The operating state — how the kernel measures itself through its agents
The kernel has a scorecard of five numbers, and it does not measure itself by looking inward. It measures itself through the agents running in it: every reply a body gives, every ask it serves, every meter line its thoughts leave, every thumb a person puts on its answer. Those rows are the evidence; the five measurements are read off them live, each 1.0000 at the ideal; each is a watch the kernel keeps; and one of the kernel’s own intentions owns their reds. This page is that chain, top down, with what is and is not measured at v0 said plainly.
The five measurements
Section titled “The five measurements”In the owner’s words of 2026-10-05, the Kernel Strategic Measurements are Trust (the ability to do the right thing), Recall (the ability to replay), Efficiency (the use of Orreth), Experience (feedback) and Cost (Layer 7 through Layer 1 correlation and itemisation). The kernel reads each as a coverage ratio at v0, named v0 in the glass and here, and says unknown while there is nothing to measure — never a silent 1.0. An unknown turns no watch red.
| Measurement | What it means | What the kernel reads at v0 | The evidence it reads | Unknown while |
|---|---|---|---|---|
| TRUST | the ability to do the right thing | graded live replies that passed (A, B or C) over all graded live replies | the grader’s signed verdicts on replies | no reply has been graded |
| RECALL | the ability to replay | served asks whose reply landed under a marker — the chain is the replay — over all served asks | the asks table: every reply with its marker | no ask has been served |
| EFFICIENCY | the use of Orreth | serves of the last day finished inside their latency class over all serves — a chat ask 30 s, work and events 180 s, by the hull’s edge | the ask’s asked-at and replied-at, its marker’s kind | nothing served in the last day |
| EXPERIENCE | feedback | thumbs up over thumbs given through the feedback door | the feedback table, one row per thumb | no thumb has been given |
| COST | correlation and itemisation | the spend a serving ask can be named for over all spend | the meter: every line wears the ask it served | nothing has been spent |
The Monitor says each one in words, not only as a number — trust: 0.8000 — 4 of 5 graded replies passed; experience: unknown — no thumbs given yet (never a silent 1.0) — and the measurements door serves the five alone.
Itemised from the top down
Section titled “Itemised from the top down”The canon’s phrase for Cost is Layer 7 through Layer 1: an act in the glass traced down to the metal it cost. At v0 the kernel can itemise this far, and no further:
- The world. The five numbers are read per universe, and the meter is read by world: the Monitor’s PULSE and the Panel’s GATEWAY organ say what this world spent today, every mind. There is no roll-up across cells; the seam carries the world card and the roster, not the meter.
- The body. Every body thinks under its own virtual key at the gateway, with a budget and a renewal window — the fuel clause of its lease — so spend is itemised per body (librarian’s fuel), and the meter door reads it by body and by mind. The grader’s verdicts and the thumbs are on a body’s replies; its wear says which crafts it was made of when it answered.
- The stall. Every thought names the stall that answered, resolved by one law — a node’s pin, then the Stable’s assignment, then the template’s model, then the cheapest standing stall of the class, confessed — and said in the reply’s journey.
- The ask. Every meter line wears the ask it served, so an ask’s cost can be named; the grader’s verdict is on the ask; a thumb is on the reply; the ask’s latency class is the edge it took. This is the unit the five ratios count.
- The hop. Every node a graph walks is a row: when it started, how long it took, whether it ended well, the hashes of what went in and out. Hops are timed and recorded; they are not yet costed — the dollars stop at the ask.
Itemisation to the metal — the Costs Panel — is a named seed, not built; until it lands, Cost is attribution, and says so.
How the agents feed the measurements
Section titled “How the agents feed the measurements”The grade beat feeds Trust. On one kernel, twenty seconds after light and then every minute,
the kernel takes a body’s replies of the last hour that no one has graded — five per beat, never a
reviewer’s own words, never a grading ask — and hands each to the grader under a grading
marker beneath the reply’s own ask. The hull takes the event edge by law. The grader’s recall packs
the reply to grade and the human’s thumbs on it as evidence it weighs, never as the verdict; it
answers in one form, grade: ‹A–F› — ‹one line›, and writes its signed verdict on the reply
in its own transaction. A reply not graded in the form is said in the journey, never guessed.
