kevin-liu-01 cotcodec — working here
Purpose
CoTCodec is a research program and executable evaluation harness for making agent-orchestration choices explicit, measurable, and optimizable. Paper 1 studies language; the shared harness and evidence ledger cover the broader orchestration-variable program.
Mental model & key files
memory.jsonis compiled project state and the active priority ledger.wiki/log.mdis the append-only operational timeline.directions/owns research hypotheses;experiments/owns preregistered runs.harness/executes conditions and benchmarks;scripts/validates and seals.research/evidence/contains portable decision bundles. Raw/local exhaust belongs under ignoreddata/, never in Git by accident.README.mdis the human entry point;AGENTS.mdis the complete operating reference; directorySKILL.mdfiles own local procedures.
Patterns to follow / invariants
- Read
wiki/SOUL.md,wiki/USER.md,wiki/HEARTBEAT.md, andmemory.jsonbefore substantive work. - Preregister falsifiers, budgets, claim boundaries, and stop conditions before observing treatment results. Never rewrite a completed experiment contract.
- Preserve negative and pre-result evidence. Create versioned reruns instead of overwriting failed output directories.
- Separate deterministic infrastructure admission from live-model or scientific claims. CPU conformance never implies H100 admission or memory quality.
- Bind source revision, tree, license, dependencies, image/model identity, and execution hashes for any result used in a research decision.
- Keep
memory.jsoncompiled truth andwiki/log.mdtimeline synchronized.
Common tasks → first action
| Task | First action |
|---|---|
| Continue research | Read memory.json priorities and the latest wiki/log.md entry. |
| Run an experiment | Read skills/run-experiment.md, then its YAML contract. |
| Add a benchmark | Read harness/benchmarks/SKILL.md and the base adapter. |
| Change orchestration logic | Read the nearest harness/**/SKILL.md and paired tests. |
| Add a memory system | Read infra/memory-baselines/SKILL.md and portfolio/source validators. |
| Ship evidence | Run the source, experiment, evidence, and portfolio validators that route the artifact. |
| Update docs | Refresh Agent-Docs, README/current-state pages, memory.json, and wiki/log.md. |
Gotchas
- Many benchmark adapters are intentionally stubs; presence is not readiness.
data/can contain multi-gigabyte models, databases, source trees, and Docker artifacts. Inspect ignored/untracked files before staging.- Exact-source lifecycle jobs often use CPU allocations on GPU hosts. A host name
containing
h100does not prove that a GPU was requested or used. - The historical
~/Documents/GitHub/kevin-wikipath may be absent locally; setKEVIN_WIKI_ROOTforscripts/run-agent-docs.tswhen needed.