Reasoning Arm Requirements & Feature List

AI Generated Human Curated

Requirements (draft v0.2, 2026-11-03)

🟡 Status: draft for discussion, nothing settled.

What v0.2 is. v0.1 (unpublished) plus José’s 17 manual edits of 2 Nov (taken as proposals) plus a review of three outside analyses (Gemini, ChatGPT, Meta Muse) checked against the session record and the published pages. Appendix A says what changed and why, line by line.

Markers (as in the session records): Published — already on ai4aging.org · Decided (José) · Proposed (José) — from José’s messages or edits, not decided · Suggested (Claude) — not decided · Open. A new tag, Review, marks a suggestion that came from the outside analyses and was checked here.

The Pepa Reasoning Arm (PRA) has three doors (an OpenAI-compatible completion door, a Pepa-native asynchronous job door, an operator panel) in front of a deterministic shell that holds no operational credentials, with the models behind it in the Mac Studio’s llama-swap slots. v0.2 tightens the contracts around the doors: who may call, how callers prove it, what each state means, what counts as a finished job, what a completion may become, and what the anchor check does and does not prove.

Ground rules carried in

Published or decided; the feature list builds on them. Rows changed by José’s edits or by the review are marked ✎ and explained in Appendix A.

Rule	Where it comes from
The arm suggests; it never decides, orchestrates or executes.	Published
“Never executes” is structural: no operational credentials, no write tools, no route to the operations plane. The Beak enforces it, not a prompt.	Published
The Head owns the decision to plan; the arm never takes a planning request from a peer directly.	Published
Planning is asynchronous and never on the voice path.	Published
✎ One job at a time. Availability is available, maintenance or unavailable; execution is idle, busy or draining. “Dead” is never a value the arm reports, because a dead process cannot report; an outside observer judges it. Published text says “the Head is the queue” — whether the arm gets a small queue at run is Open (see D-3).	Published (idle/busy, one job, Head is the queue, “no answer is its own outcome”); Proposed (José) maintenance and dead, queue at run
✎ A submission carries task, supplied context, constraints, budget, and an authority marker that is always none. Any other value is refused in code before a model sees the request.	Published; Review (refusal in code)
✎ Four typed refusals: not a reasoning task · missing information (naming what to supply) · doctrine forbids it · budget exhausted (partial work attached, marked not a deliverable). “Nothing available can achieve it” is no longer a refusal: it comes back as a done plan marked infeasible, with its reasons.	Published (five); Proposed (José) “reasoning”, strike of “nothing available can achieve it”; Suggested (Claude) the infeasible result
✎ No “can you do this?” pre-check. A help call is provided instead: it returns what the arm can do (the capability card for a verb), never a verdict on a task.	Published (no pre-check); Proposed (José) help equivalent
The requested verb travels apart from the content payload, so nothing in the content can change it.	Published (open questions)
A report is an artifact about sources, never a source. If it enters any store, it enters marked inferred.	Published
Conversation is a separate arm.	Published
Crawl (plain completions to the punt) → walk (one job in, one plan or refusal out, a human reads it) → run (house catalog, ontology, guardrails).	Published
A Pepa arm reflex: completed by the arm on its own through rules declared in advance, consulting neither the Head nor another arm. Arms may narrow it, never widen it.	Decided (José) — 1 Nov, §1.6
✎ The arm’s ledger is the EVIDENCE-class record of what the arm observed and did: written once, owned by the arm, read by the Self-Monitor and Auditor arms. It is not the Architecture Design Ledger.	Decided (José) — 1 Nov, §1.8
Loud failures: surfaced and ledgered, never a silent log line.	Decided (José) — 1 Nov, §6.3
✎ Abstention (“no fact / unclear / looked and found nothing”) is a schema-valid output, enforced by a validator outside the model. “Forbidden” is not abstention; it is the doctrine-forbids refusal.	Published invariant; Review (separation)
✎ Arms return events and/or the requested output; the Head’s transition table decides what happens next. Every run ends done, punted or error.	Head Orchestrator Design, Option C (proposed); Proposed (José) wording
Facts are slots, not prose: evidence-derived values travel as typed slots filled by code after the last model step.	Candidate rule — Head Orchestrator Design
Bilingual by default: English, Spanish, Spanglish.	Published principle

The shape at a glance

Arm-peers (Memory, Knowledge, Human Sentinel…) have no door in v0.2. The Head reaches the completion door too, for plain answers.

