Doctrine
Polymantic Systems
One substrate. Many valid views.
Here is the strangest fact in applied AI right now, made stranger by how quickly everyone stopped finding it strange: the same model, with the same weights, is a different mind in a different world.
The numbers are public. On SWE-bench Pro, Anthropic's own scaffold carries Claude Opus to 69.2 percent; under Scale's standardized SEAL harness, the best Claude run lands at 51.9. Same family, same benchmark, seventeen points apart. Morph, tracking the gap across its coding leaderboards, draws the conclusion plainly: "the scaffold around the model accounts for more variance than swapping frontier models."
Practitioners have internalized this faster than the discourse has. Anyone who has been shipping agent systems since 2023 has felt it in their bones; the craft only recently picked up a name, harness engineering, along with paper taxonomies and folk wisdom about the wrapper mattering more than the model. Anthropic's guidance on writing tools for agents is blunt: "Even small refinements to tool descriptions can yield dramatic improvements," citing a state-of-the-art SWE-bench Verified run.
Open Interpreter added the latest data point and a phrase for one corner of the practice: per-model harness emulation. A Codex-shaped world for GPT models, a Claude Code-shaped world for Claude, each model dropped into the environment it was trained to drive.
Pause on what this whole pattern actually is. An industry hand-carving a separate world for every consumer of meaning, because nothing exists underneath those worlds to compile them from. Harness engineering is real engineering, and it is also a symptom. The agent stack has discovered that every intelligence needs its own presentation of reality, and it is meeting that demand the way scribes met the demand for books: one copy at a time, by hand.
The constellation was already there
Before walking the stars, let me state the principle plainly, because the rest of this essay stands on it. Every system of meaning has two layers. The referent: the thing you are actually talking about. The view: the presentation of that thing, shaped for whoever is consuming it. The discovery, made over and over, is that these layers can separate, and that everything good happens when they do. Views can multiply without the referent fragmenting.
Hold that, and the harness numbers stop being strange. What the agent community proved with benchmark money, at least five other fields had already found on their own. They do not cite each other. The stars have been sitting there, unconnected, for over a century.
Philosophy · 1892
Frege noticed that the morning star and the evening star are the same object, Venus, known two different ways. The deep part is not the duality. It is that discovering they co-refer is real information. If meaning were singular, "the morning star is the evening star" would be as empty as "Venus is Venus." It is not. Perspective carries content the referent alone does not.
Linguistics · every conversation
One word, many related senses: polysemy. "Account" means five different things at your company, and conversation almost never breaks, because context governs which sense is active. Language runs plural meaning at planetary scale, governed by nothing heavier than context.
Databases · 1975
The ANSI/SPARC architecture split the database into one conceptual schema and many external views, each user seeing the projection that fits their work. The design was so successful it became invisible. You have used it every working day since, without thinking about it.
Ontology engineering · 1990s
Cyc, the most ambitious single-ontology project ever attempted, set out to hand-build one globally consistent knowledge base and learned that knowledge refuses the shape. It reorganized into microtheories: locally consistent bundles that may contradict each other, governed at the seams. The team that tried hardest to build one true ontology was the first to learn that one true ontology does not survive contact with real knowledge.
Sociology · 1989
Star and Griesemer watched a natural history museum coordinate trappers, amateur collectors, and taxonomists around the same specimens, and named the mechanism the boundary object: plastic enough to adapt to each group's needs, robust enough to hold one identity across all of them. Their finding compresses to three words: cooperation without consensus.
Five fields, one discovery: meaning that works at scale is plural at the surface and shared underneath. The views multiply. The referent holds. Each field built its own corner of the machine. Nobody built the general one.
The agent industry is only the latest arrival, rediscovering the principle one expensive harness at a time. That is reason enough to give the pattern a name. I have started calling systems that honor it polymantic systems.
The monomantic default
The opposite instinct has a name and a history. Master data management. The golden record. A decade of Customer 360 programs. One account status, one health score, one source of truth: the conviction that every important concept should eventually resolve to a single canonical meaning. The modern semantic layer inherited the posture; dbt's metric definitions and Looker's models exist so that every dashboard in the company agrees on one number.
For dashboards, the posture is correct. Analytics needs shared definitions. If the churn number in the CFO's dashboard disagrees with the board deck, something is genuinely broken.
Dashboards consume definitions. Agents consume worlds.
But intelligent software does not just store and report. It interprets, plans, escalates, and acts. An agent needs to know what a definition means for action: what it can touch, whose judgment applies, which exceptions matter, when to stop and ask a human.
A human expert needs provenance and the right to disagree. An evaluator needs a rubric compiled from the same semantics it is scoring. Same referent, different worlds. The moment a reasoner consumes your semantic layer, the layer stops being documentation and becomes runtime behavior.
The polymantic condition
And the organization underneath the software has been polymantic all along. Consider enterprise readiness.
| Perspective | Valid view of the same referent |
|---|---|
| Sales | A revenue blocker. |
| Engineering | A capability bundle. |
| Legal | A liability posture. |
| Finance | An unpriced support obligation. |
| Support | A future escalation pattern. |
Five perspective-scoped claims over one shared referent. Each can be valid. Each can also be wrong, stale, under-evidenced, or politically convenient. The discovery that two of them describe the same situation is itself information, the enterprise version of learning that the morning star is the evening star. Collapse the claims into one field and you do not remove nuance. You destroy the finding.
Lay the failure modes side by side. A monomantic system collapses the table and loses the perspectives. A fragmented system lets every team keep its own tool and loses the referent. Domain-driven design splits the difference by federating: bounded contexts, treaties at the borders, translation negotiated at every seam. A polymantic system makes a different move. It keeps one substrate and compiles the seams.
