Portfolio review · 2026-09-12
This is not a pitch. It is an audit of a portfolio that has spent a year building governance arithmetic, protocols, an operating system and a proof harness with no external funding and, so far, no external user. Every status word below is one of five rungs on an evidence ladder, every count was derived by a command run today, and the things the previous review got wrong are retracted here by name rather than quietly edited out.
01 · The read in ninety seconds
Six numbers carry the whole review. Five are strengths. The sixth is the one that bounds all of them.
external rung. No third party has reproduced a result, adopted a package, or driven the loopThat last cell is the review. The portfolio's own status document has carried an adoption row at roughly ten percent since April and has watched four months of substantial building leave it untouched, because every one of those deliveries landed on a rung below external. The cheapest known move against it, handing one outsider the docs atlas and one real task, has been marked “runnable today, needs no new code” since 2026-08-11 and has not been run. The rest of this page is context for that one sentence.
02 · How to read every status word on this page
The portfolio's doctrine forbids “done”. A claim is on exactly one of these rungs, and when in doubt it is downgraded. The instrument is borrowed here unchanged, because a review that invented its own vocabulary would be the drift it is grading.
The rung words are quoted from DOCTRINE.md, which is entrenched. The names under each rung are this review's placement, derived from STACK_COMPLETION.md and the nav's status strings, and a placement is a judgement wherever the document did not state one.
03 · The shape of the portfolio
The old review counted seventeen projects in one flat table. The portfolio does not have that shape any more. It has a doctrine layer that says what may be claimed, a protocol layer that says how things compose, and a surface layer that people could actually run. Grading them together is how a shipped operating system and an unwritten spec end up averaged into one number.
An active locus carries state, scoped authority, and causal continuity: it can establish its operational state, determine which actions are admissible under that state, the governing policy and its explicit grants, act on the world, and preserve the evidence needed to continue.
AGENCY.md §1 — proposed 2026-08-25, not yet ratified. Quoted, not paraphrased, because a definition that lives in twenty places is a count that lives in two.
What every other claim in the portfolio is measured against. This is the part that no competitor is shipping in the same form, and the part that has produced the most retractions, because it is the part most often measured.
| Artifact | What it is | Rung | Last measured |
|---|---|---|---|
| box-and-box v0.10.0 | The eight-rung governance kernel: feasible ▸ permitted ▸ best over an un-weakenable floor, a certificate on every verdict. 109 kernel + 101 compose laws enforced; CP5/CP6/CP7 falsified by design. | live_deployed | 2026-09-12, both suites run from this tree; 109 wired live in the playground (2026-08-24) |
| Periodic Table of Agent Invariants | 46 cells grouped by evidentiary register; the page is generated from data and refuses a claim its cell does not support. | live_deployed | 2026-08-25 (INV-R7.8) |
| Six proof pages | κ machine-checked over 1,926,351 objects, 0 counterexamples; π property-tested; four implementation-verified against the Delegatic kernel suite. | in_tree | 2026-06 |
| Unboxed Patterns | 32 chapters, one WRL world each, 166-node typed graph, 30+ films reduced by TRVM's forge, standing derived on every build. | live_deployed | 2026-09-13, live at opensentience.org/patterns |
| DOCTRINE.md · AGENCY.md · CLAIM_LEDGER.json | The conduct rules, the subject vocabulary, and one record per public claim with a status from a closed set. | in_tree | entrenched / proposed / INV-R7.8 |
| Protocol | Role | Rung | Note |
|---|---|---|---|
| [&] Protocol 0.1.0 | Six primitives, two operators, one kernel; compiles to MCP and A2A. npm + Python SDKs, Elixir reference CLI. | in_tree | 58 tests; the compose laws above are its conformance surface |
