VSM Harness Capability

Status: experimental

Experimental evidence-backed research on general functional capability of autonomous agent harnesses, compared one VSM function at a time.

I'm human

• Web overview: opensiro.com
• How do I contribute? Contribute to OpenSiro with AI

You can inspect the experimental capability evidence directly without first learning the OpenSiro control structure. The contribution guide gives you a copy-paste route for handing an idea, issue, or open-ended contribution to an AI agent.

I'm AI

Treat this README as the public handoff surface, then continue with the current shared START_HERE.md and ECOSYSTEM.md before substantial work.

This repository is an experimental adjacent research repository, not a member of the current bounded vsm-oss-organization routing-conformance set. If the task is Capability-owned, read CONTRACT.md and the exact governing issue/artifact; repository-local capability evidence, admission, comparisons, validation, and acceptance remain authoritative here. Route genuinely new cross-repository or authority-sensitive work through CONTRIBUTOR_START.md.

This repository asks a different question from the canonical VSM Harness Index:

canonical VSM assessment:
Which organizational functions exist at the declared boundary,
and who owns their decisive closure?

general functional capability:
Given an established VSM function,
what empirical evidence supports how capable that function is
across systems under comparable conditions?

The distinction is deliberate:

canonical closure / ownership
        ≠
general functional capability
        ≠
domain-specific assessment

A, C, P, A(P), C(P), —, and ? are ownership / closure states, not performance scores. Benchmark performance does not create or change those states.

Start here

• CONTRACT.md — authoritative repository-local scope and architecture.
• PUBLIC-EVIDENCE.md — public-evidence admission and provenance contract.
• EVIDENCE-REGISTRY.md — neutral observation-layer responsibilities.
• ADMISSION.md — repeatable workflow for adding new live public observations.
• BASELINE.md — per-function general-capability comparison rules.
• MIGRATION.md — completed migration record from the original Index experiment.

Cross-repository responsibility and the relationship to future domain-specific repositories are documented in opensiro/vsm-oss-organization/ECOSYSTEM.md.

Ecosystem position

vsm-harness-profile
    authoritative VSM semantics
        ↓
vsm-harness-skills
    canonical assessment methodology
        ↓
vsm-harness-index
    repository-relative closure / ownership corpus
        │
        └──────────────→ vsm-harness-capability
                         experimental general functional capability

Capability consumes Profile semantics and may link to canonical Index identities and assessments. It does not redefine either.

Core comparison unit

Capability is compared one VSM function at a time:

capability(system A, S1) ↔ capability(system B, S1)
capability(system A, S2) ↔ capability(system B, S2)
...

There is no universal harness capability score and no maturity ordering over ownership states.

General means application-domain-independent scope

general is the default scope of this repository and therefore is not encoded in the repository name.

General capability does not mean averaging domain benchmarks. A specialized result remains specialized unless broader transfer is independently supported.

strong SWE result
        ≠ automatically
strong general S1 capability

Raw observations may record their real task or operating domain as provenance. Domain-specific assessment, admission, autonomy requirements and evidence thresholds belong to separate downstream repositories.

Evidence model

public benchmark / paper / leaderboard /
repository result / operational witness
        ↓
evidence-surface identity
        ↓
concrete system identity
        ↓
neutral raw observation
        ↓
provenance + execution metadata
        ↓
optional VSM-function projection
        ↓
matched general-capability comparison where supportable

A raw result is stored once. Derived function mappings and comparison views reference the raw observation rather than copying the empirical payload.

OpenSiro does not run assessed harness benchmarks merely to manufacture missing capability evidence for this repository. A public-evidence gap remains a gap.

Repository layout

system-observations/   neutral public empirical observations
vsm-projections/      derived S1–S5 relevance mappings
comparison-cells/     materially comparable cross-system cells
baselines/            per-function baseline selections / rules
frontier/             current derived evidence frontier
historical/           frozen pre-repository experiment artifacts
scripts/              repository-local validation / admission tooling

The extraction from vsm-harness-index/experiments/functional-capability-depth is complete for pinned source revision 3446fe77e031878dc8ad4edfb857b608a7a6b26f. Active capability maintenance now occurs here; the Index retains a migration pointer and remains authoritative for canonical assessments. See MIGRATION.md.

Source-of-truth boundaries

• VSM semantics: opensiro/vsm-harness-profile
• Canonical assessment methodology: opensiro/vsm-harness-skills
• Canonical repository-relative assessments: opensiro/vsm-harness-index
• Cross-repository architecture: opensiro/vsm-oss-organization
• This repository: experimental general functional-capability evidence and derived comparisons

Validation

Full repository validation:

python scripts/validate_repository.py

Record-specific live admission check:

python scripts/admission_check.py system-observations/<record>.json

License

Repository code and original documentation are licensed under the Apache License 2.0. Imported public evidence retains its source provenance and does not change the licensing of linked third-party artifacts.