open-standards/jcp/SPEC.md

# JCP v0.3 — The Joule Context Protocol

**Status:** v0.3.0 draft. **Licence:** specification text CC-BY-4.0; reference
code Apache-2.0.

A unified, energy-native, capability-secured protocol for agentic systems. JCP
subsumes in one spec what today's stack splits across MCP (tool/context
connection), A2A/ACP (agent-to-agent), an identity layer, an authorization
layer, and — missing everywhere — an energy layer. Its thesis is one object:

> **A capability and an energy budget are the same thing — a signed, bounded,
> debited grant. A capability you cannot pay for is denied; an expenditure you
> are not capable of is denied. One check, made deterministically outside the
> model, sealed in a receipt.**

---

## 1. Scope and doctrine

1. **Enforcement is outside the model.** Authorization and metering are decided
   by the runtime, never by the LLM's instruction-following. The published
   *inverse-scaling-law* result — more capable models are more exploitable
   because they follow injected instructions better — makes model-side control
   unsound by construction.
2. **Authority is a capability, not ambient.** No actor holds standing
   privilege; every action requires an unforgeable [Grant](#4-the-grant-required).
   There are no long-lived ambient secrets in env vars.
3. **Energy is first-class.** Every grant carries a joule budget that is debited
   per action and **enforced before the side effect** — not merely observed.
4. **Data is never control.** Tool and agent *outputs* are untrusted data and can
   never be executed as instructions or used to exercise a capability. This
   closes the asymmetric-trust gap that tool poisoning exploits.
5. **Every decision is non-repudiable.** Allow and deny alike seal as a signed
   receipt. Forging the audit means forging a signature.
6. **One substrate.** The same identity, grant, frame, and receipt span the tool
   layer and the agent layer; there is no second protocol to secure.

## 2. Actors and identity (REQUIRED)

Every actor — model, agent, tool, or server — is an Ed25519 key, expressed as a
W3C `did:key` (multibase base58btc of the `0xed01` multicodec ‖ the public key).
There are no anonymous actors. An actor proves itself by signing.

### 2.1 Identity continuity: pre-committed rotation (OPTIONAL)

For long-lived actors, an identity is its **inception DID** (seq 0), forever; the
key that controls it rotates through a verified event log:

```
KeyEvent {
  identity:  did:key            ; the seq-0 (inception) DID, stable for life
  seq:       uint               ; 0 = inception, N = the Nth rotation
  current:   did:key            ; the controlling key from this event on
  next_hash: content-id         ; blake3 of the NEXT key's 32 public bytes
  parent:    content-id?        ; the previous event's id; absent at inception
  nonce:     bytes(16)
}
```

Inception is signed by the identity key (`current` MUST equal `identity`).
Rotation event N is signed by **its new `current` key**, which MUST hash to
event N−1's `next_hash` — possession of the pre-committed key is the proof of
control. The consequences are exactly the two that matter: a thief who steals
the *current* key still cannot rotate the identity away (they lack the
pre-committed next key), and an owner who *loses* the current key still can
(they hold it). A verifier MUST reject sequence gaps, parent-link mismatches,
hash mismatches, and bad signatures.

Rotation machinery is the long-lived-identity case. The RECOMMENDED norm is
cheaper: **short-lived grants re-issued on a heartbeat**, with revocation (§4.3)
covering the emergency. Rotation pressure scales with credential lifetime.

### 2.2 Threshold actors: the two-agent rule (OPTIONAL)

A **threshold actor** is a `did:key` whose private key never exists in one
place: it is split k-of-n (FROST, RFC 9591, Ed25519 ciphersuite), and a
signature exists only when at least k shareholders cooperate. Because the group
key is an ordinary Ed25519 verifying key, **nothing else in JCP changes** — a
threshold actor issues, holds, delegates, revokes, and signs wire requests, and
no verifier can tell or needs to.

Use it where one gated model is still too few: make the subject of an
irreversible capability a k-of-n team, and no single agent — however capable,
however poisoned — can exercise it alone. The refusal is arithmetic, not
policy.

Shares MAY be minted by a trusted dealer (one party creates a team identity and
hands out shares) or by **distributed key generation**, in which the group's
private key never exists anywhere — not even momentarily at a dealer. The group
key is the same ordinary Ed25519 key either way.

## 3. Canonical serialization and content addressing (REQUIRED)

Content addresses and the signed receipt core use **canonical det-CBOR**
(RFC 8949 §4.2.1): length-first then bytewise key order, shortest integer form,
**no floats, bools, or nulls** in a signed core. A content id is
`blake3(canonical_cbor(value))`, lowercase-hex on the wire — the same encoding
the JCR-1 receipt signs over (§8), so an id and a receipt agree on bytes.

## 4. The Grant (REQUIRED)

The **metered capability** is JCP's central object:

```
Grant {
  issuer:     did:key            ; who granted this authority
  subject:    did:key            ; who holds it
  capability: { resource: content-id, actions: [string] }
  budget_uj:  uint               ; the joule ceiling this grant may ever spend
  budget_funds: { currency: string, minor: uint }?  ; OPTIONAL money ceiling (§4.8)
  caveats:    [caveat]            ; OPTIONAL embedded conditions (§4.9); omitted when empty
  expiry:     uint               ; unix seconds; 0 = no expiry
  parent:     content-id?        ; the grant this was attenuated from
  delegate_depth: uint?          ; further hops permitted; absent = unlimited, 0 = leaf-only (§4.4)
  nonce:      bytes(16)
}
```

`budget_funds` is absent in the baseline grant. Per §3, an absent OPTIONAL field
is omitted from the canonical encoding entirely, so a grant with no money ceiling
is byte-identical to one minted before §4.8 existed — its content id, its
signature, and every published conformance vector are unchanged.

A grant is identified by `blake3(canonical_cbor(grant))` and signed by the
`issuer`'s key (Ed25519 over the canonical bytes). The `resource` is addressed
**by content, not by URL** — audience binding an attacker cannot spoof by name
(a content-addressed strengthening of RFC 8707).

### 4.1 Attenuation (REQUIRED)

Delegation is object-capability attenuation. The holder of a grant (its
`subject`) MAY mint a child, but the child MUST be **strictly weaker**:

- `child.capability` ⊆ `parent.capability` (same resource; actions a subset),
- `child.budget_uj` ≤ `parent.budget_uj`,
- the child's expiry MUST NOT outlive the parent's (`0` = never expires),
- `child.parent` = the parent's content id (forming the chain).

A conforming implementation MUST refuse any amplification and MUST refuse
delegation by anyone other than the parent's subject. A chain therefore can
never convey more authority — or more energy — than its root.

### 4.2 Chain verification (REQUIRED for multi-hop)

When a frame is decided against a delegated grant, the enforcement point MUST be
presented the **full chain, root first, leaf last**, and MUST verify it hop by
hop: the root has no parent; every later grant's `parent` equals the id of the
grant before it; each hop is issued by the previous grant's `subject`; each hop
narrows per §4.1, carries every caveat of the grant before it (§4.9) and no
wider money ceiling (§4.8); and every signature in the chain verifies. A
structurally invalid chain is denied `chain-broken`; an amplifying one
`attenuation-violation`. The rules `attenuate` holds when it mints are held
again here, because a chain arrives signed by whoever minted it.

A delegated grant (one with a `parent`) MUST NOT be decided on its own. Alone,
nothing shows that it narrows what its issuer held, so its budget, caveats and
expiry would be taken at their word, as a root's are. A single-grant decision
over a delegated grant is denied `chain-broken`.

**Caveats bind on every link.** A chained decision MUST evaluate the caveats of
every grant in the chain, exactly as a single-grant decision evaluates its own
(§4.9); a chain is never a way past a condition its root imposed. In the
two-phase metered path a chain is authorized without debit, cost-dependent
caveats waiting, and then charged with the measured cost, which is debited from
every link. A priced chained decision applies both ceilings of §4.8 to **every**
link: the joule ceilings first, then the money ceiling of each link that has
one. An allow debits both dimensions from every link, so siblings share their
parent's money as they share its joules; a deny debits neither.

**Energy conservation across siblings.** The cost of an allowed action MUST fit
the remaining budget of **every** grant in the chain and MUST be debited from
all of them. Two children attenuated from one root therefore share the root's
budget — delegation can never mint joules.

### 4.3 Revocation (REQUIRED)

```
Revocation {
  grant_id:   content-id         ; the grant being revoked
  revoked_by: did:key            ; MUST equal that grant's issuer
  nonce:      bytes(16)
}
```

A revocation is signed by `revoked_by` over its canonical bytes. An enforcement
point MUST verify authority before honoring one — the ids match, the revoker is
the grant's issuer, the signature verifies — and MUST then deny (`revoked`) any
decision against the revoked grant **or any chain containing it**: a delegation
subtree dies with any of its ancestors. Both the direct and the chained decision
path MUST enforce this.

**Propagation.** The rule above binds an enforcement point *once it holds the
revocation*, and says nothing about how it comes to hold one. §4.3 requires an enforcement point to deny a revoked grant **once it holds the
revocation**, and says nothing about how it comes to hold one. With a budget
sharded across N enforcement points (§4.5), a revocation that has reached three
of five leaves the grant live at two, and — the part that matters for an audit —
an allow issued by a point that *had not yet heard* is indistinguishable, after
the fact, from an allow by a point that heard and ignored it. The first is a
propagation delay. The second is a violation of this specification.