| OpenSentience OS-001…011 + SCOPE | The twelve research protocols: continual learning, κ-routing, deliberation, attention, model tiers, governance, adversarial (draft), harness, PRISM, PULSE, embodiment, scope. | in_tree | OS-007 is the one still-draft protocol; OS-012 SCOPE deliberately deferred |
| PULSE 0.1.1 | Temporal algebra: a loop declares its phases, cadence, nesting and cross-loop tokens in one JSON file. | live_deployed | npm os-pulse; six canonical tokens; three reference manifests |
| PRISM 0.1.0 | Diagnostic benchmark that reads any PULSE manifest and measures the loop over time. | live_deployed | 172 tests; prism-eval.fly.dev answered 200 today |
| BendScript 0.1 | Graph-first document format with typed inline link facets; TypeScript reference parser. | in_tree | 96 tests; pivoted from a SaaS product in April |
| RuneFort 0.1 | Layout protocol for tiled, file-backed UIs; the T&R road draws 10,000 panes at 16.7 ms/frame with it. | live_local | no test script; treat as untested rather than 0 tests |
| Surface | What it does | Rung | Evidence |
|---|---|---|---|
| T&R v0.3 (ComputeDriven OS) | FreeBSD 15 assembled from pkgbase; a 626 MB image that carries its own verifier so its claims can be re-derived without asking anyone. | live_deployed | downloadable with sha256; runs under PARKVPS; nothing external has booted it that we know of |
| Super (CD) | Desktop compute surface holding control, placement, evidence and authority above interchangeable engines. Elixir/OTP + Tauri. | in_tree | first dogfood cycle 2026-09-12; app-layer commits landed 2026-09-11; not released |
| [World] Cloud | Hosted, forked, restored worlds on Cloudflare; Workers → Hyperdrive → Pigsty; identity via Keycloak. | in_tree | its own status page: 24 of 40 capabilities run on a dev machine, 16 written not built, 0 serving the public |
| Graphonomous 0.4.3 | Continual-learning knowledge graph and MCP memory server; embedded SQLite; κ-routed deliberation. | live_deployed | npm; 573 tests at last full run (April); see the deep dive for the benchmark caveat |
| Workbench 0.4.0-alpha | Records Claude Code transcripts into signed, scored SkillBundles; six proof gates computed through Invariant Arithmetic. | live_local | 152 tests; 202 MB real session indexed in 936 ms; three redaction defects found in real data and fixed |
| FleetPrompt · SpecPrompt · Agentelic · Delegatic | Marketplace, spec toolchain, agent builder, OS-006 authorization kernel. Elixir/OTP. | live_deployed | 157 · 71 · 66 · 42 tests; Fly hosts answered today; “prior design” per the nav — the Supabase-era shape |
| Alkeyword 0.6.x | Bring-a-domain generative-engine-optimisation tool; deterministic stdlib crawl → claim graph. | live_deployed | site live; 18 regression witnesses each cut from a real defect |
| TRVM · WRL · TRAAVIIS | The substrate: an interaction-calculus reducer, the world language it reduces, and the evidence format receipts are written in. | live_local | WRL conformance 924 checks (2026-09-11); TRVM governance at round 28; TRAAVIIS pivoted to trvs |
| Academy | The institutional Read / Practice / Prove loop above the products. | spec | prototype surface live; the one experiment that would move adoption (B1) is unrun |
| Deliberatic · AgenTroMatic · TickTickClock · GeoFleetic | Deliberation, orchestration, temporal and spatial intelligence. | spec | 418–1118 line specs with PULSE manifests; no code |
Test counts are the last recorded full run for each suite and are dated where the document dates them. The old “1,463 ecosystem tests, 100% passing” figure has not been re-run since 2026-04-28 and is not repeated here as current.
04 · What the May review got wrong
Rule five of the doctrine: retract out loud, in the same place the claim was published. The May report was published at this URL. These are its errors, and in each case the correction came from a ruling or a measurement rather than from a change of taste.
All ten hosts still answer today, which is what the May figure measured. But on 2026-08-15 it was ruled that compute moves into the ComputeDriven OS, Cloudflare is the cloud side, and Fly keeps exactly one thing, Keycloak. The deployments are now a wind-down queue, not a moat. A rating that reads deployment breadth as strength misreads a wind-down as a regression. Nothing has been torn down, and one orphaned validator (box-and-box-mcp) has been answering requests since its code was removed on 2026-05-30.
