anti
e52a0e0381
BEHAVE-EXTRACTOR.md Phase A Step 5. Classifies the operator's
thinking pace between commands. Splits LW-sim / CLAUDE-FF /
CLAUDE-CL.
* _features/cognitive.py:inter_command_latency_class(ctx) emits one
Observation in {instant, typing_speed, deliberate,
llm_lightweight, llm_heavyweight, long}, computed as the median
of ctx.inter_cmd_iats bucketed against the prototype thresholds
(v0.2 split: lightweight 2-8s, heavyweight 8-30s).
* Sample-size honesty: < 5 commands halves confidence (0.40 vs
0.80) per BEHAVE-EXTRACTOR.md.
* Threshold consts (INTER_CMD_*_MAX, MIN_COMMANDS_FOR_FULL_CONFIDENCE,
plus parked Step 6/7/8 thresholds for the next three commits)
added to _thresholds.py.
Tests cover all six buckets at empirically-anchored IATs (15s ≈
Claude Opus driving recon via tmux send-keys), plus the
single-command no-IAT and low-sample-count paths.
81 lines
4.5 KiB
Python
81 lines
4.5 KiB
Python
"""Numeric thresholds for BEHAVE-SHELL primitive classification.
|
|
|
|
Each constant cites its calibration source. When the registry's
|
|
``notes:`` field disagrees with a constant here, the registry is
|
|
authoritative — fix the constant, re-run the calibration grid.
|
|
|
|
Empirical thresholds inherited from the BEHAVE prototype extractor
|
|
(``BEHAVE/prototype_extractors/shell/extract.py``); see lines 40-90 of
|
|
that file for the calibration history. Any change here must keep the
|
|
five-class grid green.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
# ── paste-burst detection (Step 1) ──────────────────────────────────────────
|
|
# A single input event with ≥ PASTE_MIN_CHARS_PER_EVENT chars is the
|
|
# paste-class proxy used by the prototype; xterm-kitty / iTerm / VS Code
|
|
# pastes arrive as one bulk write.
|
|
PASTE_MIN_CHARS_PER_EVENT: int = 4
|
|
|
|
# Consecutive paste-class events arriving within this IAT collapse into
|
|
# one PasteBurst record. 200ms is the prototype's IKI burst cap.
|
|
PASTE_BURST_MAX_IAT_S: float = 0.20
|
|
|
|
# ── motor.input_modality (Step 2) ───────────────────────────────────────────
|
|
# Paste-event ratio thresholds. ≥ 40% paste events → "pasted" (LLM-driven);
|
|
# ≤ 5% → "typed" (human at the keyboard); in between → "mixed".
|
|
# Lowered from 0.5 after the 47.6% case in sessions-2026-05-02-with-llm.jsonl
|
|
# was clearly LLM-driven but missed the 0.5 floor.
|
|
MODALITY_PASTED_MIN: float = 0.40
|
|
MODALITY_TYPED_MAX: float = 0.05
|
|
|
|
# ── motor.paste_burst_rate (Step 3) ─────────────────────────────────────────
|
|
# Same paste-event ratio re-bucketed for the "how often does the operator
|
|
# paste" axis. Coarser than input_modality on purpose: this primitive is the
|
|
# habit signal, input_modality is the dominant-channel signal.
|
|
PASTE_RATE_HABITUAL_MIN: float = 0.50
|
|
PASTE_RATE_OCCASIONAL_MIN: float = 0.10
|
|
|
|
# ── cognitive.inter_command_latency_class (Step 5) ──────────────────────────
|
|
# Bucket edges (seconds) for the median inter-command IAT. Prototype
|
|
# values; v0.2 splits the original llm_roundtrip 2-8s band into
|
|
# llm_lightweight (orchestrated agents w/ small models / terse prompts) and
|
|
# llm_heavyweight (reasoning-class agents in tool loops with text
|
|
# generation between calls). Empirical anchor: Claude Opus driving recon
|
|
# via tmux send-keys produced a median of 15.5s.
|
|
INTER_CMD_INSTANT_MAX: float = 0.30
|
|
INTER_CMD_TYPING_MAX: float = 1.50
|
|
INTER_CMD_DELIBERATE_MAX: float = 2.00
|
|
INTER_CMD_LLM_LIGHTWEIGHT_MAX: float = 8.00
|
|
INTER_CMD_LLM_HEAVYWEIGHT_MAX: float = 30.00
|
|
|
|
# Sample-size floor for inter-command IAT primitives. Below this we
|
|
# halve the confidence per BEHAVE-EXTRACTOR.md "sample-size honesty".
|
|
MIN_COMMANDS_FOR_FULL_CONFIDENCE: int = 5
|
|
|
|
# ── cognitive.command_branch_diversity (Step 6) ─────────────────────────────
|
|
# unique_first_tokens / total_commands ratio. Empirical (CLAUDE-FF vs
|
|
# CLAUDE-CL on 2026-05-02): fire-and-forget runs ~10 distinct tools (ratio
|
|
# near 1.0) → linear_playbook; closed-loop runs ~5-6 tools with the same
|
|
# tool re-invoked → adaptive_branching.
|
|
BRANCH_DIVERSITY_LINEAR_MIN: float = 0.80 # >= → linear_playbook
|
|
BRANCH_DIVERSITY_ADAPTIVE_MAX: float = 0.60 # <= → adaptive_branching
|
|
# Between is the ambiguous middle band — bias toward adaptive (the
|
|
# operator is reusing tools).
|
|
|
|
# ── cognitive.feedback_loop_engagement (Step 7) ─────────────────────────────
|
|
# Pearson r threshold for "the operator's pause grew with the volume of
|
|
# preceding output". |r| > this → significant; sign carries direction.
|
|
FEEDBACK_CORRELATION_MIN: float = 0.30
|
|
# Need at least this many (output_bytes, next_pause) pairs to even
|
|
# attempt a correlation. Below this the answer is "unknown".
|
|
FEEDBACK_MIN_PAIRS: int = 5
|
|
|
|
# ── cognitive.inter_command_consistency (Step 8) ────────────────────────────
|
|
# CV (stdev / mean) of inter-command IATs. Empirical (this corpus):
|
|
# human session CV=0.94 → variable; LLM-simulated CV=0.24 → metronomic;
|
|
# anything beyond 1.5 is heuristically "bimodal" (real bimodal detection
|
|
# via Hartigan dip is filed for v0.2).
|
|
PAUSE_CV_METRONOMIC_MAX: float = 0.40
|
|
PAUSE_CV_BIMODAL_MIN: float = 1.50
|