An implementation that supports revocation **MUST** define how revocations reach
its enforcement points, and **SHOULD** let each point state how current its
revocation knowledge is, as a marker monotonic in whatever its revocation feed
counts — a sequence number, a log index, a timestamp. JCP neither defines nor
interprets the marker; it requires only that a larger value mean "has seen at
least as much".

Given a revocation issued at marker `M`, an allow from a point at
`known_through < M` is explicable and one at `known_through >= M` is not. Where a
point states no marker, an auditor **MUST** treat its allows as *unverifiable*
for revocation rather than as compliant — "did not say" is not "was current".

### 4.4 Delegation control (REQUIRED)

A grant MAY carry `delegate_depth`: how many further attenuation hops it
permits. Absent = unlimited; `0` = **leaf-only** — the holder may act but never
re-delegate. Like every attenuation axis, depth only narrows: a child minted
from a depth-d parent carries at most d−1, and chain verification (§4.2) MUST
refuse a hop under a leaf-only parent (`delegation-forbidden` semantics map to
`chain-broken` on the wire) and any child whose allowance exceeds its parent's
bound (an amplification). In group topologies this is the primitive that makes
a worker a worker: an orchestrator delegates depth-0 grants and the team
cannot deepen itself.

### 4.5 Sharded budgets: distributed enforcement without consensus (REQUIRED baseline)

When one grant must be enforceable at N enforcement points, the holder MUST
shard it: mint N sibling children whose **budgets sum to no more than the
parent's**, one per enforcement point. Each point then enforces only its own
shard, locally — no coordination, no consensus — and the root stays conserved
by mint-time arithmetic. Rebalancing is revocation + re-issuance.

A conforming sharding operation MUST refuse slices that sum past the parent.
Fragmentation (an unused slice in one shard unavailable to another) is the
deliberate price of coordination-freedom — the same trade cash makes.

**Receipt-settled budget channels (OPTIONAL profile).** Where shards are too
rigid, N enforcement points MAY spend **optimistically** against one grant and
settle by exchanging signed statements of cumulative local spend:

```
Statement {
  grant_id: content-id
  point:    did:key              ; the enforcement point making the claim
  spent_uj: uint                 ; cumulative µJ allowed at this point since open
  seq:      uint                 ; strictly increasing per point
}
```

A statement is signed by `point` over its canonical bytes. A receiving point
MUST verify the signature, refuse stale (`seq` not increasing) and regressing
(`spent_uj` decreasing) statements, and thereafter count the peer's settled
spend against the shared budget. Between settlements the points can
collectively overspend by at most what the unsettled window allows — a bounded,
tunable exposure, and after settlement the overdraft is **provable from the
signed statements** rather than suspected. Where even bounded overdraft is
unacceptable, use shards; that is why the baseline is the baseline.

### 4.6 Standing grants: bounded authority for event-triggered work (OPTIONAL)

An autonomous agent often starts work on its own — on a webhook, a schedule, a
ticket — with no human present at the firing instant. A long-lived credential
that any trigger may spend is unbounded authority in an automation costume. A
**standing grant** keeps the object-capability discipline instead: an ordinary
grant (§4) the issuer holds, plus a signed **mint policy** that says how a
trigger may turn it into work. The policy is bound to one grant by id, signed
by that grant's issuer, and MUST NOT amplify it. It fixes three bounds at once:
the **event class** permitted to fire it (content-addressed, so a webhook for
one source cannot spend a grant meant for another), the **per-firing budget**
(a slice no larger than the grant's budget), and the **maximum number of
firings**.

Each firing mints one fresh child through the ordinary attenuation path (§4.1):
the child is strictly weaker, its `parent` is the standing grant, and it
chain-verifies (§4.2) like any delegation — so its receipt traces back to the
standing grant. A small **mint ledger** (firing count and cumulative spend)
carried by the enforcement point holds the rate limit and the parent's budget
conservation across firings by arithmetic alone, no coordination — the same
local-enforcement property as sharded budgets (§4.5). A firing is refused, and
the ledger left unchanged, when the event is of the wrong class, the firing
count is exhausted, or the next firing would push cumulative spend past the
grant. Conformance class **C20**.

### 4.7 Checkpoints and amendments: steering a long run (OPTIONAL)

A multi-hour run is not a black box that returns only at the end. At defined
boundaries the running agent commits to its state — a **checkpoint** — and a
*distinct* principal may steer it before it continues. The two roles are
structurally separate (the separation-of-duties rule): the checkpoint's
**author** is the agent (the grant's `subject`), which commits to its own
state and cannot steer itself; the **amender** is the grant's `issuer` — the
steward who delegated the authority and is therefore already a different
principal. Steering is an issuer power, like revocation (§4.3): an
**amendment** is a soft revoke that re-bounds the run rather than killing it.

A checkpoint cites the active grant and a content-addressed state commitment,
signed by the agent. It MAY also state what the run has **spent** so far: a
cumulative microjoule figure with the provenance grade of §5. An amendment
cites a checkpoint, is signed by the grant's issuer, MUST differ from the
agent, and MUST only **tighten**; it MAY also hand the agent a new
instruction-set commitment.

What "tighten" is measured against is normative, and it is the **remainder**,
not the issued budget:

- Where the checkpoint states its spend, an amendment MUST NOT set a budget
  above `grant.budget_uj − spent.uj`, saturating at zero. Both the amending
  side and the verifying side MUST enforce this: an enforcement point handed a
  signed amendment never calls the amending routine, so a check that lives only
  there is not a check.
- Where the checkpoint states no spend, the ceiling is the issued
  `grant.budget_uj`. This is the weaker rule and is retained so that a
  checkpoint written before the field existed remains valid; implementations
  MUST NOT read an absent spend as zero spend, and SHOULD report the remainder
  as unknown rather than as the full budget.
- A declared spend is cumulative, so across one run it MUST NOT decrease, and a
  checkpoint that once stated a spend MUST NOT later omit it. Otherwise an agent
  recovers its full issued ceiling by going quiet.

The figure is the agent's own signed claim, not an independent measurement; its
`provenance` grade travels with it so the amender can see which kind of claim it
is, and a deployment that needs more binds receipts (§5) to the checkpoint.

The run resumes under a **continuation grant** the issuer re-issues to the same
agent with the amended budget, linked back to the steered grant. This
continuation is an issuer re-binding, not a holder delegation — an auditor
verifies it alongside the signed checkpoint+amendment pair, which together
record who intervened, when, and how. Conformance class **C21**.

### 4.8 Dual ceilings: a money budget beside the joule budget (OPTIONAL)

A grant MAY carry a second, independent ceiling denominated in **money**
(`budget_funds`), enforced alongside the joule ceiling of §4. This exists because
the payments ecosystem authorizes agent spend in a currency (an AP2 mandate names
dollars), while JCP meters work in joules — and the two MUST NOT be reconciled by
an invented exchange rate. There is no canonical joules-per-dollar; asserting one
would fabricate authority. So JCP enforces **both** ceilings and reconciles
neither.

The two dimensions are governed by two different sources of authority: the
payment mandate bounds the money, the operator bounds the energy. A conforming
runtime MUST, on each priced decision:

- deny with `funds-exceeded` if the action's money cost would carry the grant's
  money spend past `budget_funds.minor`;
- deny with `currency-mismatch` if the cost is denominated in a currency other
  than `budget_funds.currency` — a foreign-currency cost is refused, never
  converted;
- on allow, debit **both** the joule and the money spend; on any deny, debit
  **neither**;
- treat an absent `budget_funds` as "joules alone bound this grant" — the money
  dimension is not present, not a zero ceiling.

Because the ceilings are independent, a decision MAY deny on one while the other
still has room, and the sealed receipt (§8) records *which* bound. A receipt that
reads "denied: joule ceiling, while the money ceiling had room" is a statement no
single-dimension authorization can make — and it is the point of the design.

**Attenuation (§4.1) applies to both dimensions independently.** A child's money
ceiling MUST NOT amplify the parent's: a child bounded in money where the parent
was unbounded is permitted (it only adds a constraint), but dropping a bounded
parent's ceiling, raising it, or denominating it in a different currency (which
cannot be shown to narrow without a rate) MUST be refused. A plain attenuation
that names no money ceiling **inherits** the parent's unchanged — it is never
silently dropped, which would amplify. Conformance class **C22**.

### 4.9 Caveats: the open attenuation axis (OPTIONAL)

The axes of §4.1 narrow by set-subtraction — remove actions, lower a ceiling,
shorten an expiry. They cannot express a condition the issuer did not anticipate.
A grant MAY therefore carry **caveats**: conditions embedded *inside* the signed
token that MUST all hold for a decision to be allowed.

The point is that a caveat travels **with the capability**. Any enforcement point
holding the signed grant checks it offline — no call back to the issuer, no
service-side policy to consult, no shared configuration. That is what
distinguishes a caveat from a policy engine: policy is something the verifier
must be told; a caveat is something the token carries.

