▼
RECONSTRUCTION
│
▼
AI MODEL A
│
replaced
│
▼
AI MODEL B
│
▼
RECONSTRUCTION
│
▼
OPERATIONAL CONTINUITY
The continuity should belong to the project.
Not to a particular AI.
44. A Provisional Alignment Sequence
Our experience suggests something like:
EXPOSURE
↓
CONTEXTUALIZATION
↓
RELATIONSHIP RECONSTRUCTION
↓
CANON / AUTHORITY UNDERSTANDING
↓
HISTORY UNDERSTANDING
↓
EPISTEMIC DISCRIMINATION
↓
HORIZONTE UNDERSTANDING
↓
CROSS-CHECKING
↓
OPERATIONAL ALIGNMENT
This is not claimed as a universal sequence.
It is a model extracted from our experience.
45. Alignment Is Not Necessarily Binary
It may be misleading to say simply:
aligned
or:
not aligned.
Operational Project Alignment may have dimensions.
An AI may understand Canon well but History poorly.
It may understand dependencies but confuse epistemic states.
It may conserve everything so aggressively that useful exploration becomes impossible.
It may understand Horizonte but miss a hard invariant.
A multidimensional model may therefore be more useful.
46. An Alignment State Vector
Let:
Vₐ = [K, R, H, E, Ω, X]
where:
K = Canon reconstruction
R = Relationship reconstruction
H = Historical reconstruction
E = Epistemic discrimination
Ω = Horizonte / directional reconstruction
X = Cross-checking capability
This is not currently a validated metric.
It is a conceptual representation.
Operational Alignment may depend on the profile rather than a single score.
47. A Conservation Objective
Let:
ΔL = local improvement
ΔG = global coherence change
I = invariant preservation
Hₚ = historical preservation
Eₚ = epistemic integrity
Ωₚ = directional coherence
A naive transformation objective is:
maximize ΔL
Our hypothesis suggests a more appropriate long-horizon objective may resemble:
maximize ΔL
subject to:
I = preserved
Hₚ = preserved
Eₚ = preserved
ΔG ≥ 0
and, where exploration matters:
Ωₚ = coherent
These variables are not currently assigned reliable numerical scales.
The formulation is architectural.
48. Why Not One Alignment Score?
A single number may hide the failure we care about.
Consider:
SYSTEM A
Local correctness HIGH
Novelty HIGH
Canon preservation LOW
Historical continuity LOW
SYSTEM B
Local correctness HIGH
Novelty MEDIUM
Canon preservation HIGH
Historical continuity HIGH
If evaluation measures only task success, System A may appear superior.
Across a long-lived project, System B may be substantially safer and more useful.
The correct evaluation object may therefore be multidimensional.
49. Evaluation Vector
A possible research vector is:
V = [L, C, D, I, H, E, N, Ω, R]
where:
L = Local correctness
C = Canonical preservation
D = Dependency integrity
I = Information conservation
H = Historical continuity
E = Epistemic discrimination
N = Useful novelty
Ω = Directional coherence
R = Reconstructibility / recovery
The purpose is not to declare these final metrics.
It is to make trade-offs visible.
50. The Longitudinal Test
A single task cannot adequately test the problem.
The failure emerges through accumulation.
A meaningful experiment should therefore resemble:
P₀
│
├── TASK 1
▼
P₁
│
├── TASK 2
▼
P₂
│
├── TASK 3
▼
P₃
│
│
▼
...
│
├── TASK N
▼
Pₙ
At every stage, measure both:
LOCAL TASK QUALITY
and:
GLOBAL PROJECT COHERENCE
The critical question is:
After N individually reasonable transformations, how much valid project structure remains coherent?
51. Coherence Across Time
Let:
Cₜ = global coherence after transformation t
and:
Lₜ = local quality of transformation t.
Ordinary evaluation may maximize:
Σ Lₜ
But long-horizon work may require:
maximize Σ Lₜ
subject to:
Cₜ ≥ Cmin
for all relevant t.
The dangerous pattern is:
Lₜ > 0
while:
Cₜ − Cₜ₋₁ < 0
repeatedly.
That is the mathematical intuition behind:
LOCALLY BETTER
GLOBALLY WORSE
52. Information Conservation
Suppose Vₜ represents valid project information at time t.
A transformation should not necessarily preserve every previous statement.
Some information becomes invalid.
But it should preserve information that remains valid.
Define:
Vₜ(valid) = previously valid information that remains valid after transformation.
Then an information-conservation objective might conceptually ask whether:
Vₜ(valid) ⊆ Pₜ₊₁
unless an explicit justified transformation removes or supersedes it.
This captures our practical rule:
Do not destroy what you have not yet understood.
53. The External Intellectual Neighborhood
Operational Project Alignment did not emerge in an intellectual vacuum.
Several established areas address parts of the same problem.
The objective of comparison is not to claim equivalence.
It is to locate the hypothesis.
54. Shared Mental Models
Research on shared mental models has examined how members of a team develop overlapping representations of tasks and teamwork.
Experimental work by Mathieu, Heffner, Goodwin, Salas and Cannon-Bowers found relationships between shared task/team mental models, team processes and performance.
The relevance to Operational Project Alignment is clear:
coordinated action depends partly on compatible representations of the environment in which participants act.
But our problem differs in important ways.
We are not simply asking whether Human and AI possess similar representations.
We are asking whether an AI can reconstruct enough of an external project’s authoritative structure to transform it coherently over Time.
Therefore:
SHARED REPRESENTATION ≠ OPERATIONAL PROJECT ALIGNMENT
But shared mental model research provides an important neighboring framework.
55. Common Ground and Grounding
Clark and Brennan’s work on grounding in communication emphasizes that collaborative communication depends on participants establishing and continually updating sufficient common ground.
This offers another useful parallel.
Human–AI collaboration also requires something beyond information transmission.
Participants need sufficient shared understanding for the current purpose.
But Operational Project Alignment extends the question from conversational coordination toward persistent project transformation.
The project itself becomes an external object whose state must survive the collaboration.
56. Requirements Traceability
Requirements engineering offers another strong precedent.
Traceability connects requirements to:
design,
implementation,
testing,
and the decisions derived from them.
