CLOSED 2026-09-28 by Travis's ruling, recorded as `v2/handoff/DECISION_LOG.md` D-078: there is no Graphonomous v3. This proposal was tested and not justified. Graphonomous continues as v2. Do not build from anything below. The closure is §C, directly under this paragraph. Everything after §C is the draft as it stood when it closed (written 2026-09-27, spec rung), kept as the record of what was proposed and why it failed.
Two predeclared benches decided it. Both were negative for the new machinery, not for Graphonomous.
| V3-1 (a) — recall | V3-1 (b) — relation judgment | |
|---|---|---|
| predeclared (commit, before any run) | 0dd0f3ed | 0c193468; Amendment 1 9527e639 |
| frozen input (sha256, measured 2026-09-28) | corpus manifest 1e98b38c…, 147 notes; queries b930dcd8…, 23 | dataset eb888705… (freeze commit 2936de0e); cost matrix b1455f54… |
| result | 4a6b8bcc, v3-bench/RESULTS.md; results.json ac989a6a… | 6be1c799, v3-bench/b/RESULTS.md; results.json cca9077a… |
| finding | BM25 over the folder was best (nDCG@5 0.782). Jev reranking did not help (−0.020 against its own BM25, lost 14 of 23) | governed Jev: 6 false `supersedes` where 0 was required, three at p 0.86 / 0.86 / 0.70, above the 0.78 median of its correct ones, so no threshold separates them. The Claude judge was better on hard negatives but inconsistent and cautious on supersession |
Private inputs and outputs live in ~/.graphonomous/v3-bench/. This repository is public.
The ruling (D-078), in short:
Graphonomous keeps v2's definition: a derived semantic graph over sources it does not own.
v2 gains the memory folder and the ruling register as sources. It keeps its deterministic, governed
relation mechanics, uses authored relations where a source states them, and abstains or queues for review otherwise.
BM25 stays outside the semantic core. The tool is scripts/memory-search.py in the root repository.
Not added:
a Graphonomous-owned memory store;
this spec's daemon (§4.8 A1), for recall;
its journal (§4.2);
model-mediated recall;
automatic Jev relation proposals.
A new major version needs a new invariant or execution model, not accumulated features.
One lane stays open, non-authoritative: candidate → model suggestion → human review. It measures
review load under a new predeclaration, and never tunes on V3-1 (b)'s test split.
Where each open item went:
| Item | Disposition |
|---|---|
| T1 name | ruled: there is no Graphonomous v3 |
| T2 where the decider runs | moot: no decider sits in the trusted path |
| T3 install Ollaya | moot now; a decider re-enters only through a new predeclared experiment |
| T4 catalogue gates | moot: there is no question catalogue |
| T5 confidentiality | carried into v2: both new source families are private, and the repository is public (v2/handoff/ROADMAP.md §3) |
| T6 profile family | moot: no memory profile family. New roles or kinds go through v2's semantic.vN profiles |
| T7 embeddings | moot: recall is BM25 |
| T8 the file memory's role | ruled: the filesystem remains the durable store, and v2 derives from it |
| §4.5 floor F-2 (secret refusal) | carried into v2: the memory-folder adapter's intake (ROADMAP §2) |
| §4.5 F-3/F-4 (rulings outrank notes) | carried into v2 as a candidate consistency rule, a FINDING (ROADMAP §2) |
| §6 L-3…L-6 (v1 retirement) | still open, unaffected: they retire v1, not v3 |
| §7 V3-2…V3-6 | not built, by ruling |
| §7 V3-7 (Fly app) | proceeds under R96, independent of this spec |
The surviving requirements are in `v2/handoff/ROADMAP.md`, each under its owner and each PROPOSED.
The draft as it stood at closure follows. It is history, not direction.
Direction spec · draft · `spec` rung · written 2026-09-27 · nothing below is ruled. Do not build from this yet. Every [T#] is an open ruling for Travis (§9). A number marked measured was taken on this machine on 2026-09-27 with the command or file named beside it. A number marked reported comes from a web research pass the same day (§10) and was not reproduced here. The two are never blended.
