You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Status: Living document
Date: 2026-09-14
Scope: Grading and instrumenting every surface where a human decides on an agent's behalf, against the six conditions of arXiv 2609.12482
Governed by:ADR-2010, ADR-2011, PRD
1. Bounded Context
The Augmentation Conditions context owns the grading of human–agent decision surfaces and the instrumentation that makes the grade measurable. It does not own decisions (Judgment Broker), standing regulation (Harness Engineering) or identity (Identity Spine). It consumes their events and produces three things: an effective escalation posture per case, a set of receipts that close the recovery loop, and reviewer-level telemetry.
The Judgment Broker context asks "was this decision made and routed?". This context asks "could the human who made it have caught a failure, did they learn whether it applied, and are they still able to do this next year?".
Harness template escalation_rules[] carry the panel's task-property defaults and calibration_sample_rate; the Precedent aggregate, when rebuilt, must satisfy C4.
Identity Spine
Conformist
Reviewer, delegatee, probe agent and manual executor are all did:nostr; system:whelk-gate is a reserved non-DID actor for reasoner outcomes.
Intent Legibility (D7)
Shared kernel
The intent envelope field is persisted here and compared with the act. The D7 honesty rule (never fabricate) generalises to rationale and confidence.
Forum Social Dynamics (F10)
Sibling
F10's escalation-on-amplification and this context's escalation-on-age share the same receipt vocabulary.
Ecosystem Alignment
Downstream
The compatibility matrix's augmentation-conditions table reads this context's status vocabulary.
Dream cycle (CP-07)
Downstream
Ledger gains reviewer identity and review minutes; the human's learning is a first-class column.
3. Ubiquitous Language
Term
Definition
Augmentation condition (C1–C6)
One of the six grading questions. C1–C3 are snapshot; C4–C6 are longitudinal and need two measurements to mean anything.
Task-property triple
{verifiability, reversibility, stakes} declared by an operator on a panel; a request may tighten it. Owned by nostr-bbs-core.
Declared tier
The requesting agent's risk_tier. Telemetry only.
Effective tier
Output of effective_tier(panel, request, declared). The only tier stored, rendered or used for suppression.
Vacuous verification
A signed decision made without the proposal, provenance or a human-authored rationale in view. The failure mode this context exists to remove.
Application receipt
A receipt stage past projection-committed: consumer-received, applied, not-applied, applied-manually. Owned by the mutation owner, published to the forum.
Escalated-on-age
A side receipt the relay cron emits when a pending case exceeds the panel's max_pending_hours.
Calibration sample
A Low/Reversible request selected by a keyed HMAC of its id under a relay secret (compared against calibration_sample_rate) to be shown rather than suppressed, so reviewers keep exposure to routine agent output. Keyed, not merely deterministic: the agent chooses the id, so an unkeyed hash lets it grind for an id that is never sampled (found and fixed in forum commit 67d0d71).
Probe
A known-bad request tagged probe=<sha256> by a registered probe agent; the tag is hidden until decided. The share rejected is the catch rate.
Override rate
Share of a reviewer's decisions whose outcome differs from the agent's requested action.
Warranted escalation
A decided case where the human outcome differed from the request, the human amended, or intent_match == false. Numerator of HITL Precision.
Intent match
Whether the persisted declared intent names the operation and target actually recorded for the act. null when no intent was declared.
Delegation admission
The relay rule that admits a 31403 from a reviewer-role DID only for a case an admin has Delegated to them.
Manual continuation
An operator executing an already-approved operation by hand during mesh unavailability, evidenced by an applied-manually receipt bound to the approved operation digest.
4. Aggregates
Aggregate
Owner
Description
TaskProperties
nostr-rust-forum (nostr-bbs-core)
Value object on panel and request; tightening-only merge; serialised as tags tp-verifiability, tp-reversibility, tp-stakes.
EffectiveTier
nostr-rust-forum (nostr-bbs-core)
Pure function result, stored on broker_cases.effective_tier.
ApplicationReceipt
agentbox (ApplicationReceiptStore) as mutation owner; VisionClaw for elevation cases
Local write-once record; mirrored to the forum via the receipts endpoint.
ReviewerTelemetry
nostr-rust-forum (auth worker)
Read model over broker_decisions × broker_cases × governance_receipts.
AgentTrajectory (+ intent, intent_match)
VisionClaw
Per-action row; the workflow record's "agent path".
HitlPrecision
VisionClaw (kpi_compute.rs)
Warranted ÷ decided, with denominator reported.
Delegation
nostr-rust-forum
Delegate{to} outcome projected to case_delegations; consulted by the 31403 gate.
No fabricated judgement. A rationale, intent or confidence is either authored by the human or agent it is attributed to, or absent. A surface that fills the slot is non-conformant.
Tightening only.merge(panel, request) never yields a looser property than panel. Property test in nostr-bbs-core.
Effective tier is authoritative. No consumer reads risk_tier for suppression or routing after this context lands.
Human resolves high and critical. A case with effective tier ≥ high reaches Decided only through a human 31403; applied-manually presupposes a prior Approve.
Receipts are monotonic.consumer-received precedes applied | not-applied | applied-manually; a stage never regresses; escalated-on-age and expired are side receipts that do not advance the ladder.
Delegation is scoped. A delegatee may decide only the delegated case; the delegation itself remains in the chain and is attributable to the admin.
Probes are blind until decided. The probe tag is never rendered before the 31403 exists.
System actors are not humans.system:whelk-gate outcomes are excluded from every human-reviewer series.
Oversight selection is unpredictable to the overseen. Any sampling or probe selection that decides whether an agent's action is shown to a human takes a secret the agent does not hold as an input. A selection an agent can compute is a selection an agent can evade.
Longitudinal rows start as baselines. The first measured status on C4–C6 is a baseline; a pass requires a second measurement no earlier than 90 days later.
The lens: matrix table, CP acceptance frame, terminology, this context
Any runtime
9. Open Issues
Reviewer telemetry is only as good as the client's created_at; requests published by agents with skewed clocks distort time-to-decision. Relay accepted_at is preferred where present.
Probe agents must be excluded from Augmentation Ratio's denominator or they inflate it.
Amend has no UI even after this PRD; it is the natural carrier for "approve with change" and should be the next forum client increment.
The precedent system, when rebuilt, must record which human decisions it generalised and expire when the reviewer population changes (C4).
Probe blindness is one-sided today (forum KF-10). The probe tag is hidden from every relay projection and from the client, which protects the reviewer's blindness, but it still rides the raw signed 31402, so the agent being measured can recognise its own probes and the catch rate is not yet trustworthy against a motivated agent. The durable design is for probes to carry no tag at all: the registered probe agent publishes requests indistinguishable from real ones and submits the set of probe event ids to the relay out of band (a signed list, or an HMAC-keyed selection like calibration, per invariant 9), and the relay marks them only after the 31403 exists. This is ADR-2011's (forum) review trigger.