v0.8.1 release notes

v0.8.1 generalizes how Modulatio puts a multi-piece deliverable together. The old mechanical assembly could concatenate a document; now a family of product-aware assemblers joins the actual product — text, code, or data — each by the operation its byte-nature demands. Underneath sits a new content-addressed review-ledger so QC can pass a large finished deliverable cheaply — by verifying the marks, not by re-reading every byte. Cleared fresh hull + coherence reviews (eight hull holes found and sealed across two close-out rounds).

Still Beta — the Beta calibration still applies

v0.8.1 is a hardening + generalization release; it does not move the engine’s ceilings. The honest limits on the Beta calibration page still hold. This page is the delta on top of v0.8.0.

---

Familial assemblers — join the product, not just the document

A multi-piece run produces units (chapters, modules, records) that have to become one deliverable. v0.8.0 and earlier did that one way: concatenate text. That is correct for prose and wrong for code (you don’t cat source files into one blob) or data (rows want a real merge, not a paste). v0.8.1 splits assembly into a dispatch of families, chosen by the artifact’s kind:

• document-assembly — ordered text concatenation with framing (title / separator / trailer). Prose, reports, forms, application packets. The original behavior, generalized.
• code-assembly — preserves the file tree and generates a wiring index (title, file list, entry point). The files stay separate on disk, byte-for-byte; the deliverable is the index, not a concatenated megafile.
• data-assembly — a real merge / fold: JSON arrays concatenated, CSV rows unioned, with strict parsing, dedupe, and a hard output cap.
• media-assembly — a registered seam that fails closed until its render tool lands (image / audio / video compositing is a later tier).

The family is selected by the product’s kind through the standards file (assembler_skill frontmatter) — there’s no engine routing table to drift, and the planner no longer hard-codes a single assembler. Users can add their own assembler families (a Blender assembler, a CAD assembler) without touching the engine.

Why the engine owns the join

The producer emits a small plan — a manifest naming the units and the framing — and the engine does the bulk copy, reading unit bytes from disk. Unit bytes never round-trip through the model as output tokens, so a large deliverable can’t truncate at the model’s output cap (the failure where a six-story anthology came back as two). It’s the QC / speculative-decoding thesis applied to the gather phase: cheap to emit the plan, mechanical to execute it.

---

The review-ledger — verify the marks, not the bytes

Re-reading a finished 50-page deliverable back into QC to “check it” blows the QC budget, forces a compressed partial view, and false-rejects complete work. The new content-addressed review-ledger records which bytes a QC pass blessed (qc_passed_checksum). Assembly QC then verifies structurally:

• every named unit was already QC-passed, and its on-disk bytes still match its mark;
• the manifest’s unit set equals the authoritative dependency set (no missing, no extra, no duplicate);
• the engine-recorded assembly is complete and untampered.

No LLM re-read of the assembled whole — the marks already carry the verdict. A complete book now passes its assembly review cheaply instead of being re-litigated every retry. (The code family is deliberately not eligible for this cheap pass yet — there’s no deterministic wiring validation, so code assemblies always get a full review.)

---

Also in this release

• No-regress guard — a drifted retry can no longer clobber a complete, QC-passed deliverable with a smaller or worse one; re-opens build in place (edit / revise) rather than full-regenerating, so good work isn’t thrown away.
• Graceful delivery without a renderer — if neither pandoc nor pypandoc is installed, products now ship as Markdown with a visible note (and a prompt to install pandoc for DOCX/PDF) instead of failing to deliver at all.
• Sandbox — the producer sandbox binds the virtualenv read-only so producer code can actually run python3 under the standard profile, without unmasking the home tree or exposing secrets.
• Codified-skill hygiene — a stale shared codification can no longer bury a seed-skill improvement, and an explicit user_override marker keeps a human edit sacred against re-codification.

---

Under the hood

A new assembly module holds the strategy dispatch (document / code / data / media); a review_ledger module holds the content-addressed marks and the cheap structural check. The engine wires an assembler task’s inputs from the task graph (not the producer’s manifest), pre-filters manifest units to the authoritative dependency outputs before any read, and records an engine-authored assembly proof — so a producer emitting assembled-looking text can’t bypass review. Every existing path is unchanged; document runs assemble exactly as before.