CLI / REPL

The qriton-hlm command launches an interactive REPL for basin surgery.

Usage

# Interactive mode
qriton-hlm

# Load a checkpoint on start
qriton-hlm --checkpoint model.pt
qriton-hlm -c model.pt

# Run a script
qriton-hlm --script program.hlm

# Pipe commands
echo "survey 0" | qriton-hlm -c model.pt

# Other flags
qriton-hlm --version       # Print version and exit
qriton-hlm --no-color      # Disable ANSI color output

REPL Commands

Once inside the REPL, type commands directly:

hlm:model> survey 0
  Layer 0: 47 basins found (200 inits, β=7.00)
  B0  E=-23.4521  pop=34  ████████████████░
  B1  E=-19.8734  pop=28  █████████████░░░░
  ...

hlm:model> inject 0 42 0.1
  Before: 47 basins | target is basin: False
  After:  48 basins (+1) | target is basin: True (cos=0.98)

hlm:model> verify 0 42
  Basin: True | cos=0.9823 | energy=-15.234 | iters=23

Parameters

Tunable via the set command:

Parameter	Default	Description
beta	7.0	Inverse temperature (uses learned β if available)
inits	200	Random initializations for basin discovery
strength	0.1	Surgery strength
degree	3	Polynomial degree
eps	0.005	Convergence threshold
temp	0.8	Sampling temperature for generation
tokens	50	Max tokens for generation
topk	40	Top-k sampling

Causal Commands

hlm:model> causal scan 5
Causal scan L5: 7 basins, threshold=0.15
  Causal edges found: 3
    B0 -> B3  drift=0.4231  DESTROYED
    B0 -> B5  drift=0.1823  shifted
    B2 -> B6  drift=0.2104  shifted

hlm:model> causal intervene 5 0
do(B0) = remove  L5
  Basins: 7 -> 6 (-1)
  Downstream effects: 2 basins affected

hlm:model> causal counterfactual 5 2 shift
Counterfactual: B2 -> shift  L5
  (non-destructive -- original landscape preserved)
  No downstream effects predicted.

See the full Causal Programming guide for details.

Guards

hlm:model> guard max-basins 250
  Guard set: max-basins = 250

hlm:model> guard strength-cap 0.3
  Guard set: strength-cap = 0.3

hlm:model> guards
  Active guards:
    max-basins    250
    strength-cap  0.3

hlm:model> inject 5 99 0.5
  BLOCKED by guard: strength-cap (0.5 > 0.3)
  Use --force --reason "justification" to override.

hlm:model> history
  #1  survey 5           OK
  #2  inject 5 99 0.5    BLOCKED (strength-cap)

Guard types: max-basins, min-basins, strength-cap, cosine-drift, perplexity-delta.

Features

• Readline support (arrow keys, command history)
• Checkpoint loading: full HLM3 models or W-only mode
• Operation history tracking
• Backup/restore for all layers