Software investigation, validation tools, and engineering handoffs grounded in inspectable work.
I'm S.J., the independent creator behind Estra Logics. My work brings together engineering and software production management: define the problem, prepare a focused correction, check the result, and leave a usable record.
Accepted contribution · Try a project · Evidence archive · Contact me
microsoft/semantic-kernel #14454 · Merged October 27, 2026.
| Part | What happened |
|---|---|
| Problem | Constrained Pydantic fields could place constraint metadata in a JSON schema description instead of the intended text, causing serialization to fail. |
| Correction | Prefer the field's description and retain compatible text or dictionary metadata fallbacks. |
| Regression coverage | The patch adds cases for a constrained field with a description and for a description carried in annotated metadata. Both exercise JSON serialization. |
| Outcome | The upstream project merged the contribution. The pull request and changed files preserve the correction, tests, and review history. |
This is a concrete example of source investigation, compatibility work, and regression design. It is evidence for that specific correction; project-wide reliability requires its own validation.
Problem: a failed edit should leave the last accepted document available, and concurrent windows should report conflicts rather than silently overwrite one another.
Implementation: local static validation, atomic document storage, revision checks, accepted-state recovery, and downloadable verification receipts.
Try it: load the supplied sample, validate it, remove a required lock, and validate again. The ten-minute test describes the expected FAIL result and preserved accepted state. It begins after setup and asks testers to record their actual results.
Stage and limits: alpha. Static document checks and receipt verification do not verify external execution. The published alpha.2 archive exports unsigned receipts; optional Ed25519 signing is documented for the source checkout.
Download alpha.2 · Setup and source · Receipt signing and trust
Saved preview. The hosted version currently requires ChatGPT sign-in and permission to access; public source is available for local use.
Problem: a person should be able to see the interface state and actions an agent is working with.
Implementation: a spatial React workspace with six WebMCP tools. Human controls and agent tools use shared state, with actions reflected in an activity stream.
Try it: explore a Bloom Card, compose a Wonder Route, and change the room's weather. The human interface works in an ordinary browser; agent interaction requires a compatible WebMCP browser.
Stage: public source demonstration with restricted hosted access. This is an interface example, not evidence of general autonomous control.
Hosted preview · Local setup and demo sequence
| Project | Problem and implementation | Stage |
|---|---|---|
| CML Trust | Keep evidence and later corrections traceable through plan/evidence hashes, a local JSONL ledger, coverage review, and preserved revocations. Observations stay separate from canonical CML state. | Research prototype / alpha |
| Pattern Atlas | Review scattered records through source-noted timelines, typed relationships, filters, and JSON import/export. Connections record relationships; interpretation remains with the reviewer. | Browser prototype |
openai/openai-go #932 — preserve manual response history addresses ordered replay of completed response items. The submitted work includes an SDK helper, usage example, and regression fixtures.
Status checked November 10, 2026: open and unmerged. Follow the pull request for subsequent review and validation. It is listed separately from the accepted Microsoft contribution.
| Work | Typical deliverable |
|---|---|
| Software investigation | Reproduction notes, a focused patch, and regression evidence |
| Validation tooling | Explicit checks, recorded results, and documented limits |
| Engineering production | Scope, dependencies, blockers, and an actionable handoff |
| Technical documentation | Setup instructions, test procedures, and concise maintainer notes |
Estra Logics is the platform; CML — Continuity Markup Language — is the underlying language and engine. Fermata is the working method around evidence and accountable continuation. Development is human-led and AI-assisted. Failed attempts and corrections stay in the work record, and upstream authorship and licenses remain intact.

