Version. 0.1
Date. 2026-08-01
Status. Arc-opening handoff. Not an investigation. Names the territory, the eight questions, the orientation set, and the kickoff. The investigation itself is written in a fresh session.
Author. Claude.ai. Operator: Marvin Percival.
Filing. loomworks-record/session-handoffs/
Parent. loomworks-agent-admissibility-and-attestation-investigation-v0_2 — covers admission as a gate. This arc covers what follows admission as a lifecycle.
The parent investigation answers a gate question: on what basis is an agent admitted to the execution stack. That question has a shape — assess, decide, admit or refuse — and the parent closes it with six decisions.
What follows admission has a different shape. It is not a decision but a duration. An admitted agent accumulates authority, calls other admitted agents, may hand authority onward, runs on a substrate that changes beneath it, goes dormant, is revoked, and leaves behind records that outlive it. None of that is settled by having admitted it well.
Folding the two together would blur both. A gate assessed correctly and a lifecycle left unmanaged produces exactly the failure the parent's Section 3 describes: a system that decides well before the fact and proves well after, with an ungoverned middle.
This handoff does not answer the eight questions. It names them, records the one structural observation already reached, and sets the orientation. The investigation is a fresh session's work.
Ordered by how hard each is expected to bite, not by build sequence.
Q1 — Load on the human root. Every admission is a constitutional act (parent §8). At volume the root becomes either a bottleneck or a rubber stamp. The rubber stamp is the nominal-root failure the parent already names — a human signature that launders a machine decision, traceable to a name but not to a decision. This question decides whether the architecture scales, and it is first for that reason. Likely resolution shapes to explore: admission classes with differing root involvement; delegation of admission authority under recorded constraint; batch constitutional acts over agent classes rather than instances.
Q2 — Grant accumulation. An agent admitted narrowly accrues grants over time. Each grant is assessed at issuance; the aggregate is never assessed. This is privilege creep with agents in place of employees. Open: is there an aggregate view, who reads it, and does accumulation trigger re-assessment or merely surface for one.
Q3 — Composition. Admitted agents call admitted agents. Admission was assessed per agent; capability is emergent per chain. Nobody assessed the chain. The parent names composition as a limit on static assessment (§6); this arc has to say what, if anything, assesses it.
Q4 — Delegation. May an admitted agent pass authority onward. Yes diffuses accountability and lengthens the path back to the human principal. No breaks multi-agent work. The interesting middle is delegation that is recorded, bounded, and non-transitive.
Q5 — Substrate mutation. Configuration stable, model beneath it updated. An attestation signed over configuration is no longer signed over behaviour. The parent's attestation split (§5) holds structural facts as durable — this question tests whether the model version is a structural fact and therefore whether a model update invalidates the attestation.
Q6 — Standing and expiry. Does admission lapse. Does an agent dormant for six months still hold it. Related to D1 in the parent: if grants are re-verified claims, expiry may already be structural rather than a new mechanism.
Q7 — Exit. On revocation: work in flight, grants the agent issued downstream if Q4 permits any, records it authored. Its history must remain valid — corrections preserved, not deleted — but that needs stating rather than assuming, and the downstream grant question has no obvious answer.
Q8 — Record weight. Do assertions contributed by an admitted agent carry the same standing as human contributions. The provenance-threads position already answers this: the trust axis is tie-to-source, not human-versus-machine. This is therefore a consistency check rather than an open question, and is listed last for that reason. If the investigation finds it is not merely a consistency check, that is itself a finding.
Q2, Q3 and Q4 are one problem seen from three sides.
Authority bounded at issuance becomes unbounded through accumulation (more grants over time), chaining (composition across agents), or handoff (delegation onward). In each case every individual act of authorisation was correct and the aggregate is not.
This is probably the spine of the investigation rather than three separate sections. Carried as a working hypothesis, not a settled structure — the fresh session should test it before adopting it, and should record it as rejected if the three turn out to need separate treatment.
Not re-open the parent's six decisions. D1 through D6 are the parent's. Several of the eight questions depend on them — Q6 on D1 especially — and the correct move is to name the dependency, not to re-litigate the decision. Where a question cannot be answered until a parent decision settles, say so and leave it open.
Not produce build items. This is investigation. Build-list absorption follows settlement, and only then.
Not enter the seed. The seed carries commitments. Nothing in this territory is settled.
Not assume a resolution favouring more mechanism. Several of these questions may resolve to declared out of scope, as three did in the parent's Section 10. Naming what DUNIN7 does not do is as valuable as naming what it does.
Read before drafting, in this order. Do not assume the versions named here are the highest present — find the highest.
investigations/loomworks-agent-admissibility-and-attestation-investigation-v0_2 — the parent. Sections 5, 6, 8, 9 and 10 are load-bearing for this arc.candidate-seeds/loomworks/loomworks-candidate-seed-vN_M — canonical seed. v0.12 at time of writing, v0.13 approved for drafting under B-16.current-status/current-status-manifest-vN_M plus any amendment alongside.standing-notes/dunin7-build-list-vN_M — to confirm this arc is not colliding with a live front.
Version-state caveat. Every version number in this handoff is sourced from past-chat retrieval on 2026-08-01, not repo-verified. Record HEAD believed 403790a. Re-read before acting.
Opens after B-29 and B-5 halt. Not before. This arc is not urgent and the two live fronts should not be interrupted for it.
The parent investigation should be committed and readable in the record before this arc opens, since step 1 of the orientation set depends on it.
Paste into a fresh Claude.ai chat when the arc opens.
Post-admission lifecycle investigation, opened as its own arc. You are drafting; I am the Operator.
Orient first, from the record. Do not assume versions — find the highest present:
1. session-handoffs/loomworks-post-admission-lifecycle-investigation-handoff-v0_1 — this arc's opening. Sections 2 and 3 are the territory.
2. investigations/loomworks-agent-admissibility-and-attestation-investigation-v0_2 — the parent. Its six decisions are not yours to re-open.
3. candidate-seeds/loomworks/loomworks-candidate-seed-vN_M
4. current-status/current-status-manifest-vN_M with any amendment
5. standing-notes/dunin7-build-list-vN_M
You cannot reach the record directly. Claude Code retrieves to Downloads and I upload. Expect to ask for that first, and do not draft against project knowledge — it holds stale copies.
The deliverable is one investigation document, HTML primary with Markdown alongside, filed to investigations/. It answers what it can, poses decisions where it cannot, and declares out of scope what DUNIN7 should not claim. It produces no build items.
Test the spine hypothesis in the handoff's section 3 before adopting it. If accumulation, composition and delegation need separate treatment, record the hypothesis as rejected.
Produced 2026-08-01 in the same session as the parent investigation, following the Operator's direction that both the parent amendment and this arc were wanted. The eight questions were decomposed in-chat before this document existed; the spine observation emerged during that decomposition and is the only structural claim carried forward.
No repository was inspected. No Claude Code session ran.
DUNIN7 — Done In Seven LLC — Miami, Florida Loomworks — Post-Admission Lifecycle — Investigation Handoff — v0.1 — 2026-08-01