Feature list

Out of the original 52 features proposed, José dropped 6 from consideration early but their slot was not reassigned; forty-six features in six groups remain. Δ shows the change from v0.1: new, chg (changed), = (unchanged). Tier when built: 1 enforced in code, 2 declared and enforced by discipline or deployment, 3 designed not built, 4 identified not designed. Today every row is 3 or 4.

ID	Feature	Stage	Marker	Tier	Δ
Doors
RA-01	Completion door: OpenAI-compatible chat completions, synchronous. Serves callers whose output reaches a person, or that the Head sits behind: the operator and the sensory punt now, the Head later. Not arm-peers.	Crawl	Proposed (José) own endpoint and “human”; Suggested (Claude) the reading, compatible on this door only; Open: D-1	1	chg
RA-02	Job door: Pepa-native typed API on the Agent2Agent (A2A) task lifecycle — submit (reply “accepted”, with where to check), status, result, cancel, help; asynchronous, optional long-poll	Walk	Proposed (José) precedent; Suggested (Claude) A2A	1	chg
RA-03	State endpoint: availability (available, maintenance, unavailable) and execution (idle, busy, draining), answered at once. Busy and maintenance are fast answers (503 with reason and retry-after), not refusals. Dead is not a state: no answer at all means dead, judged from outside.	Walk	Published (no answer is its own outcome); Proposed (José) maintenance, dead, fast busy; Suggested (Claude) dimensions	1	chg
RA-04	Capability card, in A2A vocabulary: verbs, models, refusal kinds, limits, versions, an example envelope per verb. The help call returns it.	Walk	Proposed (José) help equivalent; Suggested (Claude)	1	chg
RA-05	Operator tools. Walk: flat files plus a ledger viewer on the command line. Run: the panel (settings, queue, ledger view).	Walk, then Run	Proposed (José) panel; Review (defer the panel)	2	chg
RA-06	Caller registry with caller classes, and caller authentication: a per-caller token (only its hash stored in the arm) plus a source-address allow-list. Unknown or unauthenticated callers are refused before anything runs.	Crawl	Suggested (Claude); Review (authentication)	1	chg
RA-07	llama-swap on :7999 stays the engine room; Homelab tools keep using it directly; Pepa callers use RA-01 and RA-02	Crawl	Suggested (Claude)	2	=
RA-08	Model Context Protocol (MCP) face on the job door for local agents	Later	Suggested (Claude)	4	=
RA-09	Outside liveness probe: a simple external check of the state endpoint and ledger that decides “dead”. The Self-Monitor Arm takes it over once built.	Crawl	Suggested (Claude); Review	2	new
Envelope and results
RA-10	One job envelope for every verb. An authority marker other than none is refused in code. (The marker exists so a request cannot smuggle authority in; the arm itself never holds any.)	Walk	Published (fields); Suggested (Claude) the refusal	1	chg
RA-11	Verb travels apart from content; content wrapped and bannered as data	Walk	Published	1	=
RA-12	Budget plus one propagated deadline. Exhaustion refuses the job with reason budget exhausted; partial work is attached and marked not a deliverable.	Walk	Published (budget, partial work); Head design (one clock); Proposed (José) question on transactions	1	chg
RA-13	Four typed refusals (not a reasoning task, missing information, doctrine forbids, budget exhausted), each telling the caller what to do next. A closed list.	Walk	Published; Proposed (José) reasoning, strike; Suggested (Claude) closed list	1	chg