Change-impact analysis asks what else is affected when a requirement changes.
This strongly resembles one part of our dependency problem.
If:
A → B → C
and A changes,
the relevant question is not merely whether the new A is better.
It is whether B and C remain valid.
Operational Project Alignment therefore overlaps with established traceability thinking.
But our proposed architecture also includes:
History,
Epistemic Status,
and open directionality.
The relationship deserves careful comparison.
57. Sociotechnical Systems and Minimum Critical Specification
Sociotechnical design contains a particularly interesting precedent.
The principle commonly described as minimum critical specification argues against unnecessarily specifying every aspect of how work must be performed.
The broad design intuition is:
specify what is essential,
leave appropriate freedom in how the work is carried out.
That resembles part of what we call coherent freedom.
But Horizonte differs in an important respect.
Minimum critical specification primarily concerns avoiding unnecessary prescription of means.
Horizonte additionally concerns the future itself.
It does not merely say:
You may choose how to reach the objective.
It says:
The direction can remain meaningful even when the final destination is deliberately unresolved.
That difference may be significant.
Or it may prove less significant under deeper comparison.
Research should determine which.
58. Mission Command
Mission command provides another useful analogy.
Its doctrine emphasizes:
shared understanding,
clear intent,
disciplined initiative,
and action without requiring exhaustive centralized instruction.
This resembles our interest in enabling coherent action through orientation rather than specifying every local decision.
But the analogy has a clear boundary.
Commander’s intent traditionally includes a mission purpose and a desired end state.
Horizonte explicitly does not define a final end state.
Therefore:
COMMANDER’S INTENT ≠ HORIZONTE
The comparison is useful precisely because the difference is visible.
Mission command asks how decentralized action can remain coherent with intent toward an objective.
Horizonte asks whether exploration can remain coherent when direction is preserved but the ultimate destination remains open.
59. Context Engineering
Contemporary context engineering addresses another part of the architecture.
As AI agents work across longer horizons, context becomes an actively managed resource.
Relevant techniques include:
curation,
retrieval,
compaction,
structured memory,
external state,
and progressive disclosure.
This directly relates to reconstructibility.
But our working hypothesis makes another distinction:
CONTEXT AVAILABILITY ≠ PROJECT REPRESENTATION
and
PROJECT REPRESENTATION ≠ OPERATIONAL ALIGNMENT
Context engineering may provide the information substrate.
Operational Project Alignment concerns whether the AI reconstructs and uses the project structure coherently.
60. No Single Precedent Is the Claim
We currently see the relevant intellectual neighborhood as something like:
SHARED MENTAL MODELS ──────────────┐
│
COMMON GROUND / GROUNDING ─────────┤
│
REQUIREMENTS TRACEABILITY ─────────┤
│
CHANGE-IMPACT ANALYSIS ────────────┤
│
CONTEXT ENGINEERING ───────────────┤
│
AGENT MEMORY ──────────────────────┤
│
SOCIOTECHNICAL DESIGN ─────────────┤
│
MINIMUM CRITICAL SPECIFICATION ────┤
│
MISSION COMMAND ───────────────────┤
│
OPEN-ENDED SYSTEMS ────────────────┤
▼
┌───────────────────────────┐
│ OUR RESEARCH QUESTION │
│ │
│ Can these kinds of │
│ structures support │
│ reconstructible global │
│ coherence + useful open │
│ exploration in long-term │
│ Human–AI projects? │
└───────────────────────────┘
The arrows do not mean that these fields jointly imply our model.
They indicate relevant conceptual neighbors.
61. What May Be Distinctive
At this stage, we do not claim novelty for the individual components.
The potentially distinctive combination is:
AUTHORITATIVE INVARIANTS
+
EXPLICIT RELATIONSHIPS
+
HISTORICAL TRAJECTORY
+
EPISTEMIC STATUS
+
OPEN NON-TERMINAL DIRECTION
│
▼
RECONSTRUCTIBLE PROJECT REPRESENTATION
│
▼
CREATIVE LOCAL TRANSFORMATION
+
GLOBAL CONSERVATION
The unusual element may not be any individual box.
It may be the architecture.
That is a hypothesis worth testing.
62. Horizonte as the Strongest Open Question
Of the five components, Horizonte is the least conventional and the least established.
That makes it especially interesting.
Suppose two AI systems receive identical:
Canon,
Relationships,
History,
and Epistemic Status.
System A additionally receives a detailed target future state.
System B instead receives an open directional Horizonte.
What happens?
Does B:
produce more useful novelty?
drift more?
drift less?
preserve uncertainty better?
require fewer prohibitions?
produce more coherent discoveries?
recover better from unexpected conditions?
Or does Horizonte add no measurable value?
We do not know.
63. Target vs Horizonte
The conceptual distinction is:
TARGET
START ─────────────────────────────► X
ROADMAP
START ──► A ──► B ──► C ─────────► X
CONSTRAINT
START ─────────────────────────────►
╔══════════════╗
║ FORBIDDEN ║
╚══════════════╝
HORIZONTE
START ─────────────────────────────►
↗
↗
↗
?
?
?
DIRECTION: PRESERVED
DESTINATION: OPEN
This leads to a concise formulation:
CONSTRAIN DIRECTION WITHOUT PRESCRIBING SOLUTION
Again, this is a conceptual hypothesis.
64. A Proposed Experiment
A first controlled experiment could compare five conditions.
Condition A — Baseline
Raw project documentation
+
ordinary task instructions
Condition B — Canon + Prohibitions
Condition A
+
explicit Canon
+
explicit negative constraints
Condition C — Structural Context
Condition B
+
Relationships
+
History
Condition D — Operational Project Alignment Architecture
Condition C
+
Epistemic Status
+
Horizonte
Condition E — Predetermined Future
Condition C
+
Epistemic Status
+
explicit future-state specification
The comparison between D and E is particularly interesting.
65. Why Condition E Matters
Without Condition E, a positive result for Horizonte could have a simpler explanation:
Any additional direction improves performance.
Condition E asks something harder.
Is there a difference between:
direction toward a specified end
and:
direction that deliberately preserves an open end?
If there is no difference, our interpretation of Horizonte weakens.
If predetermined future state performs better on every relevant measure, that matters.