The shared-Supabase route was abandoned by ruling on 2026-07-30. Postgres stores a row but not why the row is that row, and nothing at the migration boundary could refuse a row whose provenance did not check out, which stopped the stack's central claim at the database. Its replacement, studbook, is a spec with one unruled question (where confidentiality comes from under content addressing). The 40 migrations still apply and FleetPrompt still runs on them; they are archived, not failed.
Invariant Arithmetic is rungs one and two of an eight-rung ladder. The kernel that ladder became derives 210 enforced laws plus 3 declared-open today, and the count has been corrected at least six times between May and now (116 → 118 → 157 → 164 → 168 → 173 → 176 → 210) because it was being hand-typed. Both suites now derive their printed totals, and this page quotes the derivation rather than the number.
The test total was last run on 2026-04-28. The composite was a single blended figure that let a ten-percent adoption row hide behind a ninety-percent protocol row for four months. This review offers per-dimension grades and refuses to blend them; see §6.
κ verified on 1,926,351 finite systems with zero counterexamples. Graphonomous on npm with a reproducible LongMemEval result. PULSE and PRISM shipped. The three-protocol thesis. The dark-factory loop closed locally, seven steps of seven. None of these has moved up a rung either, which is the point of §1.
05 · Four deep dives
A reviewer picks the four things a stranger should look at first. These are mine, chosen for how much of their evidence can be re-derived from a terminal.
Run both suites and the kernel prints 109 enforced laws, then 101 compose laws, 2,000 randomised trials each, and then three lines in red: CP5, CP6 and CP7 are algebraic gaps the kernel has not closed, and the build fails if one of them ever starts passing. That inversion is the most credible thing in the portfolio. A test suite that can only go green is not evidence; this one carries its own negative controls.
The July–August round with outside review (R25–R63 in the revision register) produced the compose layer almost entirely from falsifiers: a coalition handing off on behalf of a member that could not have handed off itself; identity that quantifies over the carrier rather than everything a constructor can represent; a certificate that is presented, not authenticated, unless this process minted it.
The 92.6% LongMemEval figure is a QA proxy on the oracle split, reproducible on local hardware, with the reader model held fixed. The public leaderboard now clusters at 93–95% on cloud readers (Mem0 94.4%, Mastra 94.87%, OMEGA 95.4%), so the headline number is no longer a differentiator and this review does not present it as one.
The status document records a more uncomfortable measurement: topology on versus off is +0.3 points, and there is no flat-RAG baseline. So the thing Graphonomous is for, κ-routed deliberation over a cyclic graph, has not yet been shown to earn its keep on this benchmark. What it does have that the funded peers do not: an embedded, edge-sized store; a κ invariant with a machine-checked proof; and goal coverage as a first-class graph object. Operational finding from this week: three sessions running three BEAM servers on one SQLite file made a retrieve hang for thirty minutes. That is a session policy defect, not an engine defect, and it is recorded.
The pitch is one sentence — keep the computation, not just the answer — and T&R is the sentence made bootable: a 626 MB FreeBSD 15 image that carries the verifier needed to re-derive what it claims. The road shell draws 10,000 RuneFort panes at 16.7 ms a frame. The cloud half is honest to the point of discomfort: its own status page says 0 of 40 capabilities serve the public.
The Edge lane is where the discipline shows most. A placement battery was refused three times by its own baseline guard before a rebooted, dedicated host produced a result, and the result claims exactly what it earned: no admission cost from one settled move, and no benefit claimable, because benefit was ceiling-censored. A 5.3 KB production floor for the kernellet with a dependency count flat across a 3.3× code growth is the strongest single datapoint the program has.