RA-14	Four job outcomes: done, refused, error, cancelled. Maintenance is availability, not an outcome: a job is never accepted during maintenance.	Walk	Head design (three exits); Proposed (José) four; Suggested (Claude) the four	1	chg
RA-15	Epistemic class stamped on every deliverable. Every claim carries its provenance or an explicit no_source marker. Plain completions carry the unclassified marker instead.	Walk	Published (reports); Proposed (José) acknowledgement, plain endpoint; Suggested (Claude) the markers	1	chg
RA-16	Results handed back, not kept beyond the retention windows in RA-47	Walk	Open on the published page; Suggested (Claude)	2	chg
RA-17	Publication rule: only a validated deliverable is published as done. Work accumulates inside the job; nothing partial is promoted.	Walk	Suggested (Claude); Review	1	new
RA-18	Pending capacity: zero in walk. Busy answers at once (RA-03) and the caller retries. A small bounded FIFO is a run decision (D-3).	Walk	Published (Head is the queue); Review	1	new
RA-19	Dropped by José	N/A	N/A
Services
RA-20	Completion with a guardrail profile chosen by caller class (operator: log only; sensory-punt: conversation profile; head: set per call). The response carries the unclassified marker and the job ID. A completion can never itself establish a Memory or Knowledge fact (RA-44).	Crawl	Proposed (José); Review (marker, prohibition)	1	chg
RA-21	Plan: one plan, or a plan marked infeasible with reasons (the means do not exist), or a typed refusal	Walk, then Run	Published; Suggested (Claude) infeasible	1	chg
RA-22	Analyze: crosscheck, review, compare and extract over material the caller supplies	Walk	Proposed (José); Suggested (Claude) modes	1	=
RA-23	Four validations on every analysis. Anchor (code): the cited span exists verbatim. Source (code): it belongs to the declared document and §. Verdict (reviewer model, then a human): the quote supports the finding. Timing (code plus dates): later evidence has not superseded it. Only the first two are guaranteed by code.	Walk	Suggested (Claude); Review	1	chg
RA-24	Imagine: hypotheticals, scenarios, fixture drafts; stamped hypothetical. Project-only in v0.2.	Walk	Proposed (José); Review (project-only)	1	chg
RA-25	Explain: supplied records rendered for a named audience and language; facts travel as slots filled by code	Walk	Proposed (José)	1	=
RA-26	Research: investigate a question against declared sources, report with provenance	Run	Published (candidate); reach Open	1	=
RA-27	Replay and bench: re-run a finished job with another model or prompt version. Record input hash, prompt version and model configuration so a replay is reproducible.	Walk	Suggested (Claude); Review	2	chg
RA-28	Text-to-speech (pepa-tts, :8100) listed on the capability card and written to the ledger	Crawl	Published (RV3); Suggested (Claude)	2	=
RA-29	Scoreboard: plans and analyses a human accepted, per verb and model. Deferred past the first two services.	Run	Published (plans); Suggested (Claude); Review (defer)	2	chg
Inside
RA-30	Deterministic shell, model core: intake, authentication, checks, state, budget, refusals, validation and ledger are code	Walk	Suggested (Claude)	1	=
RA-31	Fixed per-verb pipelines in walk; an agent loop only for planning with read-only tools at run	Walk, then Run	Suggested (Claude)	3	=
RA-32	Structure-first chunking by headings and § numbers; original offsets and section IDs preserved through every step	Walk	Suggested (Claude); Review (offsets)	1	chg