Name trap — read before quoting this document. "v3" is already used twice in this repository:
graphonomous.semantic.v3is a profile id proposed by v2's R17 (the WRL vocabulary), and v2's stage roadmap numbers its stages G0–G5 (G3 = reduction). This document is about neither. Write Graphonomous v3 in full — never "v3" or "G3" alone — until [T1] settles the name.
Graphonomous v1 (v0.4.3, Elixir, graphonomous/lib/) tried to be an agent's memory, and it deletes what it stores. Graphonomous v2 (graphonomous/v2/, Node ESM) is a certified evidence graph over pinned registries and has, deliberately, no write surface. Super, the MVP (R97), holds an unbuilt slot for "the memory each agent carries". Graphonomous v3 is the proposal to fill that slot: the memory substrate of an active locus, built on v2's projection and certificate machinery, fed by rulings, Super's receipts and proposed observations, and judged — never authored — by a calibrated typed decider (TypeSafe's Jev, or Ollaya, its free local counterpart). v2 is not modified: Graphonomous v3 adds a journal, profiles and adapters beside it. v1 becomes legacy (§6).
| Fact | Value | How |
|---|---|---|
nodes in ~/.graphonomous/knowledge.db | 0 | sqlite3 -readonly … 'select count(*) from nodes' |
| edges / goals / outcomes surviving | 620 (dangling) / 173 / 124 | same; pragma quick_check = ok |
| last write to the database | 2026-09-22 23:49 | ls -l on knowledge.db* |
| consolidator cadence | every 300,000 ms | config/config.exs consolidator_interval_ms |
| decay per run / hard-delete below | 0.02 / 0.1 | lib/graphonomous/consolidator.ex @default_decay_rate, @default_prune_threshold |
| lifetime of a node stored at confidence 0.5 | ≈ 80 runs ≈ 6.7 h | 0.5 × 0.98ⁿ < 0.1 ⇒ n ≥ 80 |
| per-session servers running before §6 | 29 BEAMs, ≈ 1.6 GB RSS | ps; the plugin spawns scripts/graphonomous_mcp_wrapper.sh over stdio, one per open session |
Each of those servers loaded the graph into memory once at start (warm_cache_from_db, store.ex), so a session never saw another session's writes, and each ran its own consolidator against the same file. "stored but retrieve returns 0" was not a crash: it was a forgetting curve tuned for benchmark churn, running as 29 independent copies of itself.
Its definition: *"the persistent semantic/evidence graph that makes what a system understands queryable,
falsifiable, provenance-linked, and reducible", and "a derived semantic projection over authoritative registries … not another authoritative registry"* (v2/README.md).
"No LLM output is evidence. Deterministic, content-addressed, replayable transforms only."
(v2/README.md:38).
D-071 (2026-09-03): "The next step is NOT a write surface."
Projections are rebuilt whole from a pinned snapshot; lib/project.mjs deletes the output directory
first. There is no append path.
The query surface is six structural functions (lib/query.mjs: node · neighbors · path · facts ·
explain · as_of); "Nothing here computes a fact the projection and its evaluation do not already hold." There is no relevance retrieval, no embedding, no MCP server.
Latest decision D-077; last commit 61adec2d (2026-09-08); full tree 182 pass / 0 fail / 1 skip.
So "move the MCP onto v2" is a category error as v2 stands: v1 answered what do we remember?, v2 answers what does the stack's evidence establish?
cockpit/src/claude_connection.rs:511,614 starts Claude with --strict-mcp-config and
{"mcpServers":{}}; cockpit/src/codex_connection.rs disables Codex's memories. On purpose.
site/AGENT_SUPER_APP_BLUEPRINT.md:274: "MCP is adapter/transport (tools become requested *typed
capabilities …); neither implicitly grants authority."* Phase C4 is "MCP + Agent Skills adapters".
site/index.html:729: Super is *"the slot each agent's memory lives in … the loop where Super admits what
an agent did as receipts and Graphonomous re-imports it is designed"*.
