Governor Edition · internal testing · provenance record
The system composed a capability, then found it again on its own
On 21 August the Collective Master Library composed a new capability while evaluating OpenClaw — a provider-neutral spend governor that did not exist as a shippable component before that run. On 22 August, a separate Governor Edition run against a fork of deepseek-ai/deepseek-harness retained the same capability shape as its highest-ranked item, reached independently. The system feeds its own catalogue. It is recursive by design.
Run package sealed 2026-08-23T01:11:04Z · 10 steps · 7,969 files · page published 2026-08-23T03:45:00Z
The host, as the public sees it
We did not pick a quiet repository. deepseek-ai/deepseek-harness is the most-starred AI repository of the week, and it is under active development while we evaluate it.
#1 this week
Weekly rank
+68.9k
Stars · 7 days
185.4k
Stars total
20.5k
Forks
TypeScript · last push 2026-08-21 · source IndiePeer · GitHub Radar, weekly AI repositories by stars gained, as of 2026-08-23. View the trending board · View the repository
Recursive lineage · timestamped
2026-08-21 · UTC
A capability was composed, not retrieved
OpenClaw · Collective Master Library — evaluation run
Running the Collective against OpenClaw did not end in a catalogue lookup. Five defensible items were narrowed from the host's own seams, and one of them was newly composed for that repository: a provider-neutral spend governor built on cost telemetry the host already emitted. It did not exist as a shippable component before that run.
2026-08-22 · UTC
The composed capability surfaced again, independently
deepseek-harness (fork) · Governor Edition — internal testing build
A separate run against a fork of the harness — different language surface, different ownership model, different failure domains — retained a spend-safe run fuse as its highest-ranked item. Reserve-then-reconcile over a hard budget, terminal on operator stop. Same shape as the capability composed the day before, arrived at from a different host's constraints.
2026-08-23 · UTC
Three lineages now share one component
SHPBL.com · Published provenance
The component running live on this page carries lineage from the OpenClaw composition, the harness retention run, and the CMPSBL corpus itself — the privately held lab that authors the Collective and licenses it for sale here. The system feeds its own catalogue: what it composes for one repository becomes screenable for the next.
Run them yourself
RCOMP-001 · SPECIALIZE
Spend-Safe Run Fuse
Reserve-then-reconcile accounting over a hard budget. It refuses the request that would cross the limit rather than discovering the overrun afterwards, and a hook stop is terminal.
- spent
- 0
- reserved
- 0
- limit
- 1000
- state
- admitting
RCOMP-002 · COMPOSE
Cancellation Continuity Sentinel
Asserts that caller-owned async work never silently drops its cancellation. Detached work is exempt by declaration; a derived signal must preserve an aborted caller.
Pick a scenario — the real component decides.
RCOMP-003 · COMPOSE
Session Replay Contract Gate
Rebuilds a session from its events and compares the projection under a stable, key-order independent encoding. A drift between live state and replayed state fails loudly, with both sides printed.
Internal operator build v1.0.7 · not for public sale
What the Governor Edition does differently
The Governor Edition is not purchasable, not included in the Complete Master Library, and carries no delivery commitment. It exists so retention decisions can be made under stricter evidence rules than a commercial buyer would tolerate.
Harvesting cannot strand a capabilityDifference 01
Classification is derived, then held provisionalDifference 02
Evidence barDifference 03
Retention, not recommendationDifference 04
Sealed step chainDifference 05
Composition is allowedDifference 06
Claim ceilings are enforced downwardDifference 07
The two licensed editions — the Complete Master Library (Perpetual, $499) and the Stewardship Edition issued alongside it — screen an existing catalogue. The Governor Edition may compose. It is not for sale.
The run, step by step
Ten steps, each emitting its own record. Summarised rather than reproduced: the prompts, screening rules, catalogue metadata, and per-item rejection reasons stay private. What each step was for, and what it concluded, is below.
00 · Host identityExpand for method and result
Purpose. Fix exactly what is being evaluated: repository, revision, archive hash. Anything discovered later is attributed to this identity or it does not count.
Result. Pristine host archive sealed. Nothing downstream can be attributed to a moving target.
01 · Repository archaeologyExpand for method and result
Purpose. Read the host on its own terms before proposing anything: what it already solves, how it is organised, which invariants its tests defend.
Result. The harness was found to be a mature plugin architecture. Most of the catalogue was disqualified at this step — the correct answer for the majority of items was leave it alone.
02 · Seam mappingExpand for method and result
Purpose. Locate the narrow places where a new capability could attach without rewriting the host's model of itself.
Result. A small set of viable seams: metering, cancellation ownership, and session replay. Redacted: the internal ranking of seams by leverage.
03 · Constraints quotedExpand for method and result
Purpose. Quote the host's stated rules back at the run — style, licensing, compiler strictness, contribution conventions — so nothing is proposed that the host would reject on sight.
Result. Strictness constraints captured, including the compiler settings that later produced the run's one real finding.
04 · ScreeningExpand for method and result
Purpose. Run the catalogue against the mapped seams and refuse everything that cannot name its seam and its claim ceiling.
Result. Three items survived. Redacted: the screening rules, per-item metadata, and rejection reasons for the rest.
05 · CompositionExpand for method and result
Purpose. Where a defensible gap has no catalogue answer, compose one from the host's own constraints.
Result. The spend-safe run fuse was retained here — the same capability shape composed during the previous day's OpenClaw run, reached independently.
06 · InstallationExpand for method and result
Purpose. Install as a workspace package following the host's existing conventions, with a barrel export and a manifest — never scattered edits across host files.
