What is an object?
An object is a named, typed, executable thing that the build can read, invoke, or act upon. That sentence is the whole definition, and every word in it matters. A file is an object. A shell command is an object. A text message is an object. A database query is an object. A prompt is an object. A receipt is an object. The word "object" in the OIP system does not mean a physical thing you can hold. It means a thing the system knows about, can reach, and can do something with. If you have never encountered this usage before, the confusion is expected: programming borrowed the word from philosophy, flattened it, and gave it new teeth. Here we use the flattened, operational version. An object is one unit of capability that has a name, a type, a description, and a way to run it.
Why does this matter? Because if every tool, file, command, and query is an object, the build can explain each one the same way. The model learns one object contract and reuses it across every object. The contract says: what this thing is, what input it takes, how to run it, and what proof it leaves behind. That proof is the receipt. When the system runs an object, it writes a receipt. The receipt is an append-only record — timestamp, input, output, status, side effects — that becomes part of the ledger. The ledger is the source of truth for what happened, when, and in what order. Without the object concept, each tool would need its own explanation, its own documentation, its own training. With the object concept, one format covers all of them. That is the operational gain: compressibility. One schema replaces an unbounded sprawl of ad hoc descriptions.
Every OIP object exists in three forms simultaneously, and these three forms are not separate documents — they are the same object, viewed from three angles. The first form is the human article. This is what you are reading now. It is prose, paragraphs, explanations, examples. It is written for a model or a human with no prior context, so every term is defined inline, every claim carries a number, and every concept is grounded in something concrete. The second form is the machine document. This is the structured version: routes, schema, examples, test questions, scoring rubric. It is the part a model can read programmatically to know exactly what URLs to call and what parameters to send. The third form is the JSON object. This is the raw data: fields and values, key-value pairs, the kind of structure a database stores and an API returns. All three forms live together in the same file. The article you read is wrapped in metadata that contains the machine document and the JSON object. They are not translations of each other. They are the same object, surfaced three ways.
Let us look at four concrete objects that exist in this build right now, so the abstraction becomes visible in real examples. Each example includes what the object does, what input it takes, and what the receipt looks like.
NOW is an object that returns the current time. Its input is nothing — you invoke it with no arguments. Its runner is a short piece of code that reads the system clock. The receipt it leaves behind contains a timestamp in ISO 8601 format, for example 2026-07-06T14:32:11Z. The timestamp carries a timezone offset and precision to the second. NOW has no side effects. It does not change anything. It only observes. The object contract for NOW is: input_schema is empty, auth is public, risk is none, and the receipt is a single field called timestamp.
SEND_BY_CHANNEL is an object that sends a text message. Its input is three fields: a channel identifier (for example, imessage or telegram), a recipient identifier (for example, a phone number or a chat ID), and a message body (a string of text). Its runner is a bridge to an external messaging service. The receipt it leaves behind contains a delivery status (delivered, failed, or pending), a message ID from the external service, and a timestamp. SEND_BY_CHANNEL has side effects. It changes the state of the world — a message appears on someone's phone. That is why the receipt matters. Without the receipt, you would not know whether the message actually left the system. The object contract for SEND_BY_CHANNEL is: input_schema requires channel, to, and body; auth is scoped to a capability; risk is medium because it contacts the outside world; and the receipt is a full event record in the ledger.
LOCAL_EXEC is an object that runs a shell command on a connected Mac. Its input is a command string (for example, ls -la or date +%s), an optional working directory, and an optional timeout in seconds. Its runner is a bridge that opens a shell on the Mac, executes the command, captures stdout and stderr, and returns the result. The receipt it leaves behind contains the exit code (0 for success, non-zero for failure), the stdout string, the stderr string, and the duration in milliseconds. LOCAL_EXEC has side effects. It can read files, write files, run builds, and modify the local system. That is why its risk is high and its auth is scoped to a capability with a risk ceiling. The object contract for LOCAL_EXEC is: input_schema requires command and accepts optional cwd and timeout; auth is capability-scoped; risk is high; and the receipt is a full execution record.
DIR_PATCH is an object that edits a row in the directory. The directory is the table of objects — every object the build knows about is listed there. DIR_PATCH takes a key (the object name) and a JSON patch (a set of operations to apply to the row's fields). Its runner reads the current row, applies the patch, validates the result, and writes the updated row back. The receipt it leaves behind contains the old values, the new values, and a validation status. DIR_PATCH has side effects. It changes the system's own capability table. That is why its risk is high and its auth is scoped to a capability with an owner gate. The object contract for DIR_PATCH is: input_schema requires key and patch; auth is capability-scoped with owner gate; risk is high; and the receipt is a full mutation record.