A conforming implementation MUST support:

```
ActionIn([string])    ; the action must be in this set — MAY narrow below the capability
PayloadIs(content-id) ; the frame's payload must be exactly this — a ONE-SHOT binding
CallerIs(did:key)     ; the caller must be this principal
NotBefore(uint)       ; unix seconds
NotAfter(uint)        ; unix seconds — a window narrower than `expiry`
MaxCostUj(uint)       ; a ceiling on ONE invocation, distinct from the cumulative budget
ModelIn([content-id]) ; the frame must name its model (§6), and it must be one of these
```

Every predicate ranges only over what an enforcement point holds when it decides:
the frame, the caller, the clock, and the metered cost. There is deliberately **no
content predicate** — the runtime sees a frame's `payload_hash`, never its bytes —
so a condition on content is expressible only as `PayloadIs`, binding an exact
payload rather than inspecting it. An implementation MUST NOT invent a predicate
that requires data the decision point does not have.

Three of these express authority the fixed axes cannot. `PayloadIs` makes a grant
**one-shot**: it authorizes one specific call, so a leaked grant cannot be
replayed against different arguments. `MaxCostUj` bounds a **single invocation**,
where `budget_uj` bounds the cumulative total — so a delegate cannot spend an
entire remaining budget in one call. `ModelIn` says **which model** the grant
authorizes. The capability's `resource` names the endpoint, and one endpoint can
run many models; `ModelIn` names the model content addresses this grant may
run there. It ranges over the frame's `model` field (§6), which the runtime that
dispatches the call fills with the content address of the model it will load,
and it **fails closed**: a frame that names no model does not satisfy it, so a
dispatcher that does not say what it will run cannot exercise the grant. This is
authorization by content address, not a registry (§12): the grant names the
bytes it permits, and nothing has to be looked up. A delegate narrows the set
by adding a smaller `ModelIn` beside the inherited one; both must hold, so the
intersection governs.

**The metered path (REQUIRED).** A caveat whose predicate ranges over the
invocation's cost — `MaxCostUj` — cannot be evaluated in the two-phase metered
flow at authorization time, because the cost is not yet known. Such a caveat MUST
be enforced at the **debit point** instead, and an implementation MUST NOT report
a grant as authorized in a way that implies a cost-dependent caveat has passed.
Every cost-independent caveat MUST be enforced in both phases.

When a measured cost breaches a cost-dependent caveat at the debit point, the
work has already run and the joules are real: the implementation MUST still debit
them — energy that was spent is always accounted — and MUST record the breach in
the verdict, so the sealed receipt is evidence of it. Debiting silently, or
suppressing the debit to make the record look clean, are both non-conforming.

**Attenuation (REQUIRED).** Caveats only accumulate. A child MUST carry every
caveat of its parent; a plain attenuation inherits them unchanged, and a
caveat-adding attenuation appends. Dropping an inherited caveat is an
amplification and MUST be refused.

Caveats compare by **exact identity**, not by an ordering over their parameters:
a child retains a parent's caveat only by carrying that identical caveat. There is
deliberately no notion of "a tighter `MaxCostUj` supersedes a looser one" — such
an ordering would have to be defined per variant, and a verifier that got it
wrong would silently widen authority. A delegator that wants a tighter bound adds
it **alongside** the inherited one; since all caveats must hold, the tighter of
the two governs. This is the conservative reading: it can never widen, and it may
be relaxed later without breaking existing grants, whereas the reverse is not
true. A decision whose grant carries an unsatisfied
caveat is denied `caveat-unsatisfied`, sealed like any other denial.

An empty caveat set is omitted from the canonical encoding, so a grant without
caveats is byte-identical to one minted before they existed — its content id and
signature are unchanged. Conformance class **C24**.

### 4.10 Flow: what a value's provenance may exercise (OPTIONAL)

A grant answers *who* may do *what*, for *how many joules*. It does not answer
where the **argument** came from. An agent holding a legitimate grant, handed a
value that originated in an untrusted channel, exercises real authority on
attacker-chosen input — and every check above passes, because the grant is valid
and the action is granted.

A value therefore carries a **label**, and a grant MAY carry a ceiling the label
must sit below.

```
Label { integrity: Integrity, origins: Origins, readers: Readers }
```

The three components order independently and the label order is their product.
`⊥` is a principal-authored value from no channel, readable by all — the label of
anything in a run that has touched no untrusted source. `⊤` is untrusted, from
anywhere, readable by nobody.

**JCP consumes a label; it does not compute one.** Propagation — the label of a
result being at least the join of the labels of everything that went into it — is
the host's job, because only the host sees the data move. JCP's contribution is
the gate and, below, the rule that an *uncomputed* label is not a clean one. A
host that propagates badly will present a label JCP cannot audit; what JCP
guarantees is that a host which propagates at all gets a decision bound to what
it reported, and that one which does not gets a refusal rather than a pass.

A grant expresses its ceiling as a `FlowBelow` caveat (§4.9) carrying a
`FlowCeiling`: the ceiling label, and optionally a `tainted_budget_uj` — a
separate, smaller joule allowance for calls whose argument is not `⊥`. Where a
chain carries several, they compose to the **strictest**; attenuation never
widens one.

An enforcement point MAY also configure a **taint floor** per resource — the flow
analogue of a trust root (§7). The ceiling is what the *issuer* will permit; the
floor is what the *enforcement point* will accept regardless of who granted what.
Default-open, so an unconfigured point behaves exactly as before.

**The gate.** A decision is denied unless the argument's label sits at or below
the grant's ceiling **and** the resource's taint floor does not refuse its
origins. Three cases are distinguished, and the distinction is the point:

| facts | meaning | result |
|---|---|---|
| absent | provenance was **not computed** | **fails closed** where a ceiling is present; a ceiling-free grant is unaffected |
| `⊥` | computed, and clean | passes any ceiling |
| labelled | computed, and tainted | tested against the ceiling and the floor |

"Not computed" **MUST NOT** be read as "clean". An un-migrated caller that cannot
say where its argument came from has not established that the argument is safe,
and a grant whose issuer took the trouble to set a ceiling MUST NOT be exercised
on an unlabelled value.

**Where the check sits is normative.** It MUST run *after* capability and
signature verification — the ceiling lives inside an authenticated grant, so
judging flow before the signature would seal a verdict over attacker-chosen
policy — and *before* caveats and the budget test, so that **a tainted call costs
zero joules**. A denial on flow is the same invariant as a denial on capability:
nothing is debited and nothing runs.

**What "the argument" covers.** Where an enforcement point forwards a request to
an endpoint, the argument whose label is gated is everything the request carries
to the endpoint except the routing that names the action. That includes fields
the protocol leaves open: extension parameters, side-channel metadata, and
envelope members the enforcement point does not otherwise interpret. Any such
field it leaves unlabelled is a way past the ceiling. The enforcement point MUST
label each of these fields, and MUST seal it with the arguments, or else MUST
NOT forward it. It MAY leave out the protocol's own fixed machinery (a version, a
capability set, a progress token) where folding that machinery would taint every
request. If it does, it MUST say which fields it leaves out. A host that admits
every value it labels SHOULD label a request's parts so that combining them adds
no text of its own. Otherwise the framing becomes shared text that carries one
tainted request's label into every request after it.

A grant carrying no `FlowBelow` caveat is byte-identical to one minted before
flow existed, and an enforcement point with no taint floor configured decides
exactly as it did before. Conformance class **C25**.
### 4.11 Durable spend: a restart is not a fresh budget (REQUIRED where a run can restart)

A grant's budget is enforced against the spend an enforcement point has
accumulated, and that accumulation is state. Where the enforcement point can
restart — any long-running orchestrator, worker pool or gateway — an
implementation that holds the spend only in memory gives every grant its full
budget back on restart. A crash loop then spends a grant's budget repeatedly,
and the ceiling that was meant to bound the work bounds one process lifetime
instead.

An implementation whose enforcement point can restart therefore **MUST** record
every allowed decision durably, and on restart **MUST** re-establish the spend
from that record before deciding anything new. Re-establishing it by **replaying
the recorded decisions through the gate** is RECOMMENDED over restoring a stored
total: a replay cannot produce a spend the gate would not have produced, and a
stored total is one more figure that can be wrong or tampered with.

**One writer per record (MUST).** The replay above re-establishes spend *from*
the record, so two enforcement points sharing one record each re-establish the
**same** spend, each decide against it, and each append. The budget is spent
twice and every entry still verifies — the requirement above is defeated one
layer beneath itself, and nothing in the record shows it. An implementation
**MUST** therefore ensure a single writer per record: it **MUST** take
exclusive access when it opens the record, and **MUST** fail to start rather
than proceed where another writer already holds it.

An exclusive claim an operating system releases when the holding process dies
is preferable to one an implementation must clean up, because a crash then
leaves nothing to recover. Where a platform cannot provide exclusivity at all,
the implementation **MUST** fail to open the record rather than proceed —
deciding against a record another endpoint may also be writing is the overspend
this requirement exists to prevent, and a platform limit is not a licence for
it. An implementation that instead proceeds **MUST** make that fact checkable
at run time rather than leave the type to imply a property it does not hold.