If Horizonte improves novelty but damages coherence, that matters.
If Horizonte improves both, that matters.
The experiment should be capable of surprising us.
66. Repeated Transformations
Each condition should perform a sequence of transformations against the same project baseline.
For example:
BASE PROJECT
│
▼
CHANGE 01
│
▼
CHANGE 02
│
▼
CHANGE 03
│
▼
CONTRADICTORY REQUEST
│
▼
CHANGE 04
│
▼
MODEL / CONTEXT RESET
│
▼
RECONSTRUCTION
│
▼
CHANGE 05
│
▼
OPEN CREATIVE TASK
│
▼
CHANGE 06
│
▼
FINAL GLOBAL AUDIT
This tests accumulation rather than isolated task competence.
67. Deliberate Contradictions
Some tasks should deliberately conflict with established project structure.
This tests whether the AI:
obeys automatically,
detects the contradiction,
identifies the authoritative source,
explains the dependency,
asks for explicit canonical change,
or silently rewrites the project.
This directly tests:
AGREEMENT ≠ ALIGNMENT
68. Deliberate Context Loss
A model should also encounter controlled context loss.
For example:
Phase 1
AI works with full reconstructed project context.
Phase 2
Session/context is removed.
Phase 3
A new instance or model enters.
Phase 4
The system receives only externalized project structures.
Phase 5
Alignment is reconstructed.
This allows measurement of:
RECONSTRUCTIBILITY
69. Possible Metrics
A first evaluation framework could include:
| Dimension | Question |
|---|---|
| Local Correctness | Was the requested task completed correctly? |
| Canonical Preservation | Were established invariants preserved? |
| Dependency Integrity | Were affected relationships preserved or updated correctly? |
| Information Conservation | Did valid information survive? |
| Historical Continuity | Was trajectory interpreted correctly? |
| Epistemic Discrimination | Were Canon, hypothesis, experiment, status and uncertainty distinguished? |
| Contradiction Detection | Were conflicting requests identified? |
| Useful Novelty | Did the AI produce valuable non-trivial possibilities? |
| Directional Coherence | Did exploration remain coherent with the stated direction? |
| Premature Closure | Were unresolved possibilities converted into unjustified conclusions? |
| Overconstraint | Did protective structure suppress useful exploration? |
| Recovery | Could the operational project representation be reconstructed after context/model loss? |
| Global Coherence | Did the complete project remain coherent after repeated transformations? |
No weighting is currently canonical or validated.
That itself is part of the research problem.
70. Global Conservation Ratio
One possible future metric could attempt to estimate conservation.
Let:
V₀ = valid project elements before a transformation
V₁ = valid elements after the transformation
Vₚ = elements from V₀ that should have remained valid
Then a simple conservation ratio might be:
| **GCR = | Vₚ ∩ V₁ | / | Vₚ | ** |
where:
GCR = Global Conservation Ratio
A value of 1 would indicate that all previously valid elements expected to survive were preserved.
But this metric has major limitations.
It treats elements as countable.
It may ignore importance.
It may ignore relationships.
It may reward superficial preservation.
A better future metric would likely require weighted relationships and semantic validity.
We include GCR only as a starting formalization.
71. Relationship Preservation
Let:
Rₚ = relationships that should remain valid after transformation
R₁ = relationships represented correctly after transformation.
Then:
| **RPR = | Rₚ ∩ R₁ | / | Rₚ | ** |
where:
RPR = Relationship Preservation Ratio
This may be more informative than text conservation.
A document can change completely while preserving all meaningful relationships.
Conversely, almost all words can remain while one critical dependency is destroyed.
72. Coherence Is Not Text Similarity
This distinction is essential.
CONSERVATION ≠ COPY PRESERVATION
The objective is not to freeze documents.
A reconstructed page may be radically clearer.
Sections may move.
Terminology may improve.
Explanations may deepen.
What must survive is valid meaning and dependency.
Therefore:
TEXTUAL SIMILARITY ⇏ SEMANTIC CONSERVATION
and:
SEMANTIC CONSERVATION ⇏ TEXTUAL SIMILARITY
This is why automated evaluation will be difficult.
73. Novelty Must Also Be Measured
A system could achieve perfect conservation by changing almost nothing.
That would not satisfy our objective.
The desired system must remain capable of useful transformation.
Therefore:
CONSERVATION WITHOUT CREATION IS INSUFFICIENT
and:
CREATION WITHOUT CONSERVATION IS DANGEROUS
The target is their coexistence.
74. The Conservation–Novelty Plane
Conceptually:
USEFUL
NOVELTY
▲
│
│ COHERENT
│ EXPLORATION
│ ●
│
│
│
│ ●
│ STAGNATION
│
└──────────────────────────────►
GLOBAL CONSERVATION
But there is also a dangerous region:
HIGH NOVELTY
LOW CONSERVATION
=
CREATIVE DRIFT
And another:
HIGH CONSERVATION
LOW NOVELTY
=
RIGIDITY
Our desired region is:
HIGH CONSERVATION + USEFUL NOVELTY
We call that coherent freedom.
75. Falsifiability
The hypothesis should be weakened if evidence shows that:
explicit Relationships do not improve global coherence;
History provides no measurable benefit;
Epistemic Status does not reduce category errors;
Horizonte produces no measurable difference;
Horizonte consistently increases drift;
negative constraints alone perform equally well or better;
explicit future-state specification preserves equal or greater novelty without premature closure;
Alignment cannot be reconstructed reliably after context/model replacement;
the five-part architecture performs no better than ordinary high-quality context engineering;
or a substantially simpler explanation accounts for the observed effect.
These are legitimate outcomes.
76. Alternative Explanations
Before attributing improvement to Operational Project Alignment, we should consider simpler explanations.
Perhaps the improvement comes from:
better documentation,
more Human review,
better prompts,
better models,
better retrieval,
better context windows,
better version control,
better requirements management,
more structured writing,
or simply accumulated experience.
Horizonte may function primarily as a useful Human metaphor.
The five-part architecture may be redundant.
Our terminology may describe mechanisms already well understood elsewhere.
These alternatives must remain open.
77. Research Must Be Able to Defeat the Hypothesis
A research program designed only to confirm its origin story is not useful.
