{
  "stages": [
    {
      "id": "setup",
      "label": "Setup"
    },
    {
      "id": "baseline",
      "label": "Stage 1: Baselines"
    },
    {
      "id": "diagnosis",
      "label": "Stage 2: Triage and oracle test"
    },
    {
      "id": "memory",
      "label": "Stage 3: Notes, memory and search"
    },
    {
      "id": "exec-memory",
      "label": "Stage 4: Executable memory"
    },
    {
      "id": "strategy",
      "label": "Stage 5: Self-improving strategy"
    },
    {
      "id": "weights",
      "label": "Stage 6: Weight updates"
    }
  ],
  "entries": [
    {
      "date": "2026-10-17",
      "stage": "setup",
      "kind": "milestone",
      "title": "HAI started",
      "note": "Can a cheap model get measurably better at coding over time, without retraining, by remembering what worked and checking that memory against reality?",
      "details": [
        "Coding first, because tests give a clear pass or fail. Without a reliable signal like that, an agent can't tell which of its lessons were right."
      ]
    },
    {
      "date": "2026-10-17",
      "stage": "setup",
      "kind": "decision",
      "title": "Harness: fx, kept fixed",
      "note": "fx is small, scriptable and embeddable, which makes it easy to run thousands of tasks. The harness itself won't be self-edited.",
      "details": [
        "fx is written in Zig, so a self-edit would mean a compile step that can fail. Self-improvement happens in a layer of prompts, skills and tools on top, versioned in git and easy to roll back."
      ]
    },
    {
      "date": "2026-10-17",
      "stage": "setup",
      "kind": "decision",
      "title": "Model: DeepSeek V4.1 Flash through the OpenCode API",
      "note": "Strong on coding benchmarks and very cheap to run, which matters when every experiment means hundreds of tasks.",
      "details": [
        "Its weights are open, but at 552B total parameters it isn't practical to fine-tune at home. Weight updates move to a separate stage with a small local model."
      ]
    },
    {
      "date": "2026-10-17",
      "stage": "setup",
      "kind": "direction",
      "title": "Memory should be code, not just notes",
      "note": "A memory stored as a test or probe can be run before it's trusted, and retired when it starts failing. Added as its own stage to compare against prose memory.",
      "details": [
        "Prose memories go stale silently. Runnable ones tell you when they've stopped being true."
      ]
    },
    {
      "date": "2026-10-17",
      "stage": "setup",
      "kind": "decision",
      "title": "How progress will be measured",
      "note": "Ordered task stream; every 25 tasks, a score on held-out tasks with memory frozen. Tracked: pass rate, regressions, steps per solved task, memory size.",
      "details": [
        "Stopping rule, set in advance: if memory doesn't add at least 3–5 points on held-out tasks after about 200 tasks, the core idea is weak."
      ]
    },
    {
      "date": "2026-10-17",
      "stage": "setup",
      "kind": "learning",
      "title": "Expect small gains from memory alone",
      "note": "Google's ReasoningBank added 4.6 points on SWE-Bench Verified. SICA's self-edits went from 17% to 53%, but with frontier models doing the reflecting.",
      "details": [
        "If HAI matches the first, memory is worth keeping. The bigger question is whether executable memory and strategy edits add more on top."
      ]
    },
    {
      "date": "2026-10-18",
      "stage": "setup",
      "kind": "direction",
      "title": "The goal is cost, not just accuracy",
      "note": "The aim: near-frontier results on a codebase at a fraction of the frontier's cost. The main number is now cost per solved task, shown next to accuracy.",
      "details": [
        "A harness that adds 5 points but triples the tokens per task could lose on cost. So every run records the full cost, including retrieval, search, verification and retries.",
        "DeepSeek's own numbers put bare Flash already close to the frontier at low cost. If that holds here, HAI's contribution is only the remaining gap, so a frontier model is measured on the same tasks as a reference."
      ]
    },
    {
      "date": "2026-10-18",
      "stage": "setup",
      "kind": "direction",
      "title": "Focus on one codebase at a time",
      "note": "Memory should work like a veteran on a codebase: the longer it works there, the better it knows where things live and what breaks. So tasks now run as a stream per repository.",
      "details": [
        "The test is whether pass rate on later tasks in the same repo rises as the stream goes on. Public repos prove the method; a private or very recent one is the cleaner proof, because the model can't have memorised it."
      ]
    },
    {
      "date": "2026-10-18",
      "stage": "setup",
      "kind": "decision",
      "title": "Memory has to beat a plain notes file",
      "note": "A hand-written notes file is the cheap, obvious alternative. A static-notes arm is added, so HAI memory must beat it, not just beat having nothing.",
      "details": [
        "Each memory item will record the files it was learned from and be re-checked when they change, so stale lessons are dropped instead of trusted."
      ]
    },
    {
      "date": "2026-10-18",
      "stage": "setup",
      "kind": "decision",
      "title": "Test the idea cheaply before building it",
      "note": "Two checks come first: sort the baseline's failures into missing knowledge versus weak reasoning, then give the model the right hints on those tasks and see how many it fixes.",
      "details": [
        "Search and memory add knowledge, not reasoning. If most failures are reasoning failures, or perfect hints barely help, the approach has a low ceiling and we learn that before building any memory."
      ]
    },
    {
      "date": "2026-10-18",
      "stage": "setup",
      "kind": "learning",
      "title": "Local runs work; terms need a check",
      "note": "fx runs against the model through a small local proxy, and a first 10-task trial is under way. OpenCode's terms are unclear on benchmark-scale automation, so we're asking them before any large run.",
      "details": [
        "The terms also restrict using output to train competing models, which could affect the weight-update stage. Training data will come from sources that allow it."
      ]
    }
  ],
  "next": {
    "title": "Run the two baselines",
    "note": "Bare Flash and one frontier model on the same tasks, with cost per solved task recorded, three runs each for noise. Then triage the failures and run the oracle test."
  }
}