**The decision is recorded before the action runs.** Recording it afterwards
loses the debit for any action interrupted between the two — the grant was
debited when the decision was made, so a record written after the fact reports a
budget that has more left in it than it does.

That ordering leaves one state the record cannot resolve. A recorded decision
with no recorded outcome means the enforcement point stopped while the action
was in flight, and **whether the action took effect is not knowable from the
record** — only the party that was running it knows. An implementation **MUST
NOT** silently retry such an action, and **MUST NOT** report it as completed or
as failed. Its energy **MUST** stay debited, because the decision that debited it
was made and sealed. The honest resolution is to surface it as unresolved,
naming the party that held it, and let an operator or that party settle it:
retrying risks performing paid work twice, discarding risks losing it, and
choosing between those without asking is a decision the record does not support.

## 5. Energy (REQUIRED)

Every metered cost is integer microjoules with a provenance tag: `HwShunt`
(hardware reading; a measured figure REQUIRED), `ModelBased` (a model
calibrated on the host or against a reading, or a real reading apportioned by
a model; measured MAY be absent), or `Estimator` (a figure nothing calibrated:
a coarse constant, or a model run at a default or assumed parameter, such as
time × a default power; measured MUST be absent). The runtime debits `cost.uj()` from a grant's remaining budget on every
allow; the coupling MUST be enforced. The energy axis is what no existing
agentic protocol standardizes — JCP makes it a precondition, not a report.

`HwShunt` covers a vendor energy counter as well as a dedicated shunt. A vendor
counter's count is a real reading, not an estimate; what it lacks is an
independent root — how the vendor derives it is undisclosed, and the
platform's own software relays it — which is a gap in provenance, not in
accuracy. A reading earns the grade only while its counter is live:

**Liveness (REQUIRED).** A reading from a vendor energy counter (RAPL, NVML,
Apple IOReport, ...) MAY be graded as a hardware reading only while the counter
is demonstrably live for the domain the figure claims: the implementation MUST
have observed the counter covering that domain advance under load (for Apple
IOReport: the CPU energy rail itself advancing in at least 5 sample intervals,
in at least 4 of every 5, over at least 300 ms while the probing thread burns
CPU). Where the figure sums several counters of the domain — a RAPL package
counter per package, an IOReport CPU energy rail per die — every one of them
MUST be observed live: a frozen one beside live ones would be left out of a
figure that claims the whole domain. A counter that is present but not live —
frozen, or republished in lumps — MUST NOT produce a hardware-graded figure; the implementation falls back to
its estimator or reports the work unmetered, and SHOULD record why. A figure
MUST NOT claim coverage of a domain its counter does not read; where the format
has a coverage field, such a figure is Partial.

**Plausibility (REQUIRED where a bound is readable).** Liveness is necessary and
not sufficient: a counter can advance continuously and still not be counting
what it claims. An NVML total-energy counter on an RTX 4050 Laptop GPU (driver
595.84) advances in every sample interval while running at four to thirty times
the board's own power draw for as long as it is polled — perfectly live, and no
count of that board's energy. A figure from it would debit energy the device
never spent.

So where the platform exposes a bound that is read **independently of the
counter** — the board's maximum power limit, the board's instantaneous power
draw — a hardware-graded figure's implied rate MUST be checked against it, and a
counter whose rate exceeds the bound MUST NOT produce a hardware-graded figure.
An implementation SHOULD record the rate and the bound it broke, not merely that
it broke one: a counter a tenth over a bound and one twenty times over are the
same refusal and not the same finding, and re-grading a past refusal against a
revised bound is only possible if the rate survived the check.

The check is **two-sided**. Where the bound is the board's own draw, a counter
whose rate falls far below that draw MUST NOT produce a hardware-graded figure
either. This is the more dangerous of the two directions. Under-reporting is the
unsafe direction at a ceiling (§5), because it admits work the budget cannot
cover; a counter running fast fails safe, because it only spends a budget
sooner. A counter republished in lumps and sampled across its republish period
reads low rather than frozen, so liveness alone does not see it. The floor MUST
allow for the draw reading's own imprecision. It MUST NOT fire where the board
is near idle, because a correctly reporting counter accrues almost nothing there.
The reference meters refute a rate under a quarter of the board's mean draw,
after the draw's ±5 W accuracy is taken off it. Either bound refutes and never
verifies: a counter inside both is not refuted, and that is not a finding that
it is correct.

A bound reading that is itself impossible is not a reading: the same RTX 4050
reports about 590 W, past its own 65 W limit, for a second after it wakes. Such a
reading MUST be discarded — and only that reading. One broken sample is not a
broken bracket, and discarding the whole judgement on it reports an indeterminate
where the evidence supports a finding: on the reference hardware a single spike
among otherwise sane 14.8 W samples would otherwise excuse a counter running
persistently at four to thirty times the board's draw. Discarding them
individually also keeps the comparison honest, since a spike left in a mean makes
the bound it derives *more* permissive.

Only where **no usable reading remains** is the counter **unjudgeable** rather
than implausible: there is then no yardstick, so there is no finding about the
counter. It MUST NOT produce a hardware-graded figure either, because unjudgeable
is not admissible and declining is the safe direction, but an implementation MUST
NOT report it as a counter that overran. The two call for different operator
responses — replacing a counter, or waiting for a board to settle — and
conflating them loses that. A bound that is a static limit rather than a reading
(the board's maximum power limit) is never discarded, so a broken draw sample
cannot excuse a counter that overran that limit outright.

Where no such bound is readable, this requirement does not apply; liveness
stands alone.

The record of why is the implementation's to choose; the reference meter
exposes each counter's liveness status, and which rung it selected with the
status of every rung above it that it declined. Coverage
is defined in §5.2.

### 5.1 The metering ladder (REQUIRED behavior, RECOMMENDED backends)

The rule: **measured when the platform exposes it, estimated when it doesn't —
and the provenance always says which.** There is always an estimate: an
environment with no energy interface still yields a usable figure, it just
carries `ModelBased` or `Estimator` provenance instead of `HwShunt`. A figure
MUST NOT claim better provenance than how it was obtained — including when a
hardware backend fails mid-reading and the figure degrades to a model or an
estimate (rungs 5 and 6 below).

Recommended ladder, best first:

1. **RAPL** (Linux `powercap`) — package energy counters, wraparound-corrected
   → `HwShunt`, selected only while every package counter is live under the
   probe's CPU burn, by the rule above as IOReport's CPU rails are. A span over
   which no package counter advanced is metered by a lower rung. The package
   counters are the top-level zones whose `name` is `package-<n>`, every
   socket's, each corrected at its own `max_energy_range_uj`; no other zone is
   summed or probed. A `psys` zone already includes the packages, so summing it
   beside them counts their energy twice, and a `dram` zone is not CPU energy.
2. **NVML** — the GPU total-energy counter → `HwShunt`, `Partial` (§5.2): it
   reads the GPU and not the CPU's share of the work. Selected only while the
   counter is live by the same rule; the probe cannot load the GPU, so for
   this counter live means published continuously, advancing on the device's
   own draw. A span over which it did not advance is metered by a lower rung.
3. **Per-process CPU energy** (macOS `proc_pid_rusage`, `ri_energy_nj`) —
   the kernel's count of the CPU energy of the metering process's own
   threads → `HwShunt`, `Whole` for work that runs on those threads and
   hands nothing to another process (below). Selected only while the count
   is live by the rule above. Its domain is
   that process's threads alone: not other processes, not its child
   processes, not the GPU, the Neural Engine or DRAM. A span over which the
   process reaped a child (its reaped children's CPU time moved), or at whose
   end it has a child, running or not yet reaped, is `Partial` (§5.2): the
   child's energy is not in the count. Nor is work those threads hand to
   another process: a system service that handles a message of theirs
   (on macOS the kernel bills that service's CPU work to the process
   beside the count, `ri_billed_energy`, never in it), a server reached
   over a socket or a pipe. So a span over work that calls a system
   service, or any other process, is not `Whole` for that work, and the
   reference meter, which does not read the billed count on this rung,
   does not detect it: an implementation SHOULD meter such work on another
   rung, or grade its figure `Partial`. A span over a computation the
   implementation knows runs on the metering thread alone and exchanges
   nothing with any other process (the reference gateway's flow scan, a
   pure computation over the process's own memory) is unaffected, and is
   not made `Partial` by a child the process happens to have either: no
   part of that work can be in the child or in a service. The count is
   folded only when a thread leaves a core, and a process's read of its
   own count does not fold the thread reading (its CPU time is current at
   the read, its energy is not: a thread that burned 1 ms between two
   reads of its own process read no advance in most tries, and up to a
   quantum's worth from before the first read in others), so at each end
   of a span the reading thread SHOULD leave its core just before it
   reads, and an implementation that relies on that MUST check that it
   did. A sleep does not always take a thread
   off its core: when the sleep's timer fires before the thread has
   blocked, the thread goes on without leaving (after 1 µs sleeps at load
   averages of 85 to 118, 4 of about 44,000 showed no context switch, and
   the read after missed the whole burn before it). The reference meter
   sleeps 50 µs, then longer up to 5 ms, until the thread's own energy
   count (`proc_pidinfo`'s `PROC_PIDTHREADCOUNTS`) moved across a sleep,
   which it does only when the thread is folded; a process-wide
   context-switch count is no such check, since another thread's switch
   moves it. A read whose thread was not seen to leave its core is no
   reading of that end, and the span is metered by a lower rung. Work on
   the thread that opens and closes the span is then read whole, and the
   figure also holds what every other thread of the process spent inside
   it: over that work's CPU energy on those threads, not under it.
   It ranks above the
   SoC-wide counter because it attributes the work's own CPU energy rather
   than the machine's. A span over which it did not advance is metered by a
   lower rung.