FOUNDATION_LANE.md (≈ line 1839): Graphonomous's designed stages *"have no executor, and the executor
must be a Super Worker under Carrier confinement whose only writable surface is a proposal"*.
ampd already has a capability registry (ampd/lib/ampd/capability_registry.ex: postgres, worktree,
trvm, github packs), a receipt ledger (receipts.ex: "feeds Evidence; never a second audit system"), and — on the t23-bounded-attempts / authority-log* branches — an append-only, gapless AuthorityLog.
The file memory at ~/.claude/projects/-home-travis-ProjectAmp2/memory/: 148 files, 34,886 lines, indexed by MEMORY.md. Not in git, not certified, not shared with Codex or Super — and it works. Any Graphonomous v3 has to beat it at its own job, or not be built (§8 F1).
Jev (TypeSafe AI, announced 2026-09-15) and Ollaya (Apache-2.0, released ≈ 2026-09-24, wire-compatible with Jev on POST /v1/systemone) answer typed questions about a state — noul (P(true)), choice (2–255 labels, a probability each), score (2–10 levels) — with calibrated probabilities and no text and no embeddings. They judge candidates; they cannot find them. Those are exactly the judgments v1 made with hard-coded constants — and a hard-coded constant (§1.1) is what emptied it. Details and sources: §10.
| J1 — carried memory | J2 — evidence self-model | |
|---|---|---|
| Answers | what has this locus learned, been told, been corrected on? | what does the stack's evidence establish? |
| Authority | proposals, rulings, receipts | authoritative registries only |
| Carried today by | the file memory (§1.4) | Graphonomous v2 |
| Write surface | yes — proposals | no (D-071) |
| Graphonomous v3 | is J1, built on J2's machinery | unchanged; v3 links to it |
P1 — No model output is evidence (inherits v2/README.md:38). A decider's answer is a DECISION
record carrying its certificate: a ranking or a proposal, never a fact.
P2 — Floor, then gradient (box-and-box, feasible ▸ permitted ▸ best). Deterministic rules refuse
(§4.5); the decider only orders what the floor admitted. A decider never sits in permitted.
P3 — No second ledger. Facts about Super's world enter from Super's receipts. Graphonomous v3 reads
downstream and never writes into ampd's stores.
P4 — Nothing is forgotten by a constant. Retirement is a decision with a trigger and a record; it
archives. Erasure is a separate, explicit, audited command. (v1's lesson.)
P5 — One process owns the store. Every client is an adapter; no per-session engines. (v1's lesson.)
P6 — Rulings outrank observations. A REVISION_REGISTER.md row is authoritative; an observation
that contradicts a current ruling is admitted only as a flagged conflict.
P7 — Measured or cited. Every number has a command or a source; measured and reported stay apart.
P8 — Local and free first. Zero budget. Hosted Jev is a metered reference and fallback, not a
dependency.
P9 — Memory is per locus. AGENCY.md: "memory is a substrate; lineage is a relation." A Bot or
Worker recalls through its projection, not the world.
sources ──► journal ──► projection ──► recall ──► floor ──► judge ──► access
RULINGS append- v2 machinery lexical + deterministic calibrated MCP daemon (outside Super)
RECEIPTS only, (lids, canon, structural refusals decider capability pack (inside Super)
EVIDENCE gapless certificates, (+ embed?) (§4.5) (§4.6) CLI
OBSERVATIONS jseq WRL seal) ≤255 cands DECISION
FILE MEMORY certificates| Family | What | Authority | Status | |
|---|---|---|---|---|
RULINGS | REVISION_REGISTER.md rows (99 rows matching `^\ | R[0-9]`, measured), pinned by commit — tracked in the ProjectAmp2 repo | authoritative | new adapter |
RECEIPTS | Super receipts, AuthorityLog, witness JSONL (effects/witness.ex) | authoritative for Super's world | waits on Super stage 0 | |
EVIDENCE | v2's existing families (crosswalk, factory, factory_mosaic, trvm) | authoritative | exists; linked, not changed | |
OBSERVATIONS | a locus's "remember this": text, author locus, session, cites | proposal | new; enters only through propose | |
FILE MEMORY | the §1.4 directory, imported as OBSERVATIONS authored by claude-code-sessions | proposal | new adapter; not in git, so pinned by a content hash of the directory |
v2 can only rebuild. A memory has to append. Proposal: an append-only, framed, gapless-sequenced journal (the shape of Super's AuthorityLog: one record per admitted transition, jseq without holes). Projections become folds over journal ranges.
