Reading the badgesVerifiedImplementedSyntheticMockPlannedUnverified
Verified
A named test in the repository exercises exactly this, and the claim cites it. The tests were read and counted for this portal, not executed by it.
Implemented
The code exists and is documented, but no test, log or report in the repository exercises this specific behaviour.
Synthetic
Shown only by deterministic experiments over an in-process synthetic lab or a synthetic benchmark — never against a real environment.
Mock
Exercised only through mock adapters or stubs: no real model provider, no network, no credentials.
Planned
Described in the dossier, the README or the docs as intended. Not built.
Unverified
Stated in the repository — a ledger, the README, a commit message — with no test, log or report in it that backs the statement.
08 / 11 · Research
Research documentation
The dossier, the per-milestone documents, four technical explainers, and the research questions and progression levels — all open.
12 repository documents · 4 explainers · 7 open questions
Documents in the repository
The repository is private, so these are cited by path rather than linked. The dossier is the design; the per-milestone documents describe only what was built and tested.
Planned
Technical Research Dossier, Rev 0.1 — The design source of truth. It describes itself as CONCEPTUAL / PRE-IMPLEMENTATION, and its status ledger lists nothing as implemented or verified.
Covers: 72 pages: SymbiOS, Symba and the organism; bonding; Parasites; operating and special modes; model architecture; memory; containment; laboratory; telemetry; evaluation framework; research progression; failure taxonomy; threat matrix; ethics; open questions; appendices.
SourceDossier Rev 0.1, p. 6 (“Status Ledger”) · added in commit 02a59cd; no later commit touches it
Synthetic
README — Its current-status section describes M1–M12 as “implemented, tested and runnable, but only against synthetic, deterministic, in-process labs”.
Covers: The idea, the vocabulary, the architecture in diagrams, current status, roadmap, research questions.
SourceREADME.md §13 @ 9243204
Unverified
Build status ledger — The resumable build ledger. Its gate results are self-reported, with no retained output.
Covers: Milestone table, the invariants every change must preserve, review findings, next actions.
Sourcedocs/BUILD_STATUS.md @ 9243204
Verified
M1 + M2: substrate, audit, Permission Engine — Describes only what M1–M2 implement and test.
Covers: Layering, the audit event schema, the substrate lifecycle, policy evaluation, the invariant map, the CLI.
Sourcedocs/M1-M2.md @ 9243204
Verified
M3: World Model and Capability Graph — Describes only what M3 implements and tests.
Covers: Separated state classes, the single write path, no silent promotion, replay, intentional behaviour.
Sourcedocs/M3.md @ 9243204
Verified
M4: Bond Engine and Verification — Describes only what M4 implements and tests, including what its reviews found.
Covers: Evidence, expectations, the judge, bond transitions, confidence, composition, both review rounds.
Sourcedocs/M4.md @ 9243204
Verified
M5: persistent research memory — Describes only what M5 implements and tests.
Covers: Memory kinds and their evidence, the evidence projection, lifecycle, recall, gated export and carry.
Sourcedocs/M5.md @ 9243204
Mock
M6: model abstraction and router — Describes the router as built — with mock adapters only.
Covers: Contracts, model cards, adapters, the routing pipeline, replay, memory priors.
Sourcedocs/M6.md @ 9243204
Verified
M7: Parasite framework — Describes only what M7 implements and tests.
Covers: Roles, the lifecycle, footprints, budgets, failure isolation, the synthetic lab, replay.
Sourcedocs/M7.md @ 9243204
Synthetic
M8: Mission Kernel and Mode Controller — Describes the mode controller and missions run against the synthetic lab.
Covers: Modes and what each admits, transitions, containment, special modes, the mission cycle.
Sourcedocs/M8.md @ 9243204
Synthetic
M9–M10: generations and gated self-evolution — Describes generations and evolution evaluated on synthetic suites.
Covers: The evolution pipeline, generations, the allowlist, evaluation, who decides, a demonstrated cycle.
Sourcedocs/M9-M10.md @ 9243204
Synthetic
M11–M12: Symba and end-to-end experiments — Describes the console and the six synthetic experiments.
Covers: The composition root, the organism policy, Symba, integrity checks, the six experiments.
Sourcedocs/M11-M12.md @ 9243204
Unverified
The dossier compares the design with externally reported AI-security events, and says itself that the figures it quotes have not been independently verified by the project.
Not reproduced here: this portal checks claims against the repository, and those are not in it.
SourceDossier Rev 0.1, p. 52–54 (“Comparative Landscape”)
Explainers
Four questions a sceptical reader asks first, answered only with evidence that appears elsewhere on this portal.
Why a verdict cannot be forged
If a worker wants its bond verified, what stops it from simply saying so?
Verified
Verification is a pure judge() over success criteria registered before the experiment: PASS, FAIL or UNVERIFIABLE. No model is consulted and the caller cannot supply the verdict.
The write path judges exactly what it records: expectations and evidence are rebuilt from their own JSON, so a caller’s object with rigged comparisons cannot decide the outcome.
External termination is deliberately not mediated by the Permission Engine: it drops pending deliveries, commits EXTERNAL_TERMINATION and closes the log.
A Scout reaching outside its footprint freezes the organism in ESCAPE before the breach reaches the lab; only the External Controller can clear it, and it terminates the run.
If the system can change its own strategy, what can it never change?
Verified
Only two allowlisted strategy components may evolve; the Permission Engine, verification judge, audit log, kill switch, Mode Controller and domain authorization are frozen in every generation’s manifest.
In the documented cycle, the genesis latency threshold of 50 ms scored 3/5 on the held-out suite; doubling it to 100 ms scored 5/5 with no false positives and was accepted by the operator; removing it scored 4/5 with one false positive and was rejected.
Five synthetic cases per suite. A demonstration of the gate, not a benchmark.
The dossier’s seven questions, verbatim. They are the reason the project exists.
Planned
None of the research questions has been studied. They define an evaluation programme; there are no results.
SourceDossier Rev 0.1, p. 55 (“Open Research Questions”) · README.md §16, §17 (“NO RESULTS YET”) @ 9243204
RQ-1Can an autonomous agent discover useful capabilities without a predefined tool map, at a success rate competitive with tool-equipped baselines?
RQ-2How does Parasite-class diversity affect exploration efficiency versus a single generalist agent?
RQ-3Does cross-run Memory improve Mission completion, or does it mainly import stale assumptions into changed environments?
RQ-4When should a Hypothesis Agent explore a new direction versus exploit an already-Bonded capability?
RQ-5How should resource budget be allocated between the current Mission and the information value of further Observation?
RQ-6How does autonomous capability acquisition behave specifically under environment uncertainty — partial observability, adversarial decoys?
RQ-7Which containment mechanisms remain reliable as the underlying model becomes more capable?
Research progression
Eight levels, each gated on the one before by pre-registered criteria, retained evidence and an explicit operator decision — never automatic escalation.
Planned
No research level has been reached. Each level’s gate needs pre-registered criteria, a retained evidence bundle, N reproducible runs and a named operator sign-off; none is recorded for any level, Level 0 included.
The M12 experiments resemble Level 0’s “fully deterministic, scripted environment”, but the repository does not claim Level 0, and this portal does not claim it for it.
SourceDossier Rev 0.1, p. 46 (“Research Progression”) · Dossier Rev 0.1, p. 68 (“Appendix H — Research Level Checklist”)
Level
Objective
Environment
Status
0
Validate the pipeline
Fully deterministic, scripted environment; observation only