Frontier Model Watch

frontier-model-watch

The model-side companion to ai-coding-tools-compare. Tools matter; models matter; the intersection is where the org's stack actually lives. This skill keeps the frontier-model dimension grounded specifically in "what changes for the org's Copilot deployment / AI tool posture."

When to use

• Major model release (Anthropic Opus 4.x → 4.y, OpenAI GPT-5.x → 6.x, etc.).
• Model deprecation announcement (e.g., Opus 4.6 fast 2026-06-29).
• Model availability shift (e.g., Fable 5 / Mythos 5 US export-control situation 2026-06-12).
• Benchmark publication that reframes per-model strengths.
• Annual frontier-model landscape summary.

When NOT to use

• Specific Copilot question → copilot-faq-answerer.
• AI tool comparison → ai-coding-tools-compare.
• AI governance framework → ai-governance-research.

Obsidian-first workflow (mandatory)

1. Query the vault first via vault-querier:
   • Full-text search the model / vendor key terms across vault/research/frontier-model/**, vault/facts/{model-or-vendor}/** (e.g. anthropic, openai, google), vault/facts/copilot/** (model-routing facts live there), and recent vault/digests/** (last 90 days — weekly-intelligence-digest and quarterly-ai-coding-landscape carry model news).
   • Backlink check on the model / vendor entities (e.g. [[anthropic]], [[gpt-5]]).
2. Triage findings:
   • If the vault answers the question fully → return the existing answer with source citations (vault path + original source URLs from the fact's frontmatter). No new write.
   • If partial → identify the gap. Web research targets only the gap.
   • If empty → full web research.
   • A gap means the vault has no note ≤90 days old answering the question.
3. Web research (only on confirmed gaps):
   • Use source-fetcher (with prompt-injection-guard) on tier-1 sources: each lab's own announcement / docs pages (tier-1 for THEIR model, tier-3 for cross-lab claims), the GitHub Copilot changelog for routing changes, official deprecation notices.
   • Extract claims via claim-extractor.
4. Verify load-bearing claims via verify-claim (3-vote refute):
   • A vendor's claims about its own model are tier-1 (no verification); benchmark interpretations, cross-lab comparisons, and availability/export-control analysis get the full 3-vote treatment.
5. Write the research note via digest-writer (which delegates the file write to vault-writer.write_research):
   • Path: vault/research/frontier-model/YYYY-MM-DD-{slug}.md
   • Frontmatter per research.yml schema: topic: frontier-model, question, sources, findings_count, verified_claims.
   • Body: TL;DR + Findings (with quoted anchors) + Sources (with credibility-tier badges).
6. Stage promotable claims to _inbox/frontier-model-watch/:
   • Any verified fact-typed claim (e.g. a deprecation date, a routing change) → _inbox/frontier-model-watch/{timestamp}-{slug}.md with suggested_surface: facts and suggested_path: facts/{model-or-vendor}/{predicate}.md.
   • memory-curator decides on its next sweep.

Topic taxonomy

For the research note's topic field, use frontier-model. Vault facts under vault/facts/{model-or-vendor}/.

Compliance-relevant framing per finding

For each model finding, answer:

1. Does this affect what Copilot will route to for this org? When?
2. Does this change the AI-tool TPRM file?
3. Does this require a model-risk review (if the org's posture includes one for AI tools)?
4. Is there a deprecation date that triggers user-facing communication?

Composes with

Standard Phase-1 foundation. Cross-feeds:

• weekly-intelligence-digest — recurring frontier model news.
• ai-coding-tools-compare — model side of the tool comparison.
• copilot-faq-answerer — model-selection canonical answer needs updates when models shift.

Acceptance test (for step 28 done-criteria)

One research note exercising a specific frontier-model topic with the compliance-relevant framing applied per finding.