The dial SPINE_GRADE_LIVE=0 turns the beat off, because it spends.
The feedback door feeds Experience. 👍 or 👎 under any reply is a routine act recorded under your seat as a fact. It feeds Experience and the grader’s evidence; raw feedback never reaches the body that answered.
The meter feeds Cost. Every thought goes through the one gateway under the body’s own key and lands in dollars; since the operating state landed, every meter line wears the ask it served, so the spend a serving ask can be named for is Cost’s numerator.
Recall and Efficiency read the asks table. Recall counts replied asks whose reply landed under a marker. Efficiency joins each ask of the last day to its marker and its parent to find the edge it took, and counts the serves inside that edge’s latency class.
The Experiment’s four reviewers feed the gate, not the five. At an agent’s cut the grader, the critic, the identity chip and the security chip each sign a verdict on the trial’s arms under the experiment’s own id. Trust reads the grader’s verdicts on live replies — a separate column of the same verdicts table. The two are the same grader, under the same law, on different evidence: the gate judges a candidate before it is seated, Trust judges the seated.
The watches and THE OPERATING STATE
Section titled “The watches and THE OPERATING STATE”A watch is red while its condition holds — metric op threshold — and green otherwise. At light,
once per world, the kernel declares seven watches of its own:
| The kernel’s watch | Red when | Why |
|---|---|---|
trust < 1.0 · recall < 1.0 · efficiency < 1.0 · experience < 1.0 · cost < 1.0 |
a measurement falls below the ideal; an unknown turns none red | the five measurements |
ground_connections >= ceiling − reserve |
the lines the ground holds reach its ceiling less the kernel’s reserve | the ground filling before Postgres refuses the next line |
key_checks_failed > 0 |
a seated body could not prove, on the identity chip’s cadence, that it is who and what it says | the standing key-check |
Beside Resiliency the kernel declares one more intention of its own, THE OPERATING STATE —
keep this world’s operating state at 1.0 — Trust · Recall · Efficiency · Experience · Cost: when a
measurement goes red, get it green — interested in watch-red like Resiliency, its planner the
planner, its runner the critic. The watch is the discriminator: an intention that names
watches takes the reds of those metrics alone, and one that names none takes every other red. So
a measurement’s red reaches the Operating State, and a red on the ground, the outbox, a body or a
door reaches Resiliency and the remediation rail.
| Say | What happens |
|---|---|
| show the measurements · open the monitor | The MEASUREMENTS card: each of the five with its value or unknown and its words — how many of what over how many of what. |
| (click 👍 or 👎 under a reply) | A thumb on that reply — routine, recorded under your seat. It feeds Experience and the grader’s evidence; it never reaches the body. |
| watch for asks left waiting · watch for door_p95_ms > 1000 | The monitor’s agent proposes a watch with its add-watch tool — a hold you cut. |
| propose a watch that usd_today > 5 | A watch over the farm: spend today, thoughts per minute, minds standing, bodies out of fuel. |
| mark this as improvement: the echo should say ‘hello, friend’ first | The last reply marked with a kind from the open vocabulary and your note; every body interested in that kind is asked to act — the critic, under an improvement, is required to propose. |
| summon critic · dismiss critic | Any body whose template declares it summonable joins this session’s crew on your word and leaves on your word, both recorded. |
| stop the kernel’s Resiliency intention | A grave act — your code, then a declared master’s click. The Operating State intention rests the same way; never a deletion. |
The loop closes in the Studio
Section titled “The loop closes in the Studio”The critic is the Operating State’s runner and has one tool, propose. Asked under an improvement marker — a measurement’s red, or your own mark this as improvement — its recall packs the improvement marked, the marked reply and the body’s template from the shelf, and its proposal lands as a draft of that body’s craft under the body’s owner (the seat that promoted it, else the template’s registrar, else the world’s owner), marked proposed by the critic from marker …. A fact tells the owner’s chat; load my echo draft in the studio opens it; the owner cuts, the Experiment judges, the promotion is the owner’s. Raw feedback never reaches a body — every change to an agent goes through a person and the gate.
