Registered Work · project · active

SISO Harness Lab · Agent Effectiveness

Turn observed agent failures into tested improvements, while keeping rejected and unproven results visible.

Type
project
Maturity
active
Section
Outside selected section view
Snapshot
Registered · not selected

The Library’s reading

Why it matters

The improvement loop needs to preserve evidence before drawing conclusions. The source owner reports an accepted repair that prevents racing writers from replacing a winning result and rejects malformed output without creating an artifact. That is a bounded engineering improvement. Equal-spend quality gains remain unestablished, and overlapping pilot attempts do not justify instruction promotion or a general efficiency multiplier.

Read the reasoning & boundaries ↗

Sauce

The useful things to open first. Claims stay attached to dated receipts.

For agents

SISO Harness Lab source owner · Start with the source entry, then the machine-readable dossier.

Documented entry commands: npm run project:status. These are documented source entry commands; fresh test results are recorded separately in the evidence. Follow the owning source instructions before execution; private access and campaign activation remain separate.

Relationships

related_to

GQ-002 · 10× the Agent LayerThe agent-layer efficiency question supplies research direction; Harness Lab holds scoped experiments and acceptance evidence, not a claimed answer or general multiplier.

related_to

SISO SkillsReusable skills retain their independent Skills Hub source and promotion workflow. Harness Lab evaluates interventions; it is not a second skill catalog.

related_to

SISO Session IntelligenceSession Intelligence remains an independent privacy-reduced ingestion and recall capability. Raw sessions are not copied into Library metadata.

Evidence & receipts

source_review · 2026-09-06

Eight local tests pass for quality-first evaluation, malformed input, immutable successor receipts, cold pickup, no-op repeats and exclusive-writer refusal. No fleet savings, scheduled agent, model training or public payload Release is claimed.github-api:sisodias/siso-harness-lab:a0ae561e1c6a05ef747b1cc304849e9d723fc80b

documentation · 2026-09-08

Owner-provided public-safe entry reports a local human/status surface, immutable diagnostics, an outcome grader and exclusive result publication. The accepted racing-writer repair is owner-reported, not independently replayed by the Library. Equal-spend quality gain is NOT_ESTABLISHED; the overlapping pilot does not support instruction promotion. The new detailed report is not yet available at the supplied GitHub branch path.owner-report:siso-harness-lab:2026-09-08

source_review · 2026-09-08

Private repository entry path resolves through authenticated metadata lookup; blob 129059eda192941069e8b53cb5c533e93b96e7d2. No private source bytes or local runtime address are distributed by this Work.Open receipt ↗

Provenance

Work ID
gls:work:1cc4c944-9351-466e-a442-9fe16ad145d1
Registry source
registry/works/siso-harness-lab.json
Origin
siso
License / redistribution
NONE (not_applicable)