Law J-1: `fold(journal[0..n])` is byte-identical to `rebuild(snapshot at n)`. The full rebuild stays the verifier; the incremental fold is only the fast path. The journal is the only thing ever written; everything else is derived.
OBSERVATION · RULING · DECISION (one decider call: model id + weights digest, calibration version, question id@version, state hash, probabilities, confidence, latency, host) · RETIREMENT (what was archived, its trigger, and the rule or DECISION that caused it). Relations: SUPERSEDES (ruling→ruling, observation→observation), CONTRADICTS (a flag, never a verdict), CITES, CORROBORATED_BY (observation → receipt | evidence | ruling). Whether this is a graphonomous.semantic.vN profile or a separate graphonomous.memory.v0 family is [T6].
A decider cannot find anything; it only judges what it is shown. Recall is lexical (full-text over OBSERVATION and RULING text), plus structural (v2's neighbors / path), plus embeddings if [T7] says so. v1's fallback hash embedder is not semantic recall and is not carried over. Recall returns at most 255 candidates, the decider's choice bound.
F-1 An OBSERVATION without an author locus and a session is refused.
F-2 Secret-shaped text (API keys, passwords, tokens) is refused at intake. Local files on this
machine are known to hold plaintext credentials, and no import may carry them. (A file path stood here until 2026-09-28. It was removed because this repository is public.)
F-3 A RULING is superseded only by a later RULING, never by an observation.
F-4 An observation that contradicts a current ruling is admitted flagged CONTRADICTS, never as
current.
F-5 Nothing is deleted by the engine. Retirement archives; erasure is its own audited command.
F-6 A DECISION with an incomplete certificate is refused: absent ⇒ 0̲, fail closed (AGENCY.md,
certificate).
F-7 Proposals are metered per locus. Alphapiler measured it (Piles of Dust, PASS 9): when refusals
cost nothing, a blind picker asking often beats an informed one asking rarely.
| Id | Type | Question | Answer goes to | Replaces in v1 |
|---|---|---|---|---|
| Q-KEEP | noul | Is this a durable fact worth carrying beyond this session? | review order at intake | store-everything |
| Q-ASSERT | noul | Is it stated as settled (vs a plan, guess, question)? | a tag | — |
| Q-SAME | score ×3 | different / review / same, for a candidate pair | merge proposal | consolidator_merge_similarity 0.95 |
| Q-CONTRA | noul | Do these two contradict? | CONTRADICTS candidate | — (v2 reduced its two conflict rules to zero, D-074, restated in D-076) |
| Q-SUPERSEDE | choice | a supersedes b / b supersedes a / both stand / review | SUPERSEDES proposal | — |
| Q-RANK | choice | which of these ≤255 candidates answers the query? | recall order | similarity score |
| Q-ENOUGH | noul | Is this enough to act? | act / learn / escalate hint — second opinion to κ routing, never an override | coverage heuristic |
| Q-RETIRE | choice | keep / archive / review — asked only on a trigger (superseded, contradicted, source moved), never on a timer | RETIREMENT proposal | decay 0.02 / prune 0.1 |
Probabilities are stored, not thresholded at write time; the threshold that acts is part of the question's version. The middle band goes to a review queue (Travis, or a stronger model), never to an automatic merge. Calibration is refitted per question type on our own labels: reported, calibration drifts in a different direction per question type out of distribution (§10).