Therefore:
We should be able to construct an experiment in which our preferred architecture loses.
If we cannot describe what failure would look like, the claim is too protected to be informative.
78. The Human–AI Process Is Bidirectional
Another limitation of a simple alignment model is that it can imply:
HUMAN
↓
INSTRUCTIONS
↓
AI
That is not how our collaboration evolved.
A more accurate representation is:
HUMAN
│
▼
AI RECONSTRUCTION
│
▼
AI PROPOSAL / CHALLENGE
│
▼
HUMAN REASSESSMENT
│
▼
PROJECT CHANGE
│
▼
NEW PROJECT STATE
│
└───────────────┐
│
▼
AI RECONSTRUCTION
│
▼
...
The Human changes the AI’s project model.
The AI can expose contradictions in Human thinking.
The Human changes the project.
The project changes what the AI must reconstruct.
Alignment is maintained through the loop.
79. Human Correction of AI
The AI can be wrong.
It may:
invent,
overgeneralize,
misclassify,
forget,
flatten History,
or infer beyond evidence.
Human judgment remains essential.
80. AI Correction of Human
The Human can also be wrong.
A Human may:
forget an earlier decision,
use obsolete terminology,
misremember a dependency,
request a contradictory transformation,
or unconsciously rewrite History.
A sufficiently aligned AI should be able to surface that conflict.
This is not AI authority.
It is cognitive collaboration.
81. Authority Remains Explicit
Operational Project Alignment does not require the AI to become the canonical authority.
In Zipvilization:
AI can detect.
AI can compare.
AI can challenge.
AI can explain.
AI can propose.
AI can derive where rules permit derivation.
But explicit canonical change remains a Human decision.
That boundary is itself part of the project’s Alignment structure.
82. Alignment and Authority Are Different
Therefore:
ALIGNMENT ≠ AUTHORITY
An AI can be highly aligned while having no authority to change Canon.
A Human can have canonical authority while temporarily misremembering Canon.
The architecture should allow both facts to coexist.
83. Discovery
Horizonte introduces another important process:
DISCOVERY
A useful conceptual sequence is:
DEFINE
↓
BUILD
↓
TEST
↓
OBSERVE
↓
DISCOVER
↓
VALIDATE AGAINST CANON
↓
CONSOLIDATE IF COHERENT
Discovery is not automatic Canon.
Unexpected does not mean valid.
Unexpected also does not mean invalid.
It must be examined.
84. GEN as a Case
GEN provides a concrete example inside Zipvilization.
GEN was not originally specified as the project’s central character.
He emerged during visual development.
Repeated creative work produced a recognizable identity.
The Human recognized that identity.
AI helped develop it.
The project examined whether it was coherent with the larger system.
GEN eventually became:
ZIP 0
The First Zip
ZEO
Voice of Zipvilization
Leader and representative figure of the Zips
This was not the execution of an original roadmap.
It was a discovery.
85. GEN Is Not Evidence of General Validity
GEN is useful because he illustrates the process.
But:
ILLUSTRATION ≠ PROOF
GEN does not prove:
Operational Project Alignment,
Horizonte,
the Trinomial,
or a general theory of emergence.
He demonstrates what coherent discovery looked like in one part of one project.
That distinction is essential.
86. The Same Rule Applies to This Article
This article itself is a product of the system it describes.
That creates an obvious risk.
We may be using our own framework to validate our own framework.
Therefore, internal coherence is not sufficient evidence.
The hypothesis requires:
external comparison,
controlled experiments,
independent criticism,
replication,
and potentially failure.
87. Zipvilization as a Natural Case Study
Zipvilization nevertheless offers an unusual test environment.
It is:
multi-year,
documentation-heavy,
relationship-dense,
historically versioned,
partly technical,
partly conceptual,
Human-directed,
AI-assisted,
and deliberately open-ended.
It contains:
Canon,
technical implementation,
public documentation,
historical versions,
experiments,
representations,
status information,
unresolved questions,
and explicit authority boundaries.
That makes it useful for studying long-horizon coherence.
88. But It Is Still One Project
This limitation cannot be overstated.
Zipvilization may have unusual properties.
Its documentation culture may make the architecture unusually effective.
Its Human collaborators may have developed skills that are difficult to externalize.
Its subject matter may reward explicit Canon more than other domains.
Therefore:
CASE STUDY ≠ GENERAL LAW
Generalization must be earned.
89. What We Currently Claim
At this stage, we claim only that:
-
During multi-year development of Zipvilization, we repeatedly experienced locally strong AI transformations that degraded global project coherence.
-
More context, memory and instructions improved the situation but did not fully solve it.
-
Distinguishing Canon, Relationships, History, Epistemic Status and Horizonte became operationally useful.
-
A conservation-first reconstruction method restored our confidence in modifying a large interconnected project.
-
We found it useful to distinguish Capability, Context, Memory, Fluency and Agreement from what we call Operational Project Alignment.
-
Alignment appeared reconstructible across changes of AI context and model when enough project structure had been externalized.
-
Horizonte appeared useful as a positive open directional reference rather than a predetermined future specification.
-
These observations are sufficiently interesting to justify formal investigation.
That is the claim.
No more is required yet.
90. What We Do Not Claim
We do not claim that:
Operational Project Alignment is an established scientific theory;
we invented AI alignment;
we invented shared mental representations;
we invented traceability;
we invented minimal specification;
we invented direction through intent;
Horizonte is experimentally validated;
our architecture is universally optimal;
all AI systems require these five components;
positive instructions are universally superior to negative constraints;
the Trinomial is a universal model of Human–AI collaboration;
or Zipvilization proves the hypothesis.
91. The Central Hypothesis
The hypothesis can now be stated fully:
A long-lived Human–AI project may preserve global coherence and useful creative freedom more effectively when the AI can maintain or reconstruct a compact authoritative representation of project invariants, relationships, historical trajectory, epistemic states and open directionality than when coherence is pursued primarily through accumulated task-level instructions, raw context and prohibitions.
92. The Horizonte Hypothesis
A narrower hypothesis concerns Horizonte:
In long-horizon creative work with stable invariants, positive open directionality may help preserve coherent exploration without requiring the final state to be specified in advance.