What lives in the substrate
The substrate is where meaning gets a stable address. It holds the things that must stay interoperable even while the operational view changes.
- Referents: the things the system is talking about.
- Slots: durable semantic positions where meaning can be expressed, constrained, interpreted, and governed.
- Shapes: graphs of slots that encode a domain frame.
- Claims: perspective-scoped assertions about a referent.
- Evidence: the source material behind a claim.
- Policies: rules for visibility, authority, reconciliation, decay, and mutation.
- Versions: semantic history, not just document history.
The view is everything compiled out of that substrate: a UI, an API schema, an MCP tool surface, a prompt packet, a retrieval policy, an eval rubric, a workflow, a permission set.
The database reader will object, correctly, that views existed in 1975. They did, for structure. An external schema tells you which columns you can see. It carries no perspective, no evidence, no authority, and no policy for what happens when two views disagree.
Structure was solved fifty years ago. Meaning is the layer agents consume, and meaning never got its ANSI/SPARC.
The basic polymantic architecture
The substrate preserves shared meaning. Views project that meaning into the interface required by the work.
Perspective-scoped claims
The unit that makes plurality safe is the perspective-scoped claim. A claim is not a value in a field. It carries the referent it is about, the slot it fills, the perspective it speaks from, the evidence behind it, the authority that may assert or change it, its timestamp and visibility, and the policy that decides its future.
| Claim component | Why it matters |
|---|---|
| Referent | What this claim is about. |
| Slot | Which semantic position the claim fills. |
| Perspective | Whose view this claim represents. |
| Evidence | Why the system should believe or inspect it. |
| Authority | Who is allowed to assert or change it. |
| Policy | Whether to preserve, reconcile, decay, restrict, or escalate it. |
Held this way, the system never has to pretend that five claims are the same statement, and it never loses the fact that they may be about the same operational situation. That is the difference between plurality and chaos.
Reconciliation is policy
None of this argues for worshipping disagreement. A contract gets signed or it does not. A deployment ships or it does not. Some slots must collapse to one canonical state because the organization has to move. The question is who decides, and on what grounds.
Run the readiness table forward: an agent drafting the renewal needs to know whether it may offer the enterprise tier, so the slot resolves under its policy. Engineering holds authority on capability. Finance's claim escalates because it is eight months stale. The resolution lands with provenance intact, and the plurality survives underneath it. Collapse happens because the action needed it, not because the system was uncomfortable holding two meanings at once.
- Preserve disagreement when plurality is informative.
- Canonicalize when action requires one state.
- Escalate when authority is unclear.
- Decay claims when freshness matters.
- Restrict claims when visibility is sensitive.
- Refuse modeling when legibility would do more harm than good.
These verbs have pedigree. A strand of requirements engineering spent the 1990s arguing that conflicts between stakeholder views are signal to manage, not disease to eradicate; one of its capstone papers is literally titled making inconsistency respectable. And the last verb runs deepest. James C. Scott documented what happens when institutions impose legibility on realities that worked precisely because they were informally known: the scientifically planned forest grows beautifully for one rotation and dies in the second. Some knowledge degrades when you formalize it. Declining to model is a policy decision too, and a polymantic system gets to make it deliberately instead of by omission.
The write-back problem
Agents are not just consumers of all this. They fill slots, propose claims, detect drift, stage mutations. And the moment they write, they collide with one of the oldest hard problems in data systems. Reading through a view is easy. Writing through one is not. Bancilhon and Spyratos formalized the danger in 1981: an update made through a projection is only safely translatable back to the base when it leaves everything outside the projection untouched. Most view updates are ambiguous, and the safe ones have to be specified, never assumed.
That result is forty-five years old, and the agent stack is rediscovering it the hard way, one over-permissioned write tool at a time. The write side is where governance earns its keep: which mutations this view licenses, what they may touch, and what they may not. (The write-back problem deserves an essay of its own. Consider this the trailer.)
Where this is heading
Step back and the whole agent stack looks the way the harness does. Prompts, tools, retrieval, memory, evals, schemas: each primitive quietly carries its own private semantics, each one re-describes the domain by hand, and drift is present from the first deployment because nothing underneath holds the meaning still.
Engineers already trust the fix in another domain. Event sourcing keeps one append-only log and compiles as many read models as consumers need. The shape is proven. The substrate was just never semantic.
The market has already priced half the argument. Palantir's documentation describes its ontology as "an operational layer for the organization" that "serves as a digital twin": objects and links, but also actions and functions. That is the semantic layer walking out of the warehouse, and the valuation says the walk is worth something.
But Foundry's ontology is canonical. One twin, one truth per object. The most successful monomantic ontology ever built, arriving at exactly the moment the consumers of meaning started multiplying.
So here is the trend worth naming before it coalesces on its own. Every artifact in the stack that touches meaning is becoming a compilation target. The harness was simply the first one measurable enough to prove it. Tool surfaces, context packets, eval rubrics, API schemas, interfaces: each is a view some consumer needs, and nearly all of them are still hand-made. The teams that win the next phase will not be the ones with the best hand-carved worlds. They will be the ones that own the substrate the worlds compile from.
That is the bet Penumbra is built on: intelligent software needs shared structure, not shared sameness. A substrate of referents, claims, evidence, and policy, with views compiled out per consumer, each shaped to its work, none pretending to be the only truth. Not one perfect ontology. A place where meaning has a stable address, enough structure to stay interoperable, and enough plurality to stay true to the work.
The future is not one ontology. It is one substrate, many valid views.