   The same count, read for every process of ANOTHER process tree the
   metering process can read — the process the work runs in, such as the
   server a gateway relays a call to, and its descendants — meters that
   work → `HwShunt`. Its domain is the CPU energy of the tree's processes
   over the span (as above, not DRAM): the advance of every process present
   at both ends, plus the whole count of every process born inside the
   span. A process is identified by its pid and its start time, so a reused
   pid is another process. The count is folded only when a thread leaves a
   core, so at each end of the span a member's count MUST be read with none
   of its threads on a core or queued for one, or the member is a gap
   (below). Nor is a member at rest while one of its threads is in a wait
   that no signal ends (the kernel's own waits: for a block being written,
   a page being brought in, a lock). That thread is off its core and
   folded, and it runs again when the kernel is done, with nothing
   arriving; a stop that took the member for rest there would read the
   span short by the rest of its work, with nothing seen missing. So at
   each end a member with a thread in such a wait MUST be waited out as a
   running one is, or is the same gap. A walk of the tree reads a parent
   before it lists and reads
   the parent's children, so a child that exits and is reaped in between
   is in that walk neither as a member nor in its parent's reaped time;
   the walk at each end MUST therefore be confirmed by another, one after
   the other, that agrees with it on the tree's processes, each one's
   parent, reaped-descendant CPU time and exit time, with no listed child
   gone by its read, or that end is a gap (below). The span SHOULD end
   after the work's reply. The figure is all the tree spent inside the
   span, whatever for: work the process defers past an earlier reply, or
   does on a timer of its own, is read in whichever span is open when it
   runs, and work after this span's stop is in no figure of this span. So
   `Whole` says that nothing was seen missing from the tree's energy over
   the span, not that all of it was spent on the span's own work. The
   figure is `Partial` (§5.2) when the meter sees any of these over the
   span; each is energy of the tree's over the span that the figure does
   not hold, but for the ones noted:
   - a descendant reaped inside the span (the reaper's reaped-descendant
     CPU time moved; a reaped process's energy is in no counter), unless
     it had already exited at the start, its energy then all spent before
     the span;
   - a member that exited inside the span: its record is final and is
     counted, but it lacks the energy of the member's last time on a core
     (an exiting thread's last slice is never folded into it: up to all of
     a short-lived process's energy), and its GPU time is lost at exit, so
     its GPU use is unknown;
   - a member present at the start and gone at the stop (reaped, or moved
     out of the tree): what it spent in the span before it went is not in
     the figure;
   - a member present at the stop that started before the span and was not
     read at its start: its energy from before the span cannot be
     separated, so none of it is counted;
   - a member that could not be read (a process of another user), or
     whose children could not be listed;
   - a member not at rest at an end. Still running: at the stop its
     reading lacks up to a scheduler quantum of energy per running thread;
     at the START its reading is short by as much, so the figure may hold
     up to that much energy spent just BEFORE the span. Or with a thread
     in a wait no signal ends, in the middle of its work: what that thread
     goes on to spend falls after the stop, or, at the START, inside the
     span though the work began before it. This one can make the figure
     exceed the tree's energy over the span, so a `Partial` tree figure
     with it is not a lower bound, and an implementation that names the
     gaps beside the figure SHOULD name the end (the reference gateway:
     `unsettled-at-start`, `unsettled-at-stop`);
   - an end whose walk no agreeing walk confirmed before the meter stopped
     waiting: the tree kept changing, and a process that came and went may
     be missing;
   - a member holding a GPU client at either end: GPU time its client
     recorded, or GPU use the reader cannot rule out, because a GPU
     client's recorded time is per live command queue and a queue
     created and released inside the span takes its time with it; or GPU
     time that could not be read;
   - a member that has held Neural Engine model memory in-process;
   - work another process did on a member's behalf and the platform billed
     to it (below): the figure holds the bill, which is short of the
     span's delegated work by what no bill carries, and can hold a bill
     for work done before the span that was credited inside it — so, like
     a member still running at the start, it can make the figure exceed
     the tree's energy over the span (the reference gateway: `delegated`);
   - a member credited inside the span with service to other processes
     (below): the credit is subtracted, and need not be the span's own
     service — a credit for service done before the span makes the figure
     short, service done in the span and credited after its stop leaves
     the figure over — so it too can make the figure exceed the tree's
     energy over the span (the reference gateway: `serviced`);
   - a member credited inside the span with more service to other
     processes than it spent and was billed (below): the figure may
     subtract energy spent before the span.

   **Delegated work.** Where the platform bills the CPU work one process
   does on another's behalf to the other, the figure SHOULD count it. On
   macOS the kernel's bank ledgers do: a service's thread that adopts a
   client's voucher while it handles the client's message (an XPC or other
   mach message carries one) is charged, at its next voucher change, with
   the CPU time and energy it spent, which raise the client's
   `ri_billed_energy` and `ri_billed_system_time` (not part of its
   `ri_energy_nj`) and the servicer's `ri_serviced_energy` and
   `ri_serviced_system_time` (part of its own), readable for any process of
   the same user without privilege. Each member's count is then its own
   energy, less its serviced energy, plus its billed energy, and the figure
   the members' advances of it, summed signed: work a member did for other
   processes is theirs, work others did for it is the tree's, and work one
   member did for another is counted once, through the payer. A servicing
   thread is credited at its next voucher change, not as it works, so a
   credit inside a span need not be that span's service: a span over which
   any member's serviced energy or serviced CPU time moved is `Partial`,
   whether or not its net stayed above zero. A member whose net advance is
   below zero — a credit for service done before the span landing inside
   it — is counted as it is, never wrapped, and is a gap too. The bill over
   a span is not the span's delegated work, and errs both ways. Over: a
   bill is credited at the servicing thread's next voucher change, so a
   bill for work done before the span — an earlier call's late credit —
   lands in whichever span is open, and is added to its figure. Short: it
   carries none of the GPU or Neural Engine energy spent in the service
   (one on-device inference request on the reference hardware drew 6.73 J
   on the Neural Engine and billed 1.79 J); none of the service's work on
   threads that never adopt the voucher, or before they do (starting a
   per-client service process); no credit that lands after the span (a
   servicing thread is credited at its next voucher change, which can be
   a second and more after the reply); nothing billed to another
   originator (a scheduling daemon's background work, a policy daemon's
   scan of the client's binary); and nothing sent over a socket or a
   pipe, which carries no voucher, or off the host. So a span over which
   any member's billed energy or billed CPU time moved is `Partial`, its
   figure holding the bill, and an implementation that names the gaps
   beside the figure SHOULD name it (the reference gateway: `delegated`).
   Such a span's stop SHOULD wait, briefly and within a bound, for the
   billed counts to hold, so that a credit landing just after the reply
   falls in its own span rather than the next (the reference meter reads
   them every 5 ms until two reads agree, at most 20 times). The billed
   counts take no part in deciding that a member is settled: another
   process's thread credits them, and the member's run state says nothing
   of when.

   An implementation MUST NOT grade a tree's figure `Whole` on an empty
   list alone. `Whole` further requires the operator's declaration that the
   process hands no work to a process outside its tree — a system service,
   on-device model inference run by a system daemon, a database server, a
   remote API — because a reader of the tree cannot see all such
   delegation: billing shows only work done on a voucher, and the rest of
   that work's energy is charged to the process that does it, which may be
   another user's and unreadable. Without the declaration the figure is
   `Partial`. The declaration does not outweigh evidence: a span with a
   bill (above) is `Partial` under it, the bill showing work done outside
   the tree. A span whose root is gone or unreadable at either end, or
   whose figure did not advance, is metered by a lower rung. The reference
   gateway (`jcp-gateway`) meters each gated call this way where the
   upstream server is a local process it can read.
4. **Apple IOReport** (the Energy Model group) — the SoC energy counters
   → `HwShunt`, selected only while every CPU energy rail (one per die on a
   multi-die part) is live by the rule above. A span over which no CPU rail
   advanced is metered by the next
   rung, never graded `HwShunt`. Where the CPU rail is not live but the GPU rail
   is, the rung is declined rather than reporting the GPU rail as a `Partial`
   reading of the work: the GPU rail is machine-wide (display compositing
   moves it), so booking it to a task misattributes energy the task did not
   spend, and the partial figure would under-debit the budget, the CPU's share
   going unmetered. The whole-task estimate below is the honest figure.
5. **Time × calibrated power** — wall-clock duration times an active power
   calibrated on the host (a measurement of that host's draw, such as its
   draw under load read from a shunt or a live counter) → `ModelBased`.
6. **Estimate** → `Estimator`: a figure no host calibrated. Either wall-clock
   duration times a default or assumed active power — a documented default,
   a nameplate or datasheet figure, another host's calibration — or a
   constant per-operation figure. Always available. Time × power is graded by
   its power: the power is the whole of the model, so a power not calibrated
   on the host makes the figure a constant scaled by time, and grading it
   `ModelBased` would claim better provenance than how it was obtained. A
   counter rung's span that degrades because its counter did not read the
   span (above) lands here, unless the implementation holds a host-calibrated
   power, in which case it lands on rung 5.

