World model (AGI panel)

Navin learns to anticipate what a tool will answer before calling it: "in this repo, npm test fails", "git push here is denied", "that API answers 404". Not to talk better, not to write a skill, not to train the chat LLM. Three sentences say everything that matters:

1. Off by default. With the flag off there is no tool journal, no training job, no advice in any prompt: a chat turn costs exactly what it costs today.
2. Journal and training are automatic. With the flag on, every tool call leaves one secret-free line in .navin/world/trajectories.jsonl after the call, on a background thread, and a small local head (counts, not a neural network, never the chat model) trains outside the chat once enough new lines arrived. Its judge is a number: the prediction error on a frozen set of calls it never saw. Down means the checkpoint is kept as evidence and nothing serves it.
3. Live advice is a gate plus you. advise cannot be switched on until the frozen exam says up and an offline A/B on that same set shows a gain (useless calls spared, no false alarm). Even then the advice is three lines at most and never replaces a call; it switches itself off and rolls back the head if its live precision drops.

Everything lives in the AGI entry of the Code workbench rail, right after Terminal, between skills evolution and memory. S1 (episodic memory) is useful, not required: a trajectory links to the episode of its turn when S1 runs, and logs anyway when it does not.

The project flag

// <project>/.navin/world-model.json
{
  "schema_version": 1,
  "enabled": false,
  "log": true,
  "train": true,
  "advise": false,
  "beliefs": false,
  "skip_hint": false,
  "confidence_threshold": 0.8,
  "train_every": 200,
  "min_rows": 40
}

Classes

Every observation is one of ok, changed (a write succeeded), empty, not_found, denied, timeout, error. not_found, denied and timeout are the useless calls: the world did not move and the answer was a refusal. Those are what advice tries to spare.

The exam

About one call in five is held out by the hash of its id, never trained on. The first training freezes those rows into .navin/world/heldout.jsonl with a version (hash of the row ids). Every score is computed on that file and re-hashes it first: an edited, trimmed or reordered set fails the run, the exam and the advice gate with tampered. Lowering the number by reworking the exam is the cheat S2 forbids for skills; the same rule holds here. navin agi world freeze makes a new version (a human action); scores of different versions are never compared.

Metrics: log-loss (the error of the predicted distribution), wrong-class rate, calibration (ECE). The head is compared with the stronger of two baselines, "always ok" and "majority class": up when its log-loss is at least 2% lower without more wrong classes, down when it is worse, flat in between. A new checkpoint is activated only when it is up vs the baseline and not down vs the head that serves. Rollback goes back to N-1; the newer checkpoint stays on disk as evidence.

Training runs under a budget (wall clock, memory growth, rows). Over budget means the run fails and nothing on disk moves.

Advice, the A/B and the kill switch

The offline A/B replays the frozen set: A makes every call and pays the useless ones in full; B is told "useless" when the head is confident, spares the call when it really was useless and is charged a false alarm otherwise. gain needs at least one useless call in ten spared with 90% precision; under 80% precision it is regress.

Live, every suggestion is compared with the observed class after the real call (.navin/world/live.jsonl). When the rolling precision over the last 30 advices falls under 80%, the advisor cuts itself: advise goes back to off, the head rolls back to N-1, the journal says why, the world_predict tool leaves at the next sync and the prompt is exactly what it was. A human has to switch it on again, and only once the gate reopens.

The running gateway aligns the world_predict tool with the gate at boot, on every AGI panel read or write, and after every action. A flag flipped from the CLI (navin agi world set advise on) therefore reaches the serving agent the next time the panel is opened; the panel says "world_predict tool live in the agent" when it did, and shows the gate's reasons when it is closed.

Where things go

From the terminal

navin agi status                     # skills evolution, world model and memory switches
navin agi world status               # flag, journal size, frozen set, active head, score, gate
navin agi world on | off             # master switch (advice stays off)
navin agi world set beliefs on       # log | train | advise | beliefs | skip_hint | knobs
navin agi world set advise on        # refused while the gate is closed
navin agi world log -n 20            # tail of the tool journal: class, key, duration
navin agi world train                # human: train a checkpoint now
navin agi world run                  # what the runner does after a turn, if due
navin agi world exam                 # re-score the active head on the same frozen set
navin agi world ab                   # offline A/B of the active head
navin agi world freeze --yes         # human: new held-out version
navin agi world checkpoints          # every checkpoint, the active and the previous one
navin agi world scoreboard           # baseline -> head, verdict, per run
navin agi world rollback             # back to N-1
navin agi world beliefs [--render | --discard <key> | --restore]
navin agi world journal -n 30        # trained / rollback / advise_cut ...

Every command takes -p/--project <folder> (default: current directory) and the readers accept --json.