Silo to silo.
Two harnesses, opposite depths.

Permission files. Both sides commit a Claude Code settings file, and neither treats it as a policy. Theirs is about 210 allow entries saved from click-approvals, committed once on 2026-08-20, granting push, deploy and credential commands; the permission allowlist held credentials. Ours are tiny but tracked in git, though settings.local.json is meant to stay on one machine. They got that right: their local file stayed local. Neither side has a deny list or a hook.

Dev skills. Their verify skill is a lesson learned the hard way: --only functions alone leaves Firestore calls pointed at production. Our decile-hub-interface is the same kind of guide for a system where a careless agent touches real data. Theirs sits where Claude Code finds it with no setup; ours are found only once linked, and at the snapshot none of augment-it's or memopop's project skills was. On discoverability, their simple path wins.

Commit history is the closest match. 220 of their 263 non-merge commits carry a Claude trailer, with subjects that state findings (Fix the check, not the briefs: five of eight preflight failures were mine). They got lab-notebook commits without a written convention; we wrote git-conventions down so it holds across 27 repos.

Runtime prompt builders on both sides pin behaviour in comments next to the prompt. Their grader runs three passes and averages; our writer is pinned to temperature 0. Two answers to the same worry about variance: theirs measures it, ours suppresses it.

Model routing sits behind one module on both sides; they keep the choice in runtime config, we keep it in code. Run records converge: the flight recorder and our artifact trail both write every step to disk. Don't tune on reasoning alone. Get the flight-recorder data first.

Tests of the LLM machinery exist on both sides. Their fixtures README, exam-craft edits must never break platform tests, marks the boundary between the two layers better than anything on our side.

Specs. Their dev layer is not spec-less: Design doc for review. Nothing gets built until this is signed off. The architecture doc has dated "what shipped" sections, a "Start here in a fresh session" section and epistemic tags. The discipline matches ours; the difference is structure. Their docs/ is flat and unindexed, and nothing points an agent at it.

Handoffs, browser drives, to-do lists. Their session opener carries platform to-dos but exists for exam sessions. Their browser drives run from the scratchpad and are thrown away; we codified ours in a blueprint and committed proof scripts. Both sides let to-do lists rot.

Runtime rulebooks. Their rulebook names itself the single source of truth (revise this section, not individual exams). The orchestrator's AGENTS.md has numbered principles that the code cites. Ours carries a naming hazard: in the tree, AGENTS.md also means coding-agent instructions.

Craft ledgers. Their EXAM_CRAFT.md is one 2,163-line file: stable IDs, an evidence line per finding, a part for candidate principles, not yet distilled, an append-only build log, a checked index. Our product craft is spread across 36 issue-resolution and reminder files. Ours is easier to publish; theirs does the job a ledger exists for.

Assemble, don't generate, the strongest convergence. They copy structure from a skeleton and validate invariants that are silently violable. The orchestrator fixes section taxonomy in an outline YAML with a schema because drift made merging seven ChromaDB runs produce 26 best-of catalogs instead of 9.

Generation commands and human gates. /exam-build enforces its tool list and two stops (Do not deploy, commit, or import anything into the builder. The user does that.); our memo skills cover more situations but only describe their stops. Both sides gate generation on a human, at different points.

Grading the output. Their audits check whether the model did what it was told; our five checking agents and scorers check whether the output is true and good. What we lack is their next step: running the checks many times and reporting the spread.

Pseudomonorepo rules, the shared skills library, the Chroma corpus exist because of our shape: 27 projects, nested submodules, parallel sessions. In one repo with 280 commits, git log does what the corpus does. Spec Kit is different but a candidate overbuild on our side. Legacy per-harness files are leftovers, not a choice; their single place for instructions is the better state.

Deny-all database rules, on their side, force every write, an agent's included, through the API: a structural guardrail where we would use a prose rule.

Firm-private submodules fence each client's inputs and outputs in its own repo. Their customer content lives in the app repo, which is fine for one customer.

A repo-wide instruction file. Nothing maps the API routes, libraries or the 23 ordered front-end scripts, and nothing says the old 12,897-line script is dead. With 220 agent-trailered commits, the cost of a cold start recurs. CI: their dev gates run only when remembered. A changelog, low priority: commit bodies are their only running dev memory, and one commit folded 16 days of work done "outside git".

On our side, the main lesson. A simulator plays 18 personas, from native-phd down to stalled, through a real exam and reports a matrix, the correlation with tier, a standard error per candidate and a resolution floor. A gate replay measures how often a decision flips on identical input (flip rate 2% (1/63)). Preflights check that a persona plays its brief and run the shipped prompt against the live model: A result here is a statement about production. None of this exists for memos.

On their side, an exception. /exam-build and /studio implement the same flow twice with no declared source of truth. augment-it wrote that rule at the point of use: the SKILL.md is the source of truth, this slab is its always-loaded condensation; update together.