[Alajmah]

Summary

This stacked PR adds the first structured architecture manifests to ModelMap, with field-level provenance and repository-level validation.

Architecture schema

• binds every architecture record to one model_release or model_variant through target
• adds review status, Gemma-compatible pre/post normalization flags, structured attention fields, and structured feed-forward/MoE fields
• adds field-level evidence entries (source, narrow locator, exact fields)

Validator hardening

• resolves architecture targets and evidence sources against canonical IDs
• requires every populated architecture fact to be covered by valid source evidence
• rejects evidence that names an absent or null fact
• validates zero-based inclusive block plans for ordering, bounds, and overlap

Initial manifests

Adds reviewed manifests for:

• Mistral 7B
• Gemma 2 9B
• DeepSeek-V3
• Llama 3.1 405B

The manifests use primary technical reports and a pinned first-party Meta SKU registry where exact checkpoint configuration matters.

Llama 3.1 405B ambiguity

attention.kv_heads is deliberately absent from the model-variant manifest. Meta's first-party SKU registry exposes n_kv_heads=8 for BF16 MP8/FP8 packaging and n_kv_heads=16 for BF16 MP16 for the same 405B core model. Encoding either value as intrinsic to variant:llama-3.1-405b would overstate the evidence. The manifest records the ambiguity in notes instead of inferring a value.

Tests and documentation

Adds architecture-specific regression tests for schema requirements, reference integrity, complete evidence coverage, invalid evidence paths, block-plan invariants, and the Llama KV-head omission policy. The data and ontology documentation now define architecture targeting, evidence semantics, block ranges, and omission of packaging-dependent values.

Review

The branch was reviewed at HEAD before opening this PR, including schema semantics, validator failure modes, data/source alignment, test coverage, CI integration, and documentation consistency. The existing validation workflow automatically runs the full unittest suite, corpus validator, and generated-view builds on the PR.

Stack

Base: atlas-generated-views-v0.1 (PR #10)
Head: atlas-architecture-manifests-v0.1

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a3dc8b0db6

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Guard non-list evidence before iterating

When an architecture document contains a schema-invalid value such as evidence: null, schema validation records the expected error but execution then reaches this loop and raises TypeError: 'NoneType' object is not iterable. This prevents the corpus validator from reporting the collected validation errors and makes a malformed architecture record crash validation instead of failing cleanly; check that evidence is a list before iterating it.

Useful? React with 👍 / 👎.