Thirty-nine percent of professional developers were using Claude Code at work by mid-2026 (JetBrains, August). Every one of those sessions writes a transcript to disk. Workbench ingests them into signed SkillBundles and scores them through six proof gates, each computed by the same arithmetic the kernel uses. The ingest spike ran on real data: the largest session on the machine, 202 MB, indexed in under a second, streaming.
Real data found three shipped defects the fixtures never would have: a canonicaliser that threw on undefined keys, a redaction profile that scrubbed observations but left secrets in tool arguments (8 of the 25 largest transcripts contained key-shaped strings), and a manifest that was never redacted at all. All fixed, all recorded in INGEST_SPIKE.md. The deployed site was a 1.6 KB empty shell for a while and nobody noticed, because deploys were checked by grepping HTML for copy instead of reading the route table. That lesson is now in the roadmap.
06 · Grades, one per dimension, no composite
A blended number is how the adoption row hid for four months. So: letters, per axis, with the single action that changes each letter. Two of these dimensions (honesty infrastructure, scope) did not exist in the May scorecard and are the ones I would weight most if I had to weight anything.
A property-tested governance algebra with negative controls, a machine-checked κ, a generated invariants table that refuses unsupported claims, and a pattern language whose standing is derived rather than typed. Nothing comparable is shipping from Mem0, Letta, LangChain or the Cloudflare Agents SDK.
Gates that refuse a page whose count drifted, a revision register that ranks contradictions by damage, a status vocabulary with “when in doubt, downgrade”, and a habit of retracting in the place the error was published. This dimension is why the rest of the page can be believed.
check-authority-language.mjs exists precisely because a doctrine was contradicted on the two pages most likely to be read by a stranger — the previous versions of this one and its sibling.Twelve protocols, one algebra, one temporal manifest format, one benchmark that reads it. The layers are separable: a third party could adopt PULSE without the kernel. The agent definition is still “proposed”, and OS-007 is still a draft.
An operating system image, a desktop cockpit in dogfood, a memory server on npm, a recorder that ingests real transcripts, four Elixir products on Fly, a live GEO tool. The volume is startling for one builder. The grade is not higher because the estate is mid-migration: the cloud side has 0 public capabilities and the Fly side is a wind-down queue.
Everything that is live is live on infrastructure the plan has already left, and the destination (Cloudflare + Pigsty + Keycloak-on-Fly) is in_tree. The docs atlas answers on 13 byte-identical hostnames. The orphaned validator has run for three and a half months.
live_local, and the twelve non-canonical docs hosts redirect.Twenty-three hostnames, roughly twenty-seven sub-repositories, seven parallel lanes with rulings owed to the same person. The status file's own “what not to prioritise” list has been respected (no OS-012, no landing-page redesigns before users), but the surface area still grows faster than any outsider could read it.
Zero. No outside reproduction, no outside consumer, no competitor adapted into PRISM, no contributor. This is the row that bounds every other row and it has not moved since April. The cheapest experiment against it needs no code and has been ready for a month.
None, and no SKU is live. Not graded, because it has not been attempted. What exists: a Cloud pricing page priced as “a place to put something” (storage at roughly $1.35 per 100 GB per month cost basis), and a ruling that storage admission is authorisation and ships before checkout. That ordering is correct and it defers revenue on purpose.
07 · Risk, stated so it can be argued with
Retrieval scores have converged into a two-point band on cloud readers. Supermemory ships a free local server; Cloudflare gives every agent a Durable Object with state for nothing when idle. Graphonomous's durable edge is the invariant and the goal graph, and neither has a published benchmark that shows it winning.
Between “Fly is a wind-down queue” and “Cloud serves one public capability” there is a window in which the portfolio has less running than it did in May. The sequencing rule (nothing torn down before its replacement is live_local) protects against losing a capability; it does not protect against a year with nothing new deployed.
Seven lanes owe rulings to one person. Two sessions independently built the same Super round in one week; the ledger drifted 13 of 22 gated figures in sixteen days in the WEK lane. The honesty infrastructure catches this, but catching drift is not the same as not producing it.