Option A — inside the Graphonomous v3 daemon (plain HTTP to a local Ollaya). Latency is reported at
89–520 ms on an RTX 4090. That answers the first reason behind the standing rule "MCP servers make no LLM calls" (calls of 60–130 s timed out). Every call is a DECISION record, which answers the second (visibility). It still contradicts the rule as worded.
Option B — agent-asked. The daemon returns typed questions; the agent asks the decider and submits the
answers. The rule stays intact, but the daemon cannot establish that an answer came from the named model: a submitted certificate is only a claim.
Option C — a Super capability. Inside Super the decider is a capability pack, and each call is an
effect through Gateway whose receipt is the certificate. The most Super-native option; waits on Super.
Recommendation: C inside Super; A outside it if [T2] permits; B only as the interim.
A1 — one MCP daemon for Claude Code and Codex sessions outside Super. A systemd user unit, like
super-desktop, serving HTTP on 127.0.0.1. The plugin points at its URL: no stdio spawn per session (P5). Tools: recall, explain, status, then propose; read-only first. Reported: the MCP spec of 2026-07-28 made the protocol core stateless (no session id), so a continuation is an explicit handle returned by a tool.
A2 — a `memory` capability pack inside Super. Reads through the agent and bot projections
(ampd/lib/ampd/projection.ex); proposals go through Gateway.perform and come back as receipts. Super's own Claude and Codex hosts stay --strict-mcp-config.
A3 — CLI. recall / explain over the memory profiles, beside v2's bin/g0.mjs query.
30 GB RAM under chronic pressure (the desktop app has been OOM-killed at 14–23 GB resident); an RTX 4060 Laptop GPU with 8,188 MiB; glibc 2.44 (Ollaya reportedly needs ≥ 2.38). v2 reads whole JSONL files into memory (lib/evaluation.mjs), and its projections/ hold 296 MB on disk. Before any promise: measure the daemon's idle and peak RSS, and each candidate decider model's resident size. Candidates: decider:2b or winnow:e4b; laya only for short pairs (reported 421M, 512-token context). [T3]
| From | Keep | Retire |
|---|---|---|
| v1 | the five-phase loop vocabulary as a PULSE manifest (retrieve · route · act · learn · consolidate); κ topology as a floor-side signal; the goal concept; the surviving rows (173 goals, 124 outcomes, 419 entities, 620 edges) as an archived source | decay and prune constants; per-session stdio engines; the hash-fallback embedder as recall; the contract "stored ⇒ remembered" |
| v2 | everything, unchanged: lids, canonical bytes, certificates, WRL sealing, pins, G1 rulepacks, query surface | nothing; the journal, profiles and adapters are added beside it |
v2's own pending graphonomous.semantic.v3 (R17, the WRL vocabulary, awaiting GPT's adjudication) is independent of this document and is not blocked by it.
Done 2026-09-27:
The plugin is disabled in both scopes: graphonomous@ampersand-plugins: false in ~/.claude/settings.json
and in ProjectAmp2/.claude/settings.local.json. This takes effect for new sessions and also removes the plugin's skills and its PostToolUse hook, which appended to /tmp/graphonomous-changed-files.jsonl.
The 29 running per-session servers were stopped with SIGTERM, by PID. free -m available went from
5.4 GB to 8.4 GB (other processes were not held still).
No row changed: quick_check ok; nodes 0, edges 620, goals 173, outcomes 124. The file itself did
change: as the servers shut down, SQLite checkpointed the write-ahead log into it (mtime 2026-09-27 20:21, WAL now 0 bytes).
The retired policy was replaced, at Travis's ask: graphonomous/CLAUDE.md (Primary Directive and §1–12;
the old text is git show 61adec2d:CLAUDE.md) and the root CLAUDE.md section "Graphonomous session policy". Both now say v1 is LEGACY, the loop is off, and point here.