RA-33	Reviewer model in a different slot as a check. Disagreement is a result, kept as two attributed claims; agreement is never proof. Two models may not fit beside the large slot at once: measure, and fall back to the small slot or a swap.	Walk	Suggested (Claude); Review	1	chg
RA-34	Schema-validated outputs, abstention included; validator outside the model	Walk	Published invariant	1	=
RA-35	Frozen, versioned prompt prefix per verb. At run: frozen prefix, then the stable catalog (byte-identical, sorted by entity ID), then per-job content, so the prompt cache holds.	Walk, Run	Published (frozen prefixes); Suggested (Claude); Review (layout)	1	chg
RA-36	Planted-error test fixtures for analyze: wrong quotes, unsupported claims, contradictions, superseded evidence	Walk	Suggested (Claude); Review	2	new
RA-37	Dropped by José	N/A	N/A
RA-38	Dropped by José	N/A	N/A
RA-39	Dropped by José	N/A	N/A
Safety and record
RA-40	No operational credentials, no write tools, no route to the operations plane. Enforced by deployment: the Beak between VLANs, and a filter on the Studio itself (pf or llama-swap’s own allow-list) for hosts on the Studio’s own subnet, which the Beak never sees.	Walk	Published; Suggested (Claude) the Studio filter	2	chg
RA-41	The arm’s ledger: JSON Lines, append-only, UUIDv7 job keys, correlation IDs, content hashes (not content), processing chain and versions. Field names follow the OpenTelemetry GenAI conventions proposed in the 3 Oct observability recon.	Crawl	Decided (José) the ledger; Suggested (Claude) the fields; Review	1	chg
RA-42	Loud failures to the caller, the operator tools and the ledger; household people only through the Sensory Arm	Walk	Decided (José); Suggested (Claude) routes	1	=
RA-43	The arm’s only reflexes: refusals, budget kills, drain under memory pressure — nothing outside the arm	Walk	Suggested (Claude) under the Decided reflex base class	1	=
RA-44	An interface prohibition, not a policy line: nothing the arm produces — completions included — enters Memory or Knowledge except marked inferred or hypothetical, by another arm’s decision	Walk	Published (reports); Suggested (Claude); Review	2	chg
RA-45	Settings in flat files under version control; every change is a ledgered event, kept distinct from job events; fail-closed defaults shown and explained	Walk	Suggested (Claude); Review	1	chg
RA-46	Health read from outside: the Self-Monitor Arm (and the RA-09 probe until then) reads the state endpoint and ledger, never the arm’s own account	Walk	Suggested (Claude)	2	=
RA-47	Three retentions: the ledger is append-only and permanent; job results are kept until collected or 7 days; completions are returned and not kept; supplied content has the shortest retention and the ledger stores only its hash	Walk	Suggested (Claude); Review	2	new
RA-48	Dropped by José	N/A	N/A
RA-49	Dropped by José	N/A	N/A
Host
RA-50	Shell in an LXC on its own VLAN; the Mac Studio stays a pure inference server. Host (pve2 or pve3) is decided when resource needs and network rules are known.	Walk	Suggested (Claude); Open: D-4	2	chg
RA-51	Repository prsws/pepa-reasoning-arm — simple, not clever	Walk	Suggested (Claude)	2	=
RA-52	Completion never waits behind a job. Either the punt has a model slot the job pipelines do not use, or the pipelines’ many small calls leave room for it. Measure on the Studio before choosing.	Crawl, Walk	Suggested (Claude)	2	new

