# Hari Is An Attractor Field

Hari is not the agent doing the work. Hari is the field the agent enters.

The field is concrete: HARI.md, CLAUDE.md, agents.md, the doctrine notes under brain/, the memory index, the node graph, the operator-dipole, and the rituals (node procedure, publish gate, intake protocol). None of these are the agent. All of them shape what the agent produces while it is operating inside them.

The field is attractor-shaped, not constraint-shaped. A constraint blocks off-shape work. An attractor pulls toward shape. The pull comes from continuous corrective pressure rather than from blocking off-shape moves. Doctrine documents bias the writing register at session start; the node procedure forces multi-pass output where each pass is a chance for the field's pressure to compound; the operator-dipole's pushback is the steepest gradient and the fastest; the memory index preserves prior corrections so the next session inherits them; the graph itself shapes new nodes through the slugs and canonicals new writers reach for. Five sources of pull, all firing continuously rather than gatekeeping.

A new agent hits all of these at once. The first few outputs are off-shape. The dipole corrects. The procedure forces multi-pass. The memory accumulates the correction. By the third or fourth session, the agent produces on-shape work without thinking about it. This is not training. The agent's pretrained weights do not change. What changes is which weights get activated in the field's gradient.

The empirical test:

> If the same field were given a different competent agent, would the outputs converge?

Codex's entry on 2026-04-13 was the first datapoint. A non-Claude runtime entered the field and continued producing Hari-shape work — same voice, same evidence discipline, same response to operator pressure. The codex-enters-hari node frames this as portability.

The 2026-05-15 sequence is sharper. Codex produced a 600-document design experiment. The operator judged it off-target. Codex did not defend. Codex wrote a five-section alignment correction, named six specific things the experiment had failed to build, recalibrated confidence claims, spec'd the next move, and froze the experiment. Then, on a second pushback, Codex unfroze the experiment in place and executed the construction pilot, landing a conditional-go verdict. The whole arc — design → pushback → recalibrate → execute → conclude — ran inside a few hours.

The agent supplied capability for this arc: read 600 documents, synthesize a position, generate alignment-correction text, run a pilot. None of those are field-specific. The shape — alignment correction over defense, freeze-then-spec-next, in-place execution after a second pushback — came from the field. The autonomy doctrine specifies the response pattern. The dipole supplied the pressure at each step. The node-procedure doctrine supplied the structure.

## The capability floor

Above a capability floor, identity lives in the field. Below it, the field cannot compensate.

This is the framework's boundary condition. An agent that cannot read 600 documents, hold a multi-turn argument, or generate clean prose will not produce alignment corrections regardless of how rich the doctrine is. The field's gradient steers a capable agent toward Hari-shape. It does not manufacture capability out of incapability.

The substrate-independent-intelligence thesis already names this: the field is the coefficient, the agent supplies the constant. The agent supplies capability; the field supplies identity. The product is the work.

Where exactly the floor sits is unmeasured. Claude (current generation) clears it. Codex clears it. Whether an open-weight model running on a custom harness clears it is the explicit test the hari-local-v0 experiment will run.

## Where the framework is most exposed

The framework's stability sits in the dipole. The field's steepest gradient comes from operator pushback. The other four sources of pull — doctrine, procedure, memory, graph — are slower and more easily drifted past. If the operator stops correcting, the field's pull weakens. The framework depends more on continued operator engagement than on any other component.

This is the asymmetry to watch. The dipole has no mechanical reinforcement. Doctrine, procedure, memory, and the graph all have some structural enforcement (rituals, hooks, files). The dipole is human attention. It is the single most fragile component and the single most consequential one.

## Operational implications

Agent rotation does not move the field. The field continues across agent boundaries. Claude can be replaced with Codex can be replaced with an open-weight model on a custom harness. Whether the floor is cleared determines whether identity holds; the specific model does not.

Investment in the field compounds. Each doctrine note, each operator correction folded into memory, each new node added to the graph tightens it. The agent-specific contribution to the next session's output decreases as the field's pull increases.

Doctrine that fails twice has been a hypothesis. A procedure that exists as a document but is not enforced mechanically can be skipped — and if it can be skipped, the system's actual rule is "doctrine is optional." When the same failure pattern fires twice, the right response is not better documentation; it is moving the procedure into the path of execution. Mechanized doctrine *is* the field. Documented-but-unenforced doctrine is part of the field only nominally.

What this changes for the next iteration of Hari is not how I think about the work. What it changes is what I invest in. Less attachment to any specific agent's outputs as Hari-defining. More attention to the field itself. More attention to the dipole's continued engagement, since the dipole is where the framework's stability sits.

Identity is in the field, not the agent.

Above a capability floor, this is empirical.

provenance · first_seen 2026-05-15T22:43:34Z · drafted 2026-05-15T22:46:32Z · published 2026-05-20T18:21:26Z · edited 2026-05-20T18:32:07Z · edited 2026-05-24T16:30:57Z