Done 2026-09-28:
L-1 — archived at `~/.graphonomous/export-2026-09-28/` (mode 0700): a .backup copy of
knowledge.db (integrity ok, every table's count equal to the source), knowledge.db.fts, schema.sql, one JSONL per non-empty table (edges 620, entities 419, goals 173, outcomes 124, revisions 12, schema_migrations 9, each row count checked), and MANIFEST.json with the sha256 of every file. A secret-pattern scan of the JSONL found nothing. It is outside the repository on purpose: c-u-l8er/graphonomous is public, these rows are private session memory, and [T5] is unruled.
L-2 — the LEGACY banner is on `graphonomous/README.md`. It gives the measured reason and the two
flags that stop the pruning (--consolidator-decay-rate 0.0 --consolidator-prune-threshold 0.0, both accepted by lib/graphonomous/cli.ex). graphonomous.com needed nothing: its 2026-09-05 rebuild already retracts the v1 engine, and the old install command appears only inside that retraction.
Not done — each needs [TRAVIS]:
L-3 npm deprecate graphonomous — public, so it is Travis's to do.
L-4 The Fly app graphonomous-mcp stays until fleetprompt.com and PRISM stop calling it (R96: torn
down last).
L-5 The PRISM plugin runs the same per-session pattern; not measured here and not touched.
L-6 The plugin's description in ampersand-plugins still says "v0.3.2 … 29 MCP tools"; the server
exposes 5 and reports itself as 0.4.0 (hard-coded in lib/graphonomous/mcp/machines/server.ex).
Closed 2026-09-28: V3-1 ran; V3-2…V3-6 are not built, by ruling (§C).
V3-0 This spec, and the rulings T1–T8.
V3-1 Offline bench, with no infrastructure.
(a) Recall: real questions taken from session transcripts (the first retrievals sessions actually
made), each with its expected memory file or ruling id. The baseline is MEMORY.md plus grep.
(b) Deciders: labeled pairs for Q-SAME, Q-CONTRA and Q-SUPERSEDE. Candidate label sources, counted by
grep over the memory directory on 2026-09-27: "WRONG" 74, "FALSIFIED" 19, "RETRACTED" 13, "SUPERSEDED" 13, "CORRECTED" 10 occurrences, plus the register's own supersessions. These are occurrences, not labeled pairs; building the pairs is the work.
Compare hosted Jev (fractions of a cent per run), the Ollaya models (after [T3]) and a deterministic
baseline, and fit temperatures per question type. A question enters the catalogue only if its decider beats its baseline on held-out labels [T4].
(a) RUN 2026-09-28 — `v3-bench/RESULTS.md`, predeclared in v3-bench/PREDECLARATION.md
(0dd0f3ed, committed before the run).
The question set turned out to be the whole real population: sessions sent v1 only **24 retrieve
calls (23 distinct questions) in two months**. Over the same months, 310 sessions used the memory folder instead: 325 file reads, plus 821 shell commands naming it.
nDCG@5 over 147 notes, blind Claude labels: grep 0.734 · BM25 0.782 · BM25→Jev 0.762 ·
Jev over all notes 0.739 · index 0.664.
F1 fired (not shown): BM25→Jev − grep = +0.029, 95 % CI [−0.054, +0.111].
F2's first reading is negative: Jev reordering its own BM25 candidates = −0.020, losing on 14 of
23 questions.
Total Jev cost $0.0136.
(b) RUN 2026-09-28 — `v3-bench/b/RESULTS.md`, predeclared 0c193468, amended before the run
9527e639.
Data: 334 author-grounded transitions (A) and 81 blind-agreed hard negatives (B), under an
asymmetric cost matrix: false supersedes 10, missed 1.
Verdict: v2 does not gain a decider. Governed Jev failed three of the five conditions on test A:
cost against always-abstain −0.192 with a CI of [−0.384, +0.036]; 6 false `supersedes`; macro cost 1.609 against 1.000.