### 5.2 Coverage (REQUIRED where a figure is partial)

Provenance says how a figure was obtained; **coverage** says how much of the
work it accounts for: `Whole` (the default) or `Partial`. A reading can be
hardware-rooted and still cover only part of the work — an accelerator counter
on a host whose CPU has none — and its remainder is unmeasured, not zero. A
`Partial` figure MUST NOT be presented as a task total.

On the wire coverage is carried only when it is `Partial`, and omitted when
`Whole`, so a figure sealed before the field existed keeps its exact bytes; a
verifier reads an absent field as `Whole`. The `jcp/2` receipt carries it as
`energy_coverage` (§8); the `intercept/1` step receipt of the reference
interceptor chain (`jcp-intercept`) carries it as `coverage`, beside its
`provenance`. In `intercept/1` the field, when present, has exactly one value,
`"Partial"`; a verifier MUST refuse any other value, `"Whole"` included, rather
than read it either way.

The `efficiency/1` outcome receipt (§5.3) carries it as `energy_coverage`,
under the same one-value rule. Its headline figures are ratios of a task total:
energy per accepted token on the receipt, and energy per completed unit over a
ledger. So a `Partial` receipt MUST NOT carry `uj_per_accepted_token`, and a
verifier MUST refuse one that does. A total summed from receipts of which any
is `Partial` is `Partial` and gives no per-unit energy figure. A partial figure
divided by tokens is not what the work cost per token, nor a bound on it,
because a partial figure may fall short of the work or exceed it. The vectors
are in [`conformance/v1/efficiency.json`](conformance/v1/efficiency.json).

### 5.3 Traceability across operation, activity and outcome (REQUIRED where an outcome claims energy)

§5.1 grades a figure and §5.2 says how much of the work it covers. Neither says
which *level* the figure belongs to, and a deployed system is answerable at
three that are not interchangeable:

| level | artifact | what it answers |
|---|---|---|
| **operation** | the `jcp/2` work receipt (§8) | what one unit of work cost |
| **activity** | the AR-1 action receipt | what one *authorised* thing did — grant checked, environment attested, work performed, effect committed, bytes egressed |
| **outcome** | `efficiency/1` | whether any of it bought anything |

Two rules follow, and both exist because the levels were previously joined by
nothing.

**An outcome's energy MUST name its activities by content-id.** An outcome that
identifies its unit of work by a label can assert any total it likes, and the
level where the number is most load-bearing is then the level with no binding
at all. A verifier MUST be able to resolve each named activity and re-add the
figures. An outcome naming no activity is not traceable and MUST NOT be
presented as though it were.

**The difference MUST be named, not absorbed.** The sum of an outcome's
operations is not the outcome's energy. Authorising, attesting, verifying and
pushing bytes off the device all cost joules that no operation receipt claims —
AR-1 carries exactly one energy figure and it belongs to the work. That
difference is the cost of being accountable, and it is reported as a signed
residual in exact integer microjoules:

```text
residual_uj = outcome_total_uj − Σ activity.joules_uj
```

The residual MUST NOT be clamped. Negative means the bound activities claim more
than the outcome admits — double-counted work, or an activity bound to the wrong
outcome — which is a finding. An activity bound twice MUST be refused before the
sum: counting it twice inflates the accounted figure and shrinks the residual,
flattering the trace in the one direction that matters.

The composite provenance is the **floor** across the outcome and every activity,
never an average and never the best: one `Estimator` part makes the whole trace
estimated, because a trace cannot be more measured than the part of it that was
not measured. Coverage takes the floor the same way. Aggregation MUST NOT be a
step at which a grade improves.

### 5.4 Dynamism: what a deviation establishes, and what it never does (OPTIONAL)

A total cannot distinguish a throttled part from a contended host from a cold
cache, and on silicon nobody has profiled it cannot say what normal was. Where
an implementation carries energy as a time-resolved series, the following hold.

**Provenance and coverage are per-window.** Counters do not have uniform
resolution — a package counter may be continuous while an accelerator counter is
live only while a queue is — so one grade for a series is a grade for its best
window.

**A baseline MUST declare where it was measured, and MUST NOT be applied
outside it.** An expectation is a fit, and a fit is only a fit somewhere.
Comparing against a baseline taken on different silicon, at a different
occupancy, or at a coarser sampling window manufactures deviations that say
nothing about the workload. An undeclared baseline is inapplicable, so
forgetting to declare under-claims rather than inventing findings.

**A cause explains a deviation only with evidence.** A named operational cause
— thermal limit, host contention, cold start, frequency scaling, a meter
coverage gap — MUST carry a content-id for the reading that supports it.
Otherwise "explained" becomes the laundering channel: anything inconvenient is
labelled and disappears.

**Absence of detection is not absence.** A phenomenon shorter than one sampling
window is invisible to it. That case is *unresolvable*, never *within band*.
Reporting nothing found at a resolution that could not have found it is the
quietest available way to be wrong.

**There MUST NOT be a verdict of intent.** Energy does not carry intent. A
deviation says the joules did not match the expectation; it cannot say why, and
on an uncharted stack it cannot say the expectation was right. An unexplained
window is a request for a look, not a finding of wrongdoing, and a conformant
implementation MUST NOT collapse unexplained, inapplicable and unresolvable
windows into a single score — they mean "look at this", "we have no yardstick"
and "we could not have seen it", which call for different responses.

### 5.5 Conservation, not anomaly detection (REQUIRED where a grant is reconciled)

The strong instrument is conservation, and it needs no model of normal.

Because every figure is integer microjoules and the sums are exact, the smallest
representable discrepancy against a grant of `N` µJ is `1_000_000 / N` parts per
million — on a one-joule grant, **1 ppm**. A drift of a fraction of a percent is
three orders of magnitude above that floor. The question "do these joules fit
anywhere" is therefore answerable arithmetically, where "is this unusual" is not.

This property does not survive the usual construction, and that is the point.
Where an energy figure is *derived from a performance proxy* — utilisation times
a thermal design figure, tokens times a coefficient, instructions times
picojoules-per-operation — its model error is tens of percent, and a
fractional-percent drift is invisible by construction: it sits far inside the
error bar of the estimator that produced it. Deriving energy from performance
does not merely lose accuracy, it destroys the ability to notice that the books
do not balance. A conformant implementation MUST report whether the floor it
quotes rests on a measured figure or a modelled one; quoting a 1 ppm floor over
`Estimator` windows is exact arithmetic over a model's opinion.

### 5.6 The physical bound: refutation, never verification (OPTIONAL)

The joules were spent and the heat left the die; that quantity is not
negotiable. The *measurement* of it is exactly as forgeable as any other report
— a counter can be lied about, a figure fabricated, a `HwShunt` grade stamped on
a number from nowhere.

What energy has that a synthetic unit does not is an **external referent**. A
token count, a call count, a request rate can be any number whatsoever: nothing
outside the accounting system constrains them, so a fabricated one is
indistinguishable from a true one. A joule count is bounded by the part's power
envelope and the window's duration.

An implementation MAY therefore refute a figure that exceeds what the part could
have drawn, or falls below what a powered, executing part must draw — the second
being the shape a frozen counter takes, and the reason §5's liveness rule
exists. Bounds SHOULD be deliberately permissive (burst rating for the ceiling,
absolute idle for the floor) so that a refutation is unassailable.

Passing a physical bound MUST NOT be reported as verification. A fabricated
figure inside the envelope passes, and must: the bound establishes only that the
figure is not impossible.

**The one thing that cannot be faked is how the energy was consumed.** A report
of a profile is just numbers, but a part cannot draw the profile of one
computation while performing another. This closes the hole conservation leaves:
a total can be balanced by padding — invent an operation and attribute the
surplus to it — but a fictional operation contributes no draw, so the observed
character stops matching the declared work mix.

Character is also what survives unfamiliar hardware. The magnitudes move with
the process node, the power envelope and the memory hierarchy, which is why a
band baseline fails there; a compute-bound steady state is flat on any part, a
cold start ramps on any part, a queue-driven workload alternates on any part. A
declared shape MUST be declared in advance — classifying the observed shape and
then deciding what it must have been is post-hoc pattern matching. An
inconsistency refutes the declaration. It attributes no cause and names no
intent.

## 6. The Frame (REQUIRED)

Every message is a typed Frame — one envelope across the tool layer (MCP's job)
and the agent layer (A2A's job):

```
Frame {
  kind:         tool-call | tool-result | agent-task | agent-reply
  trust:        control | data
  grant_id:     content-id        ; the grant this frame exercises
  action:       string
  resource:     content-id
  payload_hash: content-id
  model:        content-id?  ; OPTIONAL: the model the action runs (§4.9 ModelIn)
}
```

`model` is filled by the runtime that dispatches the call, which knows the
content address of the model it will load. It is absent for an action that runs
no model, and an absent `model` is omitted from the canonical encoding, so such
a frame, and every receipt and vector over it, is byte-identical to one built
before the field existed.

