The loop is a pumped, lossy resonator — a laser, not an echo chamber

The principle, stated as physics not metaphor

A mirror cavity that only bounces is a resonator, and a resonator does two real things — neither of them "create energy":

1. It selects. Round-trips reinforce the modes that constructively interfere

(standing waves that fit the room) and cancel the rest. A smear of light becomes a few sharp, coherent modes. That is coherence + selection, not amplification.

1. It stores with loss. Real mirrors leak; "bounce forever" is really "bounce

until it decays." The cavity's Q is just how slowly it loses.

A cavity lases — light that grows on each pass — only with a gain medium and an external pump, bled by a leak. The round-trips don't make the energy; the pump does, and the leak is what keeps the beam anchored instead of runaway. Pull the pump and the most perfect mirror room rings down to dark. Seal the leak and it is not a laser — it is an echo chamber that converges, beautifully, on nothing.

The exact mapping to the Convergence loop

The whole difference between a laser and an echo chamber is the last row. A cavity with gain but no loss term does not amplify signal — it amplifies runaway internal consistency: maximally coherent, maximally self-reinforcing, anchored to nothing. That is exactly the confident-but-unanchored axis of the groundedness canary ([[sigma0-two-canary-axes]]) — i.e. hallucination. Recurrence without a pump and a leak is the precise mechanism the North Star forbids; recurrence with them is "high exploration + mandatory verification" restated in physics.

The concrete leak handle already exists — and is unspent

The loss term is not hypothetical. It is the per-token surprise signal:

• surprise_i = −log₂ p(token_i) [bits] — the per-token code length — implemented in

token-surprise.js and wired into the groundedness canary as signals.modelUncertainty (groundedness-canary.js:148). High surprise concentrated on content tokens = the model was guessing the specifics it stated fluently (token-level reading of semantic-entropy confabulation, Farquhar et al., Nature 2024). It is the one live carry-forward from the E1 kill doc (docs/research/2026-06-28-csf-tesseract-novelty-and-e1-kill.md §7 — "the lapse FIELD is real even though depth-as-storage was killed").

But the valve is closed. No production caller parses provider logprobs into it — fromOpenAILogprobs / fromOllamaLogprobs are referenced only by the module and its tests, so modelUncertainty is always 0 in the live loop. The leak is installed and plumbed to the canary, but no energy bleeds through it. The resonator is, today, a sealed room.

Why our owned local model is the cavity that can lase

The leak only opens where we can read the light leaving the mirrors — the per-token logprobs. By that module's own design note, Anthropic exposes none (graceful no-op). A local PLT decode stream emits them natively. So the model verified under ADR-0011 is not merely the resonator (recurrence); it is the only member whose decode we control well enough to open the leak. Owning the model is what lets us close the loop: recurrence (resonance) + Observe (pump) + surprise-damped Verify (leak), in one cavity we own end to end.

The actionable hypothesis (do NOT build yet — Stage-0 