Interfaces

Its own endpoints; OpenAI compatibility only on the completion door. Everything new speaks a small Pepa-native job API on the A2A task lifecycle (submitted, working, completed, failed, cancelled, rejected), so a typed envelope, typed refusals and an asynchronous lifecycle do not have to bend into the chat shape.

Door	Who uses it	Shape	Notes
Completion (RA-01)	Operator, sensory punt, later the Head	/v1/chat/completions, /v1/models, synchronous	The punt moves by changing a URL and adding its token. Every call is ledgered; the reply carries the unclassified marker.
Job (RA-02)	Operator now; the Head when it exists	Submit → “accepted, check here” → status → result; cancel; help	Busy and maintenance answer at once.
State and card (RA-03, RA-04)	Callers, the probe, the Self-Monitor Arm	Two read-only endpoints	No work submitted.
Operator tools (RA-05)	José	Flat files and a ledger viewer; the panel at run	Operator-only, house network only.

Caller classes (RA-06).

Class	Example	Doors	Verbs
operator	José	Completion, job	All
head	The LangGraph Head, once built	Completion, job	All
sensory-punt	PSA’s call_external_llm	Completion only	Completion

Note: Arm-peer is removed in v0.2. Another arm that needs reasoning goes through the Head, as published.

What stays outside. llama-swap on :7999 remains the Studio’s engine room for Homelab tools. Once the punt has moved to the completion door, Pepa hosts are blocked from :7999 by two mechanisms, because they live on two different networks: the Beak for hosts on other VLANs, and a filter on the Studio itself for hosts on its own subnet (RA-40).

Naming. Caller class replaces “authority level” for the guardrail dial. The record now uses “authority” in three senses (a mandate to act, the EVIDENCE authority class of a record, and the envelope marker that is always none), so the glossary should fix all three.

State model (RA-03, RA-14)

Dimension	Values	Who reports it
Availability	available · maintenance · unavailable	The arm, at once
Execution	idle · busy · draining	The arm, at once
Job outcome	done · refused · error · cancelled	The arm, on the job
Dead	—	Never the arm. An outside probe judges it from no answer or a stale ledger.

Degraded (a slot missing, the reviewer unavailable) is shown as extra health detail on availability, not a fourth axis. Superseded is a Head concept and waits for the Head.

Services

Verb	Goes in	Comes out	Epistemic class	Good first job
completion	Messages from a registered caller	One answer	Unclassified; never storable as fact	The Sensory Arm’s punt, as today
plan	Task, context, constraints, budget	A plan, or a plan marked infeasible, or a refusal	Proposal	Sequencing the PSA re-adoption work
analyze	A question or rubric plus the material	Findings, each with verdict, anchor and four validations	Interpretation, anchored	Cross-check the 2 Nov session record’s Summary against its Running record
imagine	A premise and a frame	Scenarios, variants, drafts	Hypothetical	What-if walks through the eight use-case EPCs (project only)
explain	Records, audience, language, level	An explanation whose facts all come from the records	Interpretation; facts as slots	A pepa_verify failure event in plain Spanish
research	A question and permitted sources	A report with provenance per claim	Interpretation about sources	Reading-list triage

Analyze, in detail. Modes: crosscheck (each claim gets one verdict: supported, contradicted, partly supported, not found, stale), review, compare, extract (marked inferred). The 2 Nov failure is a harness problem: a small model given a long page in one pass loses the middle. The pipeline splits by structure, extracts one claim per Summary bullet, finds candidate entries by § and keyword in code, asks the model for a verdict and a quoted span, then checks. The validation ladder (RA-23) is honest about its limits: a real quote can still be misread, so the anchor check proves the quote exists and the source check proves where it is from, while the verdict and timing checks still need the reviewer and the reader. A verdict whose quote is not found becomes unanchored.

Explain keeps fail-closed slots. Values arrive as slots filled by code after the model writes, and the output fails if the model restates a value. A validator that tried to tolerate restatement would miss values written out in words (“veinte”, “las dos y media”), and the Head design’s stricter rule is the 68 °F lesson.

Job envelope and result

Envelope field	Holds	Marker
job_id	UUIDv7 set by the caller (idempotency)	Suggested (Claude)
caller_id	Registered caller. The proof is the token in the request header, not a field.	Suggested (Claude); Review
verb, mode	The verb is a field of its own, never inferred; mode for sub-kinds	Published; Suggested (Claude)
task, content[], constraints	Instruction; material with source and provenance, wrapped as data; limits	Published
budget, deadline	Done condition and one propagated clock	Published; Head design
model, language, audience	Preference from the class’s allow-list; for explain and refusal text	Published; Suggested (Claude)
authority	Always none; anything else refused in code	Published; Suggested (Claude)