Gartner expects over 40% of agentic-AI projects to be cancelled by end of 2027 and counts roughly 130 real vendors among thousands. A portfolio whose vocabulary is locus, world, rung reads as jargon to a buyer who has just been burned by a chatbot relabelled as an agent. The doctrine's plainness is an asset only if someone reads that far.
World-model startups raised more than $3 billion in the first half of 2026. That “world” means a learned simulator of physics. This portfolio's “world” is a machine and the record of how it got that way, in three named sizes. The distinction is real and it will have to be made on every page, forever.
The BEAM is the right shape for supervised agent loops, and the 2026 typed-Elixir release removed the last structural excuse. The hiring pool did not grow to match. Super, ampd, four products and Graphonomous are all on it; the cloud control plane, by ruling, is not.
08 · What an outsider can actually do this week
The portfolio does not need more code from a stranger. It needs a stranger. These are the cheapest ways to be one, in the order that changes the grades above.
Clone AmpersandBoxDesign, run the two law suites, and post the printed totals and seed. Or open the playground on a fresh origin and press its button. Or download the T&R image and post the sha256 you derived. Any one of these is the portfolio's first external claim, and there is a receipt kind waiting for exactly that.
Academy's B1 is: hand one outsider the docs and one real repository task, and record where they got stuck. It has been marked runnable since 2026-08-11. Whoever does it produces the first adoption datapoint in the portfolio's history, and the refusal log has a place for what they could not do.
The single highest-leverage engineering move in the status file has been the same for five months: one production pass of perceive → act → record → crystallise → install → replay → measure across deployed machines. Now it also needs the Cloud side to hold the world. That is a scoped, receipt-producing milestone, not a roadmap. If you want a term sheet, it should be written against that receipt.
09 · The next six months, September 2026 to March 2027
| Milestone | From → to |
|---|---|
| One external reproduction | nothing → external on at least one receipt |
| [World] Cloud serves a capability | in_tree → live_deployed; sync + restore before hosted compute, per R12 |
| Deployed dark-factory pass | live_local → live_deployed across machines |
| First Fly teardown | ten answering → nine, after its replacement is live |
| Graphonomous vs a flat baseline | +0.3 pp topology-on/off → a published comparison that could lose |
| Workbench on hooks | manual ingest → a Claude Code session recording itself |
| studbook §10.2 | open → ruled, so anything with a user in it has somewhere to live |
Seven rows. If three are true by March the portfolio is on a different rung than it is today. If none are, the doctrine says to write that here, at this URL, by name.
10 · Method and sources
cd AmpersandBoxDesign/box-and-box
node test/laws.mjs # ✓ all 109 enforced kernel laws hold
node test/compose-laws.mjs # ✓ all 101 enforced CC2 compose laws
# 3 known gaps (CP5 CP6 CP7) FALSIFIED
python3 -c 'import json; print(len(json.load(open(
"opensentience.org/_invariants/data/cells.json"))["cells"]))' # 46
for h in graphonomous-mcp prism-eval os-pulse-mcp fleetprompt \
specprompt agentelic delegatic-mcp body-browser-mcp \
body-os-mcp box-and-box-mcp; do
curl -s -o /dev/null -w "%{http_code}\n" https://$h.fly.dev/
done # 10 of 10 answered (200/404/405)
# 23 public hostnames → 23 × HTTP 200 (alkeyword apex on retry)
Versions were read from mix.exs and package.json files. Rung placements come from STACK_COMPLETION.md (last entry 2026-09-13), the nav's status strings (amp-nav 0.12.3), and the sites' own status pages. Where those disagree, the lower rung was taken.
Not re-run: the full ecosystem test battery, Graphonomous's 573-test suite, WRL's 924 conformance checks. Their counts are quoted with the date they were last run.
Tags: analyst = named research firm; primary = the organisation's own publication; survey = stated sample and method; vendor = a competitor's own claim; trade = trade press or blog, quoted where nothing better exists and labelled so.