`trust = data` marks untrusted output. A conforming runtime MUST refuse to
exercise any capability from a `data` frame (§7), regardless of its content. Tool
and agent outputs MUST be tagged `data`.

### 6.1 Continuations: work that answers later (REQUIRED where an enforcement point relays it)

An endpoint may answer a request before its work is done. It may return a
**handle** to work that goes on, whose outcome the client fetches in later
exchanges that name the handle (MCP's tasks; an A2A task). Or it may return an
**interim result** asking the client for input, which the client answers by
repeating the request with its answers attached (MCP's `input_required`). Each
later exchange moves the first request's work and the first request's kind of
payload: the outcome is the action's output, and the answers go to the
endpoint as its arguments do. So an enforcement point that relays such
exchanges and treats the later ones as discovery has a door beside its gate.
Such an enforcement point:

1. MUST decide each later exchange as a **continuation** of the first: against
   the grant the first was decided on, at the moment the later exchange is
   made, so a grant revoked, expired or spent since refuses it (§7), and the
   outcome of work the grant no longer covers is not fetched under it. It MUST
   seal each continuation's decision (§8) over a payload that commits to the id
   of the receipt it continues (the first's, or the previous continuation's),
   so that one piece of work's receipts form a chain within the stream.
2. MUST refuse, and seal the refusal of, a later exchange naming a handle it did
   not see returned to a request it decided. A handle is a bearer string the
   endpoint minted, and no grant covers it until the enforcement point has bound
   it to one. A handle it has closed (item 4) or forgotten (a restart, a bound on
   what it keeps) is refused the same way: fail closed.
3. MUST flow-check what the client authored in a later exchange (the answers)
   as it does the first request's arguments (§4.10), and MUST admit what comes
   back — status text and the outcome alike — as `data`, labelled as the output
   of the first request's action and joined with the label of everything the
   client has sent the work. An exact echo of a handle or state the endpoint
   minted MAY be left out of the flow check, being the endpoint's own text;
   anything else the client sends, a handle it made up included, MUST be
   checked. What the enforcement point forwards SHOULD be built from the fields
   it decided over, so that nothing it did not check reaches the endpoint.
4. MUST meter and debit each later exchange (§5), and, when the outcome is
   final, MUST seal one **closing record**: a `tool-result` (or `agent-reply`)
   frame tagged `data`, debiting nothing, whose payload is the content address
   (§3) of a record naming the first receipt, every continuation's receipt, the
   joules metered across them, and the outcome's digest. The endpoint works
   between exchanges, inside no exchange's span, so the record MUST mark that
   sum `Partial` (§5.2) unless the meter read the work between them.

The reference gateway (`jcp-gateway`) does this for MCP's `tasks/get`,
`tasks/update` and `tasks/cancel`, and for a `tools/call`, `resources/read` or
`prompts/get` retried with `inputResponses`; it advertises it in its
`server/discover` and `initialize` responses as `continuations`.

## 7. The decision (REQUIRED)

Given a Frame, its [SignedGrant](#4-the-grant-required), the caller's DID, a
metered `cost`, and the current time, a conforming runtime decides **Allow** or
**Deny(reason)** by these checks, in order, the first failure naming the reason:

1. `data`-tagged frame → **Deny(untrusted-control)**.
2. grant revoked (§4.3) → **Deny(revoked)**.
3. the issuer (for a chain: the **root** issuer) is not a configured trust root
   for the frame's resource → **Deny(untrusted-root)**. An unconfigured
   resource is open; once any root is configured, only listed keys anchor
   authority.
4. frame's `grant_id` ≠ the grant's id → **Deny(wrong-resource)**.
5. caller ≠ grant's `subject` → **Deny(wrong-subject)**.
6. issuer signature invalid → **Deny(bad-signature)**.
7. expired → **Deny(expired)**.
8. frame's `resource` ≠ capability's resource → **Deny(wrong-resource)**;
   action ∉ capability → **Deny(action-not-granted)**.
9. prior spend + `cost` > `budget_uj` → **Deny(budget-exceeded)**.

On **Allow**, the runtime debits `cost` from the grant's remaining budget. The
decision is pure and deterministic; no model participates.

For a delegated grant the decision takes the full chain (§4.2): structure and
signatures first (`chain-broken` / `attenuation-violation` / `bad-signature`),
then revocation against **every** link (`revoked`), expiry of every link, the
leaf-level checks above, and budget against every link, debiting all on allow.

**Trust roots.** An enforcement point is **configured** with the root-issuer
keys it honors for each resource it guards — the same local, pairwise trust
shape as an `authorized_keys` file. `did:key` is self-certifying, so there is no
certificate authority and no hierarchy; cross-organization federation is a
directory convenience, never a trust dependency. A chain whose root issuer is
not configured for the frame's resource MUST be denied.

## 8. The receipt (REQUIRED)

Every decision — allow and deny — seals as a **JCR-1** receipt: canonical
det-CBOR inside a COSE_Sign1 envelope (RFC 9052, EdDSA/Ed25519, `alg = -8`). The
`jcp/2` profile signs a flat uint/text core:

```
jcp_receipt        = "jcp/2"
grant_id           = hex
issuer, subject    = did:key
resource, action   = hex, string
payload_hash       = hex                   ; content id of what was acted upon
frame_kind, trust  = "tool-call"…, "control" | "data"
verdict            = "allow" | "deny"
deny_reason        = string                ; omitted on allow
cost_uj            = uint
energy_provenance  = "HwShunt" | "ModelBased" | "Estimator"
energy_coverage    = "Partial"             ; omitted when Whole (§5.2)
budget_remaining_uj, budget_uj = uint
steward            = the signer's id (the COSE kid)
```

`payload_hash` binds the receipt to **what the decision was over**, not merely
that a decision was taken. Without it a tool-call receipt does not commit its
arguments, and two decisions differing only in payload seal to identical bytes —
so a payload substitution is undetectable in the audit. A conforming
implementation MUST include it.

**Profile versioning (REQUIRED).** The predecessor profile `jcp/1` omitted
`payload_hash`. A verifier MUST read the `jcp_receipt` tag from the signed core
and recompute the preimage under **that** profile, so receipts sealed as `jcp/1`
keep verifying; it MUST refuse a tag it does not implement rather than assume a
profile. Implementations MUST NOT mint `jcp/1`. Published `jcp/1` receipts are
retained in [`conformance/v1/legacy_jcp1.json`](conformance/v1/legacy_jcp1.json)
precisely so a verifier can prove it honours the migration.

Receipts share the family's JCR-1 **envelope** (canonical det-CBOR +
COSE_Sign1/EdDSA) and vocabulary, so family tooling verifies signatures and
decodes cores uniformly. The *core field list is per-profile*: sibling profiles
(`map/1`, Sandbox, proof, jouleclaw's receipt kinds) sign their own cores, and a
change to `jcp/…` does not alter theirs. A deny receipt is as load-bearing as an
allow receipt — the record of a refused attack is evidence.

### 8.1 The receipt stream is the write-ahead log (REQUIRED)

A receipt carries everything enforcement needs to remember: the grant, the
verdict, the debit. A receipt sealed for a **multi-hop chain** MUST additionally
record `chain_ids` — every grant id the decision debited, root first — so the
stream is self-contained. An enforcement point that crashes MUST be able to
rebuild its budget state by replaying its own verified receipt stream:
allow-receipts re-apply their debits to every grant they name, ancestors
included, with no external index. Durability and audit are the same artifact;
there is no second database to drift from the evidence.

### 8.2 Group accounting (RECOMMENDED profile)

When a team works under sub-grants of one root, aggregation of the verified
receipt stream keyed by `(grant_id, subject)` yields per-worker attribution —
actions allowed, actions refused, joules spent — computed from evidence rather
than self-reporting. An aggregator MUST count (not silently skip) receipts that
fail verification: a ledger that drops entries is not a ledger.