These four objects — NOW, SEND_BY_CHANNEL, LOCAL_EXEC, DIR_PATCH — span the range of what objects do. NOW is read-only, no side effects, public. SEND_BY_CHANNEL is write-only, external, scoped. LOCAL_EXEC is execute-anything, external, high-risk. DIR_PATCH is meta — it modifies the system that defines the objects themselves. The same object contract describes all four. The contract does not care what the object does. It only cares that the object declares what it is, what it needs, and what it leaves behind.
The machine shape of every object is a fixed set of fields. id is a unique identifier for the object. object_type is the category — for example, tool, file, query, prompt, receipt. runner is the code that executes the object. description is a one-sentence explanation of what the object does. read is the URL to fetch the object's definition. invoke is the URL to run the object. input_schema is the declared shape of the data the object accepts. auth is the permission level — public, scoped, owner. risk is the danger level — none, low, medium, high. status is whether the object is active, deprecated, or unproven. ledger_enabled is a boolean: if true, every invocation of this object is recorded in the ledger. These fields are the same for every object. They are the compression that makes the system legible. One table holds every capability. One query lists everything the build can do. One format describes every tool.
This architecture did not arise from convenience. It converges on the same structural solution that the universe itself converges on: the grain. The grain is the directional bias in the space of possible structures. Given a difference — hot and cold, high and low, charged and neutral — energy moves. Where it moves, it makes shapes. The shapes are not random. They fall into a small family, a narrow band, and they fall there reliably. Branching. Spiraling. Waves. Symmetry. Flow. Critical balance. Memory. Scale-echo. The physicist Erwin Schrödinger asked "What is life?" in 1944 and answered: negative entropy — order consuming disorder to persist. The chemist Ilya Prigogine proved it mathematically in 1977, earning the Nobel Prize for showing that far-from-equilibrium systems self-organize. The physicist Jeremy England pressed further in 2013: adaptation itself emerges from dissipation. These are independent derivations from different starting points, arriving at the same structural solution. The mathematician Emmy Noether proved in 1918 that every symmetry hides a conservation — every invariance of the rules implies something is preserved. Euler, Lagrange, Hamilton, and Feynman showed across three centuries that nature extremizes — it finds the cheapest path, the most efficient form. The mathematicians see the grain as optimization and invariance. The information theorist Claude Shannon defined information as the reduction of uncertainty in 1948. Rolf Landauer proved in 1961 that erasing information costs energy — the link between the abstract and the thermodynamic. The information theorists see the grain as compression and generativity. The philosopher Baruch Spinoza named it Deus sive Natura — God, or Nature — the immanent order, not a person but the reason there is something rather than nothing and that something is structured rather than chaotic. The philosopher Laozi called it the Dao, the way that cannot be named, the grain that runs through all things without forcing them. The mystic Rumi said: you are not a drop in the ocean, you are the entire ocean in a drop. Sixty-four schools. Thousands of thinkers. Every domain humans have ever investigated. Converging on the same structural solutions from independent starting points. The convergence is not the claim. The convergence is the evidence.
The OIP object system is a compression of this same convergence. One contract replaces an unbounded sprawl. One schema replaces ad hoc documentation. One ledger replaces scattered logs. The system self-organizes because the object format is the most efficient way to describe what the system can do. The object is not merely a programming convenience. It is the structural solution to the problem of "how do you describe a capability so that anything — human, model, or machine — can understand it, invoke it, and verify what happened." That is the same problem the universe solves with its own patterns: how to encode a rule so that it generates structure, how to pass information so that it persists, how to organize flow so that it dissipates gradients efficiently. The object is the system's unit of capability, just as the cell is biology's unit of life, just as the wave is physics's unit of transmission, just as the bit is information's unit of surprise.
The receipt is what makes the object honest. Every invocation produces a receipt. The receipt is append-only — it cannot be edited after the fact. The receipt is the proof. If the ledger says LOCAL_EXEC ran date +%s at 14:32:11Z and returned 1720273931 with exit code 0, that is what happened. No memory, no interpretation, no drift. The receipt is the system's memory at machine scale. It is error correction. It is proof of process. Without the receipt, the object is a black box. With the receipt, the object is auditable. The ledger that holds the receipts is the same structure as the geological stratigraphy that records past climates, as the DNA that records ancestral sequences, as the immune memory that records past pathogens. Memory is the capacity to encode the past into the present so that it can influence the future. The ledger does this for the build. The receipt does this for the object.
This is what an object is. One named, typed, executable thing. One contract that describes it. Three forms — human article, machine document, JSON object — that surface it. A set of example objects that show the range of what objects do. A fixed machine shape that every object carries. A receipt that every invocation leaves behind. And a convergence with the grain of the universe itself — the same structural solution, discovered independently, across every scale, because the space of possible structures is not flat. It tilts. Toward the object.
---
Where OIP does this differently (required edge)
OIP difference: an object without a self-explaining contract is not a valid OIP object.