Result field	Holds
status	done · refused · error · cancelled
refusal	One of four kinds, plus what to supply next when information is missing
deliverable	Typed by verb; for a plan, possibly marked infeasible
epistemic_class	proposal · interpretation · hypothetical · unclassified
claims[]	Each claim with provenance or no_source; for analyze, the anchor, verdict and four validation results
assumptions, uncertainties, invalidators	Confidence about the result in words, not a self-rated score
partial_work	Only on a budget-exhausted refusal; marked not a deliverable
processing_chain, versions, input_hash	Model, quantization, slot, prompt version, shell version, hash of the input
budget_used, timings	What the job cost

Inside the arm

A deterministic shell around a model core. Plain Python for the shell; fixed pipelines per verb in walk; LangGraph only where a verb needs a loop (planning with read-only tools at run). Models are checks, not routers: the caller declares the verb, so there is nothing to classify, and inferring intent at this layer runs into the no-inferred-intent directive. A reviewer in a different slot, preferably a different model family, checks the first model’s findings (RA-33). Every model output is validated against its schema; after a capped number of retries the job ends in error, and free text is never passed through. Whether rapid-mlx’s constrained output is good enough to replace validate-and-retry is to be measured on the Studio.

Where it runs

Leaning, unchanged: the shell in an LXC on a Proxmox host in the Pepa stack, on its own VLAN; models stay behind llama-swap on the Studio. The deciding reason is that it turns “never executes” into a firewall rule. Bare on the Studio puts the arm on the same subnet as Home Assistant, so the rule would rest on discipline; a VM on the Studio takes memory from the model slots on a 64 GB machine. The cost of the LXC is a second box that can take the arm down, which is acceptable because it is off the voice path and fails loudly. The host (pve2 or pve3) waits for resource needs and network rules.

Operator tools and settings

Walk: policy in flat files under version control, a command-line ledger viewer, and a kill switch (drain, then stop). Run: the panel (state and queue, engine room, callers, services, guardrail profiles, retention, messages, scoreboard, ledger). Either way, every change is a ledgered event with old value, new value, who and when, kept apart from job events. Defaults fail closed and say why.

Build order (Review)

1. Shell, caller authentication, state endpoint, ledger (RA-30, RA-06, RA-03, RA-41).
2. Completion door; move the punt and confirm PSA’s calling contract is untouched (RA-01, RA-20, RA-52).
3. Close :7999 to Pepa hosts with both mechanisms; start the outside probe (RA-40, RA-09).
4. Job door with one verb: analyze/crosscheck on the 2 Nov session record, with planted-error fixtures (RA-02, RA-22, RA-23, RA-32, RA-36).
5. Reviewer slot and replay bench (RA-33, RA-27).

Deferred until two services have proved their contracts: the panel, MCP, an internal queue, LangGraph, the scoreboard, research beyond supplied sources.

Pending Decisions

#	Question	Suggested answer	Why
D-1	What does “human callers” in RA-01 mean?	Callers whose output reaches a person or sits behind the Head; no arm-peers	Keeps the punt working and stops arms chaining unlogged reasoning through a chat endpoint
D-2	Operator as interim submitter	Yes, recorded as a dated development exception	Honest about the published rule instead of quietly reinterpreting it
D-3	Queue at run: the Head (published) or a small bounded queue in the arm (José’s edit)	Keep the Head as queue unless a concrete need appears	Priority and supersession live where authority lives; changing it needs a dated entry on the page
D-4	Where the shell runs	LXC on its own VLAN; host later	The firewall rule is the enforcement
D-5	Close :7999 to Pepa hosts	Yes, after the punt moves, with both mechanisms	One mechanism leaves a gap on the Studio’s own subnet
D-6	Auditor Arm and analyze	No use of this arm’s models by the Auditor for now	Shared models share failures
D-7	Imagine’s reach	Project only	Household-facing storytelling edges into the Conversation Arm
D-8	LangGraph	Plain Python now; at run only for the planning loop	Same library as the Head, so the lease rule can carry over
D-9	Retention	RA-47	Supplied content may include health talk (D-11)
D-10	First build slice	Build order above	Each step is testable alone and none breaks the punt
D-11	“Nothing available can achieve it” as a plan result instead of a refusal	Yes	It comes from a real attempt and deserves provenance. Your strike suggests this was the intent: please confirm.
D-12	Publishing the changes	The four-refusal list, the state model and any queue change need a dated entry on the Reasoning Arm page	They change published text