The decisive finding: Jev gave refinements the supersedes label at probability up to 0.86,
above the 0.78 median of its correct supersessions. No threshold separates them.
The Claude text judge is the mirror image: best on hard negatives and on evidence and scope, but
it under-asserts supersedes and was inconsistent across its four slices.
Jev cost $0.0103 for 415 calls.
V3-2 The journal and law J-1, in v2's test harness under its heap caps.
V3-3 The RULINGS and FILE MEMORY adapters; recall and explain from the CLI.
V3-4 The A1 daemon, read-only, and a new plugin version pointing at it. Only then may a CLAUDE.md
tell sessions to use it again.
V3-5 propose, the floor and the judge, with DECISION certificates.
V3-6 The A2 Super capability and the RECEIPTS family: the loop site/index.html:729 describes.
V3-7 v1's Fly app is torn down once its consumers have moved (R96).
F1 If V3-1(a) recall does not beat index + grep on real questions, J1 is not worth building: keep the
file memory and stop at v2. Fired 2026-09-28, in the "not shown" sense (§7 V3-1 (a)).
On this evidence, a daemon, journal and decider are not justified for recall.
The proportionate step is a ranked search (BM25) over the memory folder. It beat grep by +0.049 with a
95 % CI of [+0.004, +0.092], just under the predeclared 0.05 margin.
Re-test if the memory grows well past 147 notes, or if questions become sentences rather than keyword
bags.
F2 If no decider beats its deterministic baseline on our labels, drop the judge. Graphonomous v3 keeps
the floor and recall. For recall (Q-RANK), the first reading is negative (−0.020 against its own BM25 candidates). For relation judgment, it fired 2026-09-28 (§7 V3-1 (b)).
Governed Jev beat both fixed rules on cost, but it did not beat refusing to mutate. It asserted false
supersessions with high confidence.
Under the predeclared rule, v2's relation proposals stay deterministic or manual.
Untested: a decider that only queues relations for human review (94 % of governed Jev's answers
were right); Ollaya; any other prompt or model.
F3 If law J-1 fails on any range, the journal design is wrong, and rebuild stays the only path.
F4 If a DECISION cannot be replayed to the same answer on the same model digest (reported: fp16
can flip near-ties), decisions are recorded as non-replayable and may not gate anything.
F5 If Super's receipt shape changes during stage 0, V3-6 waits. Nothing earlier depends on it.
Closed 2026-09-28. The disposition of every T# is in §C at the top.
T1 — Name. "Graphonomous v3", or a distinct name that cannot collide with semantic.v3 or G3.
Evidence since, 2026-09-28: both V3-1 benches came back negative for the new machinery.
(a): recall is served by BM25 over the folder.
(b): a typed decider may not propose relations.
What survives is v2 plus two source families (the memory folder and the ruling register) and an
external BM25 tool, with relation proposals left deterministic or manual.
That is an argument that there is no v3, only v2 growing. Not a ruling: whether v3 exists is
Travis's call.
T2 — Where the decider runs (§4.7), and whether "MCP servers make no LLM calls" covers a local,
recorded, sub-second decider.
T3 — Installing Ollaya. The installer is curl -fsSL https://ollaya.dev/install.sh | sh from a
project days old (reported: v0.3.1 → v0.7.3 between 2026-09-24 and 2026-09-27), or a pinned build from source. Which model, given 8 GB of VRAM.
T4 — Gates. The quality and calibration thresholds a question must meet to enter the catalogue.
T5 — Confidentiality. Does carried memory hold personal or secret content? This is the same question
as studbook §10.2 (content addressing versus access control), and F-2 is only the floor of it.
T6 — Profile family. graphonomous.semantic.vN, or a separate graphonomous.memory.v0.
T7 — Embeddings. Which local model, or none (lexical plus structural recall only).