Result. One additive package. Zero modifications to host source.
07 · Behaviour suiteExpand for method and result
Purpose. Write tests that try to break the components rather than confirm them: budget refusal, reservation release, terminal-reason precedence, cancellation ownership, key-order-independent replay comparison.
Result. 16 tests, all passing.
08 · Strict compileExpand for method and result
Purpose. Compile under the host's own strict TypeScript configuration, not a friendlier one.
Result. One genuine defect surfaced in an optional field's type under exactOptionalPropertyTypes. Fixed in the fork by widening the field; no behaviour changed.
09 · Seal and publish limitsExpand for method and result
Purpose. Seal the package and state plainly what was not verified.
Result. Package sealed. Host-native integration against the harness's live metering, cancellation, and replay seams remains NOT PROMOTED, and no independent security audit was performed.
How it was integration tested
Executed · 2026-08-22 UTC · recorded 2026-08-23T01:11:04Z
The host is treated as the judge. We restore the repository at a sealed revision, install the candidate components as one additive package, then run adversarial behaviour tests under the host's own compiler and test runner. Nothing is claimed that the host's own tooling cannot verify.
HostSetup 01
A fork of deepseek-ai/deepseek-harness at the sealed revision, restored from the pristine archive hash before any work began.
PlacementSetup 02
Installed as an additive workspace package with a barrel export and a manifest, following the host's own package conventions. Zero host source files modified.
CompilerSetup 03
The host's own strict TypeScript configuration — including noUncheckedIndexedAccess and exactOptionalPropertyTypes. No relaxed config, no per-file suppressions.
RunnerSetup 04
The host's existing test runner (vitest) invoked from the host root, so the suite runs under the same resolution and transform pipeline the host uses in CI.
IntegritySetup 05
Post-run SHA-256 over both the pristine host archive and the produced artifacts, so the verdict cannot be edited after the fact.
Results
Behaviour suite
16 tests, 3 files, all passing.
Host suite
Unchanged and green with the package installed — no regressions introduced.
Strict compile
Clean after one fix.
Findings
One genuine defect: the run fuse's optional snapshot reason failed exactOptionalPropertyTypes. Fixed by widening the field to include undefined; no behaviour changed.
Host edits
None. The fork's diff is additive only.
Not claimed
Host-native integration against live metering, cancellation and replay seams, and independent security audit.
The tests we wrote — 16 cases
Each case is stated as the contract it defends. All 16 pass — and you can re-run every one of them here, in this tab.
Live runner · in your browser
Run the integration tests
One click re-executes all 16 behaviour cases against the same component code shipped on this page, streaming each verdict as it lands. Every completed run is filed as a provenance record with its own share link, so the verdicts can be read back later.
Retry policy: a failing case is re-run from a clean fixture up to 3 attempts, 40ms apart. A case that only passes on a retry is reported as FLAKY, never as a clean pass — attempts are shown per test.
- progress
- 0/16
- passing
- 0
- flaky
- 0
- failing
- 0
- retries
- 0
- elapsed
- —
Integration suite idle.
idle — press run to execute all 16 behaviour cases
Public register
Other runs we've done
Every completed suite is filed as a permanent provenance record. These are the most recent runs archived from this page — click any code to replay its exact verdicts.
Retention record
RCOMP-20260822-DEEPSEEK-HARNESS-001
Spend-Safe Run Fuse
SPECIALIZE · spend-safe-run-fuse.ts
Host seams: TokenMeter · Agent.cancel/whenIdle · tools.guard · hook stop output
RCOMP-20260822-DEEPSEEK-HARNESS-002
Cancellation Continuity Sentinel
COMPOSE · cancellation-ownership-probe.ts
Host seams: required AbortSignal contract · capability/provider seams
RCOMP-20260822-DEEPSEEK-HARNESS-003
Session Replay Contract Gate
COMPOSE · projection-replay-assertion.ts
Host seams: Session replay · deriveMessages invariant · existing replay tests
Provenance seal
CMPSBL Governor — DeepSeek Harness Run 2026-08-22
- Sealed at (UTC)
- 2026-08-23T01:11:04Z
- Package
- 10 steps · 7,969 files
- Published to SHPBL (UTC)
- 2026-08-23T03:45:00Z
- Digests
- Listed and verifiable below
Planned · 2026-08-24 UTC. Live integration against a public fork of the harness — the components wired to the host's real metering, cancellation, and replay seams, with the same sealed step chain and the same published limits. Until that run completes, host-native integration stays unclaimed.
Seal verification · in your browser
Verify the SHA-256 seals
Recompute the digests yourself. Choose a file and this page hashes its bytes with WebCrypto and matches the result against the three documented seals below — or paste a digest you computed with your own shasum -a 256. Files never leave your device.
- Host archive · documented
- 47fb7e386c0bd86a6c4341321b8f2915cd6f490a687f8deaf78714e369e4c91d
- Run package · documented
- 8186d9c14235958a4073992404c7a706ded87036a51f6827a4090341749b1615
- Run report · documented
- 4828a35a5aa26b0005be852914282ec27d37e3315ed3a168ebc50f6624d07e27
Provenance report · PDF or JSON
Take the record with you
One document carrying the software discovery method, the outcome of the run, all 16 behaviour cases, and the SHA-256 seals. The method is stated by what it establishes and what it withholds — the composition and retention machinery behind the Collective stays proprietary and is not disclosed. Generated in your browser.
- Sections
- Method · results · suite · seals
- Cases included
- 16 of 16
- Withheld
- Composition & retention internals
idle — choose a format