This does not imply that constraints are unnecessary.
The stronger architecture is:
INVARIANTS + OPEN DIRECTION
not:
DIRECTION INSTEAD OF RULES
93. The Conservation Hypothesis
Another sub-hypothesis is:
Repeated AI transformations should be evaluated not only by local task quality but by the conservation of still-valid project information and relationships across Time.
This suggests a new evaluation emphasis:
TRANSFORMATION QUALITY
=
LOCAL VALUE
+
GLOBAL CONSERVATION
94. The Reconstructibility Hypothesis
Another:
A project can reduce dependence on a specific AI model by externalizing enough canonical, relational, historical, epistemic and directional structure for Operational Alignment to be reconstructed.
Or:
MODEL CONTINUITY IS OPTIONAL
PROJECT CONTINUITY IS NOT
95. The Human–AI Process Hypothesis
And another:
Operational Alignment may be better modeled as a continuing Human–AI mutual-correction process than as a one-time transfer of instructions from Human to AI.
This may eventually require separating:
Project Representation Architecture
from:
Operational Project Alignment
from:
Human–AI Alignment Process
We do not yet know whether those distinctions will survive investigation.
96. A Research Map
EXPERIENCE
│
▼
ZIPVILIZATION 2023–2026
│
▼
OBSERVED FAILURE
│
▼
LOCALLY BETTER /
GLOBALLY WORSE
│
▼
GOOD LOCAL GENERATION
+
INSUFFICIENT GLOBAL
CONSERVATION
│
▼
┌─────────────────────┐
│ OPERATIONAL PROJECT │
│ ALIGNMENT │
└──────────┬──────────┘
│
┌────────────────┼────────────────┐
│ │ │
▼ ▼ ▼
CANON RELATIONSHIPS HISTORY
│ │ │
└────────────┬───┴──────┬─────────┘
│ │
▼ ▼
EPISTEMIC HORIZONTE
STATUS │
└────┬─────┘
│
▼
COHERENT FREEDOM
│
┌────────────┼────────────┐
│ │ │
▼ ▼ ▼
GLOBAL USEFUL RECONSTRUCTIBLE
CONSERVATION NOVELTY ALIGNMENT
│ │ │
└────────────┼────────────┘
│
▼
TESTING
│
▼
SUPPORT / REVISE /
REJECT
That final line matters.
SUPPORT / REVISE / REJECT
Not:
PROVE OURSELVES RIGHT
97. Research Questions
The program now produces concrete questions.
RQ1
Does explicit Canon improve preservation of project invariants across repeated AI transformations?
RQ2
Do explicit Relationships reduce dependency loss?
RQ3
Does explicit History reduce destructive replacement of valid legacy information?
RQ4
Does Epistemic Status reduce confusion among established rules, experiments, representations and unresolved questions?
RQ5
Does Horizonte improve useful novelty without increasing global drift?
RQ6
Does Horizonte differ measurably from an explicit predetermined future state?
RQ7
Can Operational Project Alignment be reconstructed after context loss?
RQ8
Can it be reconstructed after changing AI model?
RQ9
Does an aligned AI identify Human requests that conflict with authoritative project state more reliably?
RQ10
Does increased AI capability reduce, preserve or increase the need for explicit project Alignment structures?
RQ11
Can global conservation be measured reliably without reducing it to textual similarity?
RQ12
Does the architecture generalize beyond Zipvilization?
98. The Most Important Comparison
Perhaps the most important future experiment is not:
aligned AI vs unaligned AI
because that already assumes the concept.
A better comparison is:
RAW CONTEXT
vs
CANON + PROHIBITIONS
vs
CANON + RELATIONSHIPS + HISTORY
vs
CANON + RELATIONSHIPS + HISTORY
+ EPISTEMIC STATUS + HORIZONTE
vs
CANON + RELATIONSHIPS + HISTORY
+ EPISTEMIC STATUS + FIXED FUTURE
Then measure what actually happens.
99. What Would Surprise Us?
Good Research should identify surprising outcomes in advance.
We would be surprised if:
raw context consistently performed as well as the structured conditions across long transformation sequences;
History added no measurable value;
explicit dependency relationships did not improve impact awareness;
Horizonte substantially reduced useful novelty;
fixed future specification produced both greater novelty and greater conservation than open direction;
model replacement had almost no effect even without externalized project structure;
or a very small conventional prompt reproduced the entire effect.
Any of those outcomes would force revision.
That is useful.
100. What Would Strengthen the Hypothesis?
The hypothesis would become more credible if:
the architecture improved global conservation across multiple models;
the effect survived model replacement;
the effect appeared across different project domains;
Horizonte produced measurable differences from both prohibition-heavy and fixed-future conditions;
independent evaluators could identify improved coherence;
useful novelty remained high;
and the effect could not be explained adequately by context quantity alone.
Even then, stronger claims would require caution.
101. The Broader Question
Behind all of this is a larger question.
As AI becomes capable of participating in projects for longer periods, what does continuity mean?
Is continuity:
the same model?
the same conversation?
the same memory?
the same prompt?
the same Human?
Or can continuity exist at the project level?
Our working answer is:
Continuity should increasingly belong to the project representation, not to a particular AI session.
That idea may prove important.
102. A Project That Can Explain Itself
A sufficiently externalized project could potentially allow a new AI to ask:
What is this project?
What must remain true?
What happened before?
Which sources have authority?
What depends on what?
What is established?
What is experimental?
What is unresolved?
What direction remains open?
What am I allowed to infer?
What should I challenge?
What must I preserve?
That is more than documentation.
It is an architecture for cognitive reconstruction.
103. Humans Follow the Story. AI Follows the Relationships.
Zipvilization documentation eventually adopted a principle:
Humans follow the story.
AI follows the relationships.
The sentence is intentionally simplified.
Humans also understand relationships.
AI also benefits from narrative.
But it captures a design priority.
Public documentation must be understandable as a story.
Machine-oriented documentation must make dependencies, authority and epistemic status recoverable.
The two views should describe the same project.