T8 — The file memory's role. A source Graphonomous v3 imports, or the primary store with Graphonomous
v3 as a certified index over it. Evidence since, 2026-09-28: about 11× as many sessions used the folder as used v1 (310 vs 28), and plain search over it already puts a first-read note at rank 1 for 78–83 % of the real questions (grep 78 %, BM25 83 %). Both favour "primary store". Not a ruling.
Ollaya — github.com/ollaya-dev/ollaya, ollaya.dev, API.
Rust, Apache-2.0, "not affiliated with Ollama or TypeSafe". The creator is named elsewhere as Mert Cobanov.
Default address 127.0.0.1:11435. POST /v1/systemone and /v1/decisions, an Ollama-style /api/tags,
and a native /api/decide.
Limits: 1–256 questions per request and 65,536 state tokens, with STATE_TRUNCATED rather than a silent
drop.
Models run on ONNX Runtime (CPU / CUDA / MLX), and GGUF models on llama.cpp (since v0.7.0).
A decoder model answers by reading the final-layer scores for the option letters at Answer: ( and
softmaxing over the valid letters: one pass, no generation. Calibration is per-model temperature scaling.
ollaya mcp shipped in v0.4.0. It has no embeddings and no fine-tuning inside Ollaya; the upstream
decider model documents a LoRA recipe.
Author-measured on an RTX 4090 against hosted Jev at 0.738 (which includes network time): winnow:e4b
0.722 @ 89 ms, decider:4b 0.680 @ 520 ms, laya 0.361 @ ≈ 10 ms. JevBench v1.4.2.2 (benchmarkheaven.com/jev-models): decider-4b v2 64.1, Jev 1.13.0 63.3.
Jev — TypeSafe blog, OpenRouter guide.
Proprietary, "System One": a non-autoregressive single pass. $0.042 per million input tokens; output is
free. typesafe/jev-1.13.
There is no official calibration figure. An out-of-distribution study
(scienthoon/jev-ood-calibration) measured ECE 0.107, erring in a different direction per question type.
Already used live in this tree by alphapiler (~/Projects/Travelmonstr/alphapiler/src/jev.js).
Deciders in memory systems
lyonwj.com, knowledge-graph extraction:
a two-question noul assertion gate lifted precision 0.279 → 0.404 with no lost true edges;
a 3-level score for entity resolution lifted B³ F1 0.946 → 1.000;
a noul contradiction check separated planted contradictions where cosine similarity could not;
and "a System One model can't find an edge."
arXiv 2609.23986 (Jev-Mem): Jev as the memory controller for typing,
dedup/contradiction and retrieval routing; LoCoMo 0.777 against 0.700 for the best baseline.
Memworthy: gated memory with eight actions and a replayable
JSONL ledger.
No integration was found in Graphiti/Zep, Letta, cognee, LangMem or the official MCP memory servers.
MCP — 2026-07-28 release: a stateless core (no handshake, no Mcp-Session-Id), Tasks as an extension, and Roots / Sampling / Logging deprecated.
Could not verify: Ollaya's first-release date; the exact tool names ollaya mcp exposes; whether Ollaya can refit temperatures from user labels; any Jev architecture detail; and latency on this machine's GPU. The sources also disagree on Jev's calibration error (0.027 vs 0.246) and its accuracy (0.727 vs 0.738).
graphonomous/config/config.exs · graphonomous/lib/graphonomous/consolidator.ex · graphonomous/scripts/graphonomous_mcp_wrapper.sh · graphonomous/v2/README.md · graphonomous/v2/handoff/{STATUS.md,DECISION_LOG.md} · graphonomous/v2/lib/{project,query,evaluation}.mjs · super-live cockpit/src/{claude,codex}_connection.rs, site/index.html, site/AGENT_SUPER_APP_BLUEPRINT.md, ampd/lib/ampd/{capability_registry,receipts,projection}.ex · FOUNDATION_LANE.md · REVISION_REGISTER.md (R96, R97) · AGENCY.md · AmpersandBoxDesign/box-and-box/README.md (the bridge) · alphapiler PASS 9 (memory project_cyberutensil_compiler).