Appendix A — what changed from v0.1

Source	Item	Change in v0.2
José’s edit 1	States; “dead”; caller-is-queue struck	State model (three dimensions); dead judged from outside; published “Head is the queue” restored in the text and flagged as Open (D-3)
José’s edit 2	Authority marker struck	Kept: authority: none, other values refused in code
José’s edit 3	“Reasoning”; “nothing available” struck; ”???” on budget	Adopted “reasoning”; the strike read as a plan result (D-11); budget exhaustion refuses with partial work not promoted
José’s edit 4	“A -help equivalent”	RA-02/RA-04: help call returns the capability card
José’s edit 5	“Job queue” at run	Open (D-3), zero pending in walk (RA-18)
José’s edit 6	“What is EVIDENCE-class record?”	Answered in the ground rule, from §1.8
José’s edit 7	Abstention ”/ forbidden”	Abstention kept epistemic; “forbidden” routed to the doctrine refusal
José’s edit 8	“events and/or the requested output”	Kept
José’s edit 9	RA-01 “human callers”	Read as D-1; arm-peer class removed
José’s edit 10	RA-02 “find precedent/best practices”	A2A task lifecycle
José’s edit 11	RA-03 maintenance, dead, fast busy	RA-03
José’s edit 12	RA-10 “(explain)”	Explained in RA-10
José’s edit 13	RA-12 “debatable… transactional”	RA-12, RA-17: only validated deliverables are published
José’s edit 14	RA-13 five struck, “as identified”	Closed list of four (RA-13)
José’s edit 15	RA-14 four states, maintenance?	Four outcomes; maintenance is availability
José’s edit 16	RA-15 acknowledgement of missing citations; plain endpoint	no_source marker; unclassified marker on completions
José’s edit 17	RA-20 “(explain)”	Explained in RA-20
Review	Caller identity vs authentication	RA-06
Review	Completion as a governance bypass	RA-15, RA-20, RA-44
Review	Anchor check overclaims	RA-23 validation ladder
Review	Queue semantics	RA-18, D-3
Review	Transactional publication	RA-17
Review	Ledger and retention	RA-41, RA-47
Review	Reviewer memory limit	RA-33
Review	Prompt cache layout	RA-35
Review	Planted-error fixtures	RA-36
Review	Same-subnet bypass of the Beak	RA-40
Review	Completion starved by jobs	RA-52
Review	OTel field names	RA-41
Declined	Explain restating values with after-the-fact validation	Kept fail-closed slots
Declined	Meta Muse: revert “events and/or the requested output”, keep the authority edits softened, keep all five refusals	Kept your wording on the first; kept authority: none; adopted a closed list of four with infeasible as a plan result
Sent elsewhere	ha_control allow-list fails open; verify lost on a Home Assistant restart; service-channel confirmation	Sensory Arm backlog (already in the 2 Nov record)
Rejected	Gemini’s build plan (“asynchronous completions”; planning bound to Head transitions)	Contradicts the synchronous completion door and Option C

Abbreviations

For more Pepa terms refer to the Pepa Glossary

Short	Spelled out
A2A	Agent2Agent protocol — vocabulary and task lifecycle, not its transport
EPC	Event-driven process chain
FIFO	First in, first out
JSONL	JSON Lines
LXC	Linux container, as run on Proxmox
MCP	Model Context Protocol
OTel	OpenTelemetry
PRA, PSA	Pepa Reasoning Arm, Pepa Sensory Arm
RA-nn	(R)easoning (A)rm feature IDs
RV3, RV4	Reasoning Arm versions in the PRA Version History
VLAN	Virtual local area network