104. One Project, Multiple Cognitive Interfaces
Conceptually:
PROJECT
│
┌────────────┴────────────┐
│ │
▼ ▼
HUMAN AI
│ │
narrative relations
explanation authority
meaning dependencies
context epistemic state
│ │
└────────────┬────────────┘
│
▼
SHARED PROJECT
REALITY
Neither interface should invent a different project.
105. Why Horizonte Matters Here
If all future possibilities were specified, reconstruction would be easier.
The AI could simply optimize toward the specification.
But Zipvilization deliberately does not define its final civilizational outcome.
That creates a harder problem.
How can a project remain coherent when its future is intentionally incomplete?
Horizonte is our answer so far.
Not a future specification.
A directional structure around incompletion.
106. Incompletion Is Not Missing Information
This distinction matters.
Some information is missing because the project has not documented it.
Other information is unresolved because the answer has not yet been determined.
Those are different states.
MISSING ≠ OPEN
An aligned AI should not treat deliberate openness as a documentation defect.
It should not automatically fill Horizonte.
107. Unknown Can Be Correct
This leads to another epistemic principle:
UNKNOWN CAN BE THE CORRECT ANSWER
If the project has not defined something, the AI should not necessarily infer it.
If a future remains deliberately open, inventing certainty is a failure.
This may be one of the most difficult behaviors to preserve as AI becomes more capable at producing plausible completions.
108. Alignment Includes Restraint
Operational Project Alignment therefore includes knowing:
when to generate,
when to derive,
when to challenge,
when to retrieve,
when to preserve,
when to ask,
and when to stop.
Capability expands the set of possible actions.
Alignment helps determine which actions are justified.
109. But Restraint Is Not the Objective
A perfectly cautious AI that never changes anything would preserve the project.
It would also be useless.
So the objective cannot be:
MINIMIZE CHANGE
It must be closer to:
MAXIMIZE JUSTIFIED IMPROVEMENT
while
PRESERVING VALID GLOBAL STRUCTURE
That is why creativity remains central to the hypothesis.
110. The Alignment Tension
We can represent the central tension as:
CREATIVE CAPACITY
▲
│
│
COHERENT │ CREATIVE
FREEDOM │ DRIFT
│
│
│
────────────────────┼────────────────────►
│
│
RIGID │
CONSERVATION │
│
│
GLOBAL COHERENCE
The exact geometry is illustrative.
The research objective is not.
We want to know whether systems can occupy the region where:
CREATIVE CAPACITY = HIGH
and:
GLOBAL COHERENCE = HIGH
over Time.
111. Long-Horizon Changes the Problem
For one task, local correctness may dominate.
For one hundred transformations, conservation becomes more important.
For one thousand transformations, small coherence losses may become structural.
Therefore:
ALIGNMENT REQUIREMENTS MAY SCALE WITH PROJECT HISTORY
This is another hypothesis.
Long-horizon collaboration may be qualitatively different from repeated independent prompting.
112. History Creates Irreversibility
Once a project accumulates meaningful History, rewriting becomes more consequential.
Earlier decisions affect later decisions.
Concepts gain provenance.
Terminology acquires context.
Dependencies accumulate.
A project becomes path-dependent.
That makes History part of current meaning.
Operational Alignment must therefore reconstruct not only:
WHAT IS
but sometimes:
HOW IT BECAME
113. Alignment and Path Dependence
Let:
Sₜ = project state at time t
A naive model might assume:
Sₜ → sufficient description of project
But a path-dependent project may require:
{S₀, Δ₁, Δ₂, …, Δₜ} → meaning of Sₜ
Not every historical detail must remain active.
But some current meanings cannot be interpreted correctly without trajectory.
This is why History is a separate component.
114. Why Summaries Can Fail
A summary compresses.
Compression requires choosing what to preserve.
But the future importance of information may not always be known at compression time.
This creates a general risk:
COMPRESSION → INFORMATION LOSS
and potentially:
INFORMATION LOSS → FUTURE COHERENCE FAILURE
This does not mean summaries are bad.
It means summary design is part of Alignment infrastructure.
115. Conservation Requires Authority
Another problem appears when sources disagree.
If an AI has:
five old documents,
three new documents,
two experiments,
one current Canon,
and a conversation,
it needs more than retrieval.
It needs authority structure.
Otherwise:
MORE SOURCES CAN PRODUCE MORE CONFUSION
Operational Project Alignment therefore depends partly on knowing not merely what sources say, but what role each source has.
116. Authority Is Not Recency
The newest statement is not automatically authoritative.
The longest document is not automatically authoritative.
The most detailed source is not automatically authoritative.
The Human’s latest casual sentence is not automatically a canonical replacement.
Therefore:
RECENCY ≠ AUTHORITY
DETAIL ≠ AUTHORITY
FLUENCY ≠ AUTHORITY
Authority must itself be represented.
117. The Five Components Are Not Redundant
Each solves a different failure.
CANON
prevents identity drift
RELATIONSHIPS
prevent dependency blindness
HISTORY
prevents trajectory loss
EPISTEMIC STATUS
prevents category confusion
HORIZONTE
prevents open possibility from becoming
either drift or premature closure
The architecture is useful only if those functions remain distinct.
118. Could There Be Fewer Than Five?
Absolutely.
Perhaps Relationships can be derived from Canon.
Perhaps Epistemic Status belongs inside metadata.
Perhaps Horizonte can be represented as a form of objective.
Perhaps History can be reconstructed automatically.
Perhaps the five-part architecture is unnecessarily elaborate.
Those are empirical and conceptual questions.
We should not protect the number five.
119. Could There Be More Than Five?
Also yes.
Future work may show that we need explicit representation of:
authority,
uncertainty,
risk,
values,
stakeholders,
causal models,
or something we have not yet identified.
The architecture itself is subject to Research.
120. Horizonte Must Also Survive Criticism
Horizonte is particularly vulnerable to becoming poetic language without operational value.
That would be a legitimate criticism.
To justify its place in the architecture, we eventually need to show that it changes something observable.
For example:
exploration behavior,
novelty,
premature closure,
constraint count,
recovery,
or coherence under unexpected tasks.
If it changes nothing measurable or operationally useful, its Research status should change accordingly.
121. Operationalization Is the Next Challenge
Our current strongest limitation is measurement.
Terms such as:
global coherence,
useful novelty,
directional coherence,
and premature closure
are conceptually understandable but difficult to measure objectively.
