Original, clean-room, dependency-free Node.js implementation of the technique described in ../DEFENSIVE-PUBLICATION.md.
This is illustrative prior-art code, not production code. It reproduces no proprietary source, performs no I/O, requires no external packages, and contains no secrets. Its only purpose is to enable the disclosure — to show that the invention reduces to practice.
| File | What it shows |
|---|---|
band-router.js |
The full mechanism: a per-action-type band store, the routing trichotomy (route → AUTO / ESCALATE / REVIEW), outcome recording into a recency-weighted sliding window, and rolling recalibration with clamps, hysteresis (maxStep), minWidth, and a minSamples gate — all configurable globally and per action type. |
example.js |
A runnable scenario and self-check: drives chat (high success) and email (low success) through outcome streams, prints how each band moves and how a fixed confidence routes, and asserts invariants R1, R6, R7, R11. |
Requires Node.js 18+ (uses only the standard library).
cd src
node example.jsExpected: the chat band loosens (bounds drop → more AUTO), the email
band tightens (bounds rise → more REVIEW/ESCALATE), the two types stay
isolated, invalid confidence routes to REVIEW, and the script prints
All self-checks passed.
| Claim element | Code |
|---|---|
| Per-action-type bands (Claim 1, 3, 15) | states: Map<actionType, BandState>, cfgFor, stateFor. |
| Routing trichotomy (Claim 1, 9) | route() returning AUTO/ESCALATE/REVIEW. |
| Outcome + sliding window (Claim 1, 4) | recordOutcome() with recency-decayed ring buffer. |
| Human-free recalibration, tighten/loosen (Claim 1, 2) | recalibrate() with the signed delta rule. |
| min-samples gate (Claim 5) | if (window.length < minSamples) .... |
| recalc every M (Claim 6) | sinceRecalc >= recalcEvery. |
| clamps + ordering (Claim 7) | clamp, enforceOrdering. |
| hysteresis / max-step (Claim 8) | boundedStep. |
| audit hook (Claim 10, 11) | audit('recalibrate', …), getStatus/getAllStatus. |
| seed unseen type (Claim 12) | seedLow/seedHigh defaults in stateFor. |
| fail-safe (Claim 13) | invalid confidence → REVIEW. |
| emergent gradient (Claim 14) | demonstrated by example.js divergence. |