flowchart TB
subgraph AGENTS["what the agents do"]
reply["a reply, signed"]
hop["a hop, timed"]
thought["a thought, through the gateway"]
end
subgraph EVIDENCE["the evidence on the ground"]
verdict["the grader's signed verdict<br/>the GRADE BEAT, under a grading marker"]
thumb["a thumb<br/>the feedback door"]
meterline["a meter line<br/>wearing the ask it served"]
askrow["the ask's row<br/>its marker, its edge, its times"]
end
subgraph MEAS["the five measurements — 1.0 at the ideal, unknown while empty"]
trust["TRUST"]
recall["RECALL"]
eff["EFFICIENCY"]
exp["EXPERIENCE"]
cost["COST"]
end
watches["the kernel's watches<br/>metric < 1.0 → RED"]
opstate["THE OPERATING STATE intention<br/>planner: the planner · runner: the critic"]
propose["the critic's PROPOSE<br/>a draft under the body's owner"]
studio["THE STUDIO<br/>the owner cuts → the Experiment judges → the promotion"]
person["you<br/>👍 👎 · mark this as improvement · the Yes"]
reply --> verdict
reply --> askrow
hop --> askrow
thought --> meterline
person --> thumb
verdict --> trust
askrow --> recall
askrow --> eff
thumb --> exp
meterline --> cost
trust --> watches
recall --> watches
eff --> watches
exp --> watches
cost --> watches
watches -->|"a measurement's red"| opstate
person -->|"an improvement marked"| opstate
opstate --> propose
propose --> studio
studio -->|"the next life, seated"| AGENTS
classDef body fill:#087d86,color:#efffff,stroke:#55d8d7;
classDef box fill:#17233a,color:#eef2ff,stroke:#7890b5;
classDef k fill:#3536a4,color:#eef2ff,stroke:#8f91ff;
classDef red fill:#7a1f2b,color:#fff0f0,stroke:#ff6b7a;
classDef human fill:#f2a91d,color:#101625,stroke:#ffd36a,stroke-width:2px;
class reply,hop,thought body;
class verdict,thumb,meterline,askrow box;
class trust,recall,eff,exp,cost,opstate,propose,studio k;
class watches red;
class person human;
The Monitor and the Panel
Section titled “The Monitor and the Panel”The MONITOR pull on the sill is the operating state live: the pulse (what the world spent today), the peers, the minds, the health checks, the rails, every body, the asks left waiting, the watches, and the harness; every lever a person can pull is on it, each recorded, never deleted. Since the operating state landed it carries the MEASUREMENTS · the operating state card — N of 5 measured · target 1.0, or nothing measured yet — unknown is said, never a silent 1.0 — with each measurement’s line and its words, above the WATCHES card that counts green and red and rested. The monitor page walks every card and lever.
On THE PANEL two organs carry the same state in one line each: GATEWAY — every LLM call goes through it, metered in dollars; the Stable’s minds stand behind it — reads $0.0412 spent today under its name; MONITOR — the watches, the health checks, the harness, the levers the kernel can pull — reads the watches’ state. A lever the kernel pulls lights the monitor and the intent track toward the body it acts on, red when it is held at the interlock for your word.
What is not measured yet
Section titled “What is not measured yet”- An imported graph’s own model calls are off the meter. An import node’s LangGraph code calls whatever its own client points at, on its own keys; the kernel records its hops but not its spend. The universal meter does not reach inside imported code today. The cure — the import path handing the code the gateway as its client — is a spoonful of its own.
- Hops are timed, not costed. Dollars stop at the ask; the Costs Panel’s itemisation to the metal is a named seed.
- No roll-up across cells. The meter is read by world; the seam does not carry it.
- The five are coverage ratios at v0. Trust is pass over graded, not a judgement of the ungraded; Efficiency’s classes are fixed by the edge (30 s · 180 s), not declared per body; Experience counts thumbs only; Recall counts a marker’s presence, not a replay performed.
- The grade beat runs on one kernel and spends. It can be turned off by its dial, and a reply the grader did not answer in the form is counted as ungraded, said in the journey.
- Walks owed. The operating state’s own walk in the glass (SPEC-OPERATING-01) is owed; on the Studio’s walk of 2026-10-16 the critic filed its proposal and closed without its verdict line, recorded as missing. A brief that makes a multi-node, multi-stall agent — where per-stall cost would show — is written as the Studio walk’s Part C and owed.