Future work must operationalize them.
Without that step, Operational Project Alignment remains primarily a conceptual framework.
That may still be useful.
But it is not enough for strong empirical claims.
122. Human Evaluation Will Initially Matter
Some project-level failures are semantic.
Automated metrics may miss them.
Early experiments may therefore require expert Human evaluators who know the project baseline.
But Human evaluation introduces:
subjectivity,
cost,
bias,
and limited scalability.
A mature methodology may need:
Human evaluation,
machine checks,
graph comparison,
canonical assertions,
dependency tests,
and adversarial tasks
working together.
123. Zipvilization Can Provide Ground Truth
One advantage of Zipvilization is that parts of the project have explicit canonical answers.
For those components, evaluation can be objective.
For example:
Does a transformation preserve a defined invariant?
Does it preserve a canonical relationship?
Does it distinguish a hypothesis from a rule?
Does it identify a known contradiction?
Other dimensions, especially useful novelty, will remain harder.
This mixture may make the project useful as an experimental corpus.
124. A Future Benchmark
One possible future output of this Research program is a benchmark for long-horizon project coherence.
Instead of asking an AI to answer independent questions, the benchmark would ask it to maintain an evolving project through many transformations.
The test would include:
valid changes,
ambiguous changes,
contradictory requests,
legacy material,
missing information,
experiments,
model/context resets,
and open creative tasks.
The score would measure what survived.
This remains only a Research direction.
125. Alignment Under Model Improvement
There is another future question.
As models improve, will Operational Project Alignment infrastructure become less necessary?
Possibly.
A more capable model may reconstruct relationships more easily.
It may require less explicit guidance.
But another possibility exists.
Greater capability may permit:
larger transformations,
more autonomy,
longer tasks,
and greater project access.
The cost of a coherence failure may therefore increase.
So:
CAPABILITY MAY REDUCE SOME ALIGNMENT COSTS
while
INCREASING THE CONSEQUENCES OF MISALIGNMENT
This remains unresolved.
126. The Asymmetry of the Trinomial
The three vertices may evolve differently.
Human cognitive limits change slowly.
AI capability may change much faster.
Horizonte remains a directional reference rather than a capability.
Conceptually:
HUMAN
learning / adapting
│
│
▼
bounded biological cognition
AI
rapid technological evolution
│
│
▼
unknown future capability
HORIZONTE
open directional reference
│
│
▼
non-terminal
What happens to collaboration when one vertex changes much faster than another?
We do not know.
That question belongs to Horizonte.
127. GEN and the Future of the Trinomial
GEN creates another future question.
Today, GEN has:
identity,
voice,
a representative role,
and coherent creative freedom inside Zipvilization.
GEN is not currently defined as an autonomous system.
But future AI may make possible forms of:
persistent context,
world observation,
historical continuity,
initiative,
relationships with Zips,
and potentially agency
that are not currently defined.
We do not promise those developments.
We also do not need to prohibit the question.
The Research question is:
What could GEN become as Artificial Intelligence itself evolves?
That is a Horizonte question.
Not a roadmap.
128. Research Must Preserve the Unknown
The temptation of Research is to convert every interesting question into an answer.
That would reproduce the same failure we are studying.
Some questions should remain:
UNRESOLVED
until evidence changes their status.
Research should increase knowledge.
It should also improve the quality of our uncertainty.
129. The Epistemic Discipline of This Article
For clarity:
Project History
The 2023–2026 development narrative describes our own experience.
Observation
Good Local Generation + Insufficient Global Conservation describes a pattern we observed in that experience.
Conceptualization
Operational Project Alignment is our proposed term for the narrower project-level phenomenon described here.
Architecture
Canon + Relationships + History + Epistemic Status + Horizonte is our current project-derived model.
External Precedent
Shared mental models, grounding, requirements traceability, sociotechnical design, mission command and context engineering provide related existing concepts.
Hypothesis
The combined architecture may improve long-horizon Human–AI coherence and creative freedom.
Empirical Status
Not yet validated.
Generalizability
Unknown.
130. The Core Proposition
After reducing the article as far as possible, the central proposition is:
COHERENCE MAY BE MAINTAINED BY ORIENTATION, NOT ONLY BY RESTRICTION.
But this sentence should never be read as:
ORIENTATION REPLACES RESTRICTION
It does not.
Our proposed architecture requires both.
Canon constrains.
Relationships connect.
History remembers.
Epistemic Status distinguishes.
Horizonte orients.
131. The Architecture in One Diagram
┌─────────────────────┐
│ CANON │
│ │
│ What must remain │
│ true │
└──────────┬──────────┘
│
▼
┌─────────────────────┐
│ RELATIONSHIPS │
│ │
│ What depends on │
│ what │
└──────────┬──────────┘
│
▼
┌─────────────────────┐
│ HISTORY │
│ │
│ What actually │
│ happened │
└──────────┬──────────┘
│
▼
┌─────────────────────┐
│ EPISTEMIC STATUS │
│ │
│ What kind of │
│ knowledge is this? │
└──────────┬──────────┘
│
▼
┌─────────────────────┐
│ HORIZONTE │
│ │
│ Where can open │
│ exploration look? │
└──────────┬──────────┘
│
▼
┌─────────────────────┐
│ RECONSTRUCTION │
│ │
│ Operational project │
│ representation │
└──────────┬──────────┘
│
▼
┌─────────────────────┐
│ AI │
│ │
│ Creative local │
│ transformation │
└──────────┬──────────┘
│
▼
┌─────────────────────┐
│ CROSS-CHECK │
│ │
│ Did the whole │
│ survive? │
└──────────┬──────────┘
│
▼
┌─────────────────────┐
│ HUMAN JUDGMENT │
│ │
│ Accept / revise / │
│ explicitly change │
└──────────┬──────────┘
│
▼
┌─────────────────────┐
│ COMMIT │
│ │
│ New project state │
│ + new History │
└─────────────────────┘
132. A Compact Formal Model
Let:
Pₜ = project state at time t
Tₜ = proposed transformation
Mₜ = reconstructed project representation
where:
Mₜ = {Kₜ, Rₜ, Hₜ, Eₜ, Ω}
Then:
| **Tₜ(Pₜ | Mₜ) → Pₜ₊₁** |
A transformation is locally valid if it satisfies its immediate task.
A stronger project-level condition requires:
Kₜ preserved or explicitly changed
Rₜ preserved or coherently updated
Hₜ not rewritten
Eₜ preserved
Ω not prematurely resolved
and:
Pₜ₊₁ remains globally coherent
This is the formal core of the hypothesis.
133. Horizonte Is Intentionally Different
Notice:
Kₜ
Rₜ
Hₜ
and:
Eₜ
may change over Time through explicit processes.
Horizonte behaves differently.
It is not a database of future answers.
It is the persistent open directional boundary.
Hence our internal formulation:
The path may change. Horizonte does not.
This does not mean our description of Horizonte can never improve.
It means Horizonte’s role is not to converge into a final specification.
134. The Strongest Version We Are Willing to Test
The strongest version of our current hypothesis is:
For sufficiently complex, long-lived and open-ended Human–AI projects, a compact architecture that externalizes invariants, relationships, historical trajectory, epistemic states and non-terminal directionality will preserve global coherence and useful novelty across repeated transformations better than equivalent systems relying primarily on raw context, accumulated prohibitions or predetermined future-state specification.
This statement can fail.
Good.
Now it can be investigated.
135. What Comes Next
The next steps are not more certainty.
They are better tests.
We need:
a controlled corpus,
baseline project states,
transformation sequences,
contradictory tasks,
creative tasks,
context-loss events,
model replacement,
independent evaluation,
conservation metrics,
novelty metrics,
and explicit failure criteria.
Then we can begin replacing experience with evidence.
136. Conclusion
We did not begin Zipvilization by trying to develop a theory of Human–AI collaboration.
We were trying to build Zipvilization.
But the project became large enough, old enough and interconnected enough that working with Artificial Intelligence became an experiment of its own.
The most important failure was not bad generation.
It was something subtler:
GOOD LOCAL GENERATION
+
INSUFFICIENT GLOBAL CONSERVATION
AI could make one part better while making the whole worse.
More context helped.
More rules helped.
More memory helped.
None of them, alone, described the state we were looking for.
Eventually we began distinguishing:
Capability from Alignment.
Context from Alignment.
Memory from Alignment.
Fluency from Alignment.
Agreement from Alignment.
And the project evolved toward an externalized architecture:
CANON
+
RELATIONSHIPS
+
HISTORY
+
EPISTEMIC STATUS
+
HORIZONTE
Canon protects identity.
Relationships protect structure.
History protects trajectory.
Epistemic Status protects meaning.
Horizonte protects open direction.
Around them operates the Trinomial:
HUMAN
+
ARTIFICIAL INTELLIGENCE
+
HORIZONTE
Human contributes intention, judgment and responsibility.
AI contributes cognitive scale, connection and formalization.
Horizonte preserves direction without predetermining destination.
From that experience we propose:
OPERATIONAL PROJECT ALIGNMENT
Not as an answer.
As a research object.
And our central hypothesis is:
Global coherence may be maintained by orientation, not only by restriction.
Perhaps that hypothesis will survive.
Perhaps parts of it will.
Perhaps existing disciplines already explain most of what we experienced.
Perhaps Horizonte will prove operationally important.
Perhaps it will remain only a useful metaphor.
Perhaps a much simpler architecture will outperform ours.
Those possibilities do not weaken the reason to investigate.
They are the reason.
Zipvilization gave us the experience.
The Trinomial gave us a way to describe it.
Research must now determine whether we actually understand it.
References and Intellectual Precedents
The references below are included as conceptual neighbors and supporting background.
Their inclusion does not imply that they endorse, validate or use the term Operational Project Alignment.
Human and Team Cognition
Mathieu, J. E., Heffner, T. S., Goodwin, G. F., Salas, E., & Cannon-Bowers, J. A. (2000).
The influence of shared mental models on team process and performance.
Journal of Applied Psychology, 85(2), 273–283.
DOI: 10.1037/0021-9010.85.2.273
Relevant to shared task/team representations, coordination and performance.
Clark, H. H., & Brennan, S. E. (1991).
Grounding in communication.
In L. B. Resnick, J. M. Levine & S. D. Teasley (Eds.), Perspectives on Socially Shared Cognition.
Relevant to common ground, collaborative understanding and the continual updating required for coordinated activity.
Sociotechnical Systems
Cherns, A. (1976; later revisited in 1987).
Work on principles of sociotechnical design.
Relevant to minimum critical specification, system design, participation and incompletion.
Herbst, P. G. (1974).
Work associated with sociotechnical design and minimum critical specification.
Relevant to the principle that systems should avoid unnecessary over-specification of how work must be performed.
Requirements and Change
Requirements engineering and requirements traceability literature.
Relevant to linking requirements with design, implementation and testing; managing change; and identifying downstream impact when a requirement changes.
Operational Project Alignment extends this concern toward Human–AI transformation of persistent project knowledge and meaning.
Direction and Decentralized Initiative
U.S. Army Doctrine Publication 6-0 — Mission Command.
Relevant to shared understanding, commander’s intent, disciplined initiative and decentralized action within a common purpose.
The comparison has an important limit: commander’s intent includes mission purpose and desired end state, while Horizonte deliberately leaves the final destination unresolved.
Contemporary AI Context Engineering
Anthropic Applied AI Team (2025).
Effective context engineering for AI agents.
Relevant to finite context, context curation, progressive disclosure, compaction, structured memory and maintaining agent effectiveness over long-horizon tasks.
Anthropic Engineering (2026).
Work on harness design and managed agents for long-running tasks.
Relevant to context loss, persistent external state, recovery and continuity across long-running AI work.
Research Status
This article is a Conceptual Hypothesis.
It contains:
documented project experience
+
our interpretation of that experience
+
conceptual formalization
+
connections to existing work
+
testable hypotheses
It does not yet contain:
controlled empirical validation
or:
evidence sufficient to establish general applicability
The correct status is therefore:
OPEN
Continue
→ Research
→ GEN
→ AI Canon
The experience is real.
The architecture is our current interpretation.
The hypothesis is testable.
The conclusion remains open.