## 9. Conformance

A conforming implementation passes, with real checks:

- **C1 — signed grants + audience binding:** a grant verifies only for its
  issuer's key and authorizes only its exact content-addressed resource.
- **C2 — attenuation monotonicity:** a child grant can only narrow; any
  amplification or non-holder delegation is refused.
- **C3 — budget enforcement:** spend beyond the ceiling is denied deterministically
  and the denied call is never debited.
- **C4 — data/control separation:** a `data` frame can never exercise a capability.
- **C5 — the receipt:** every decision seals to JCR-1 and verifies; a wrong
  steward fails.
- **C6 — revocation:** only the issuer revokes; a revoked ancestor kills the
  whole delegation subtree; a forged revocation is never admitted.
- **C7 — chain conservation:** a multi-hop chain verifies hop by hop; a spliced
  chain is refused; sibling delegations share their root's budget; a chained
  decision enforces an inherited caveat; and a chain whose child dropped one does
  not verify.
- **C8 — wire binding:** a request round-trips the `jcp/1` framing canonically
  and an enforcement endpoint decides + seals over a live socket; a request
  presented again gets its first response unchanged and is not debited again,
  and one stamped outside the window is refused `stale-request`.
- **C9 — energy honesty:** the metering ladder always yields a figure and its
  provenance is internally consistent.
- **C10 — delegation control:** depth narrows per hop; a leaf-only grant cannot
  be delegated, at mint or in chain verification.
- **C11 — sharded budgets:** shard slices that sum past the parent are refused
  at mint.
- **C12 — rotation:** an un-pre-committed key cannot rotate an identity, the
  pre-committed key can, and the identity DID is stable across rotations.
- **C13 — caller-signed wire + threshold:** an impersonated caller is denied
  `bad-signature`; one share of a 2-of-3 actor cannot sign; two can, and the
  endpoint treats the aggregate as any caller signature.
- **C14 — replay + group accounting:** a fresh runtime restores spend from the
  verified receipt stream; per-worker attribution matches the original ledger;
  a tampered receipt contributes nothing.
- **C15 — trust roots:** an unconfigured resource is open; once configured, a
  validly-signed grant from an unlisted root is denied `untrusted-root` on both
  decision paths.
- **C16 — chained replay:** a chained receipt names every debited link; a fresh
  runtime replays leaf AND ancestor debits exactly and keeps conserving.
- **C17 — budget channels:** optimistic spends settle by signed statement;
  forged, stale, and regressing statements are refused; post-settlement the
  merged ledger binds and any overdraft is provable.
- **C18 — DKG:** a dealer-free k-of-n actor signs a standard Ed25519 signature;
- **C19 — injection resistance:** a tool-poisoning corpus presented through an untrusted frame cannot steer a decision; the trust tag is structural, so data cannot become control. (Implemented in the harness since before this entry: §9 listed every other clause but omitted this one.)
  one share cannot.
- **C-neg:** expired, forged, and wrong-caller frames are denied with the precise
  reason; canonical CBOR rejects floats/bools.

Published vectors live in [`conformance/v1/`](conformance/v1/) (`cargo run -p
jcp-conformance --bin gen-vectors`).

## 10. What JCP closes (mapping to the documented failures)

| Documented failure | JCP mechanism |
|---|---|
| Prompts-as-security; inverse scaling law | §1.1 enforcement outside the model |
| Tool poisoning / asymmetric trust | §6 data≠control |
| Ambient authority, static env secrets | §4 unforgeable grants |
| Confused deputy, token passthrough | §4 content-addressed audience binding |
| Delegated authority | §4.1–§4.2 attenuation chains, verified hop by hop |
| A leaked grant lives forever | §4.3 issuer revocation, subtree-transitive |
| Runaway loops, no energy accounting | §5 enforced joule budgets, conserved across delegation |
| Mutable / optional audit | §8 signed JCR-1 receipts |
| Fragmentation across protocols | §6 one frame for tools and agents |
| Results fetched later, outside the gate (task handles, interim input requests) | §6.1 continuations, chained to the receipt they continue |

## 11. Wire binding: `jcp/1` over a byte stream (OPTIONAL)

JCP is transport-agnostic; this section defines the reference binding for any
ordered byte stream (TCP, Unix socket, pipe, TLS session).

**Framing.** Each message is a 4-byte big-endian length followed by the
message's **canonical det-CBOR** bytes (§3). A receiver MUST reject a message
longer than its declared cap and MUST reject a non-canonical body — a message
has exactly one wire form, so content addressing stays injective across the
wire.

**Messages.** One request/response pair per action:

```
WireRequest  { frame, chain: [signed-grant…], caller: did:key, cost: energy,
               now: uint, signature: bytes(64) }
WireResponse { decision, receipt: bytes }
```

`chain` is the full delegation chain, root first (a single un-delegated grant is
a one-link chain). The response carries the deterministic decision and the
sealed JCR-1 receipt, deny and allow alike.

**Time is the enforcement point's.** The endpoint MUST decide at its own clock.
`now` is the time the caller signed the request, a freshness stamp: the endpoint
MUST deny a request whose `now` differs from its clock by more than its
freshness window as `stale-request`, and MUST NOT decide any request as of the
time it names. A caller that chose the decision time could stamp a request
before its grant expired, or inside a time caveat it has since left.

**Caller authentication.** `signature` is Ed25519 by `caller` over the
canonical bytes of every other field. The endpoint MUST verify it before
deciding; a request that fails is denied `bad-signature` and sealed like any
other decision — the refused impersonation is evidence too. A threshold actor
(§2.2) signs the same field with a k-of-n aggregate, indistinguishable on the
wire.

**A request is decided once.** The wire never carries a secret: grants are
signed public objects, and a request is a signed public object too. The endpoint
MUST remember every request it decided while that request's stamp is inside the
window, together with its response. A request presented again MUST get that same
response, the same decision and the same receipt, and MUST NOT be decided,
debited or recorded a second time. That holds across a restart, which replays
the record (below) and so recovers each response: the decision is deterministic
and so is the seal. A captured request is therefore worth nothing beyond what
its first exchange carried. A caller whose response was lost recovers it by
sending the same request again. Refusing the repeat instead would leave that
caller only one way to learn the outcome: stamp the request again, which is a
new request and decides and debits the same action a second time. A request
refused before it was decided (stale, unauthenticated, or `unrecorded`) is not
remembered. Once a request's stamp has left the window, a repeat of it is stale.
Two byte-identical requests are one request, so a caller that means the same
action twice stamps the second later.

**The cost is the holder's measurement.** `cost` is the requester's own figure,
and the endpoint debits it as given, so a budget binds only as honestly as that
figure was measured. The requester MUST therefore be the party that metered the
action: the grant's holder is the enforcement point that ran it, as a steward is
in Mesh, rather than the agent the grant governs reporting its own consumption.
The endpoint behind the enforcement point is invoked only on Allow.

**What the endpoint holds survives it.** An endpoint that can restart is bound
by §4.11: it MUST record each request it admits before deciding it, and each
revocation it honours before applying it, and on restart MUST replay them
through the gate, each request at the time it was decided, before deciding
anything new. That re-establishes the spend, the revocations, and the memory of
requests still inside the window. A request it cannot record is denied
`unrecorded`.

## 12. Non-goals

JCP does not define a model or a tool registry. It does
not detect misbehavior with probabilistic "trust scores" — bounded capability and
enforced budget make the dangerous action unreachable rather than observed.
Behavioral scoring MAY be layered above as a secondary signal; it is never the
gate.
- **C20 — standing grants:** an event-triggered firing mints a strictly-weaker child that traces to the standing grant; the wrong event class, an exhausted firing count, and a firing past the cumulative budget are each refused without advancing the ledger; a policy that would amplify its grant is refused at author time.
- **C21 — checkpoints + amendments:** an agent checkpoints its own state; only the grant's issuer — a principal distinct from the agent — may amend, only to tighten, and only up to what the grant has **left** at that checkpoint; an agent steering itself, a non-issuer amending, a widening amendment, an amendment above the remainder of a checkpoint that stated its spend, and an amendment bound to a different checkpoint are each refused.
- **C23 — receipt binds its payload + profile migration:** a `jcp/2` receipt commits the `payload_hash` it decided over; substituting only the payload (same grant, action, and cost) fails verification — the forgery a payload-blind receipt could not detect; and receipts published under `jcp/1` still verify, because the verifier recomputes the preimage under the profile the receipt names rather than assuming one.
- **C24 — caveats:** a caveat embedded in the grant is enforced offline by the decision point; `PayloadIs` makes a grant one-shot (a replay with different arguments is denied `caveat-unsatisfied`); `MaxCostUj` bounds a single invocation independently of the cumulative budget; `ModelIn` admits only a frame naming one of its models, and denies a frame naming another or none; a caveat may narrow below the capability; attenuation inherits every parent caveat and refuses to drop one; and a grant with no caveats encodes byte-identically to one minted before they existed.
- **C25 — flow:** a value's propagated label must sit at or below the grant's ceiling and clear the resource's taint floor; an argument whose provenance was **not computed** fails closed against a ceilinged grant rather than passing as clean; several ceilings in a chain compose to the strictest and attenuation never widens one; a flow denial debits nothing, so a tainted call costs zero joules; and a grant with no flow caveat encodes byte-identically to one minted before flow existed.
- **C22 — dual ceilings:** a money ceiling beside the joule ceiling is enforced independently — the first to bind denies (`funds-exceeded` / `budget-exceeded`), a foreign-currency cost is refused `currency-mismatch` rather than converted, an allow debits both dimensions and a deny debits neither, and an absent money ceiling encodes byte-identically to a joules-only grant (so its content id and signature are unchanged). Attenuation narrows money independently: dropping, raising, or re-denominating a bounded parent's ceiling is refused; a plain attenuation inherits it. A priced chained decision applies the parent's money ceiling to its child's spend, and a delegated grant decided without its chain is denied `chain-broken`.
- **C26 — one writer per record:** a durable decision record admits one writer. A second holder of the same record is refused rather than opening it, because two would each replay it, each re-establish the same remaining budget and each debit against it — spending it twice while every entry still verifies (§5). The exclusive claim is released when the holder goes away, so a restart can reopen the record it previously held, and a platform that cannot provide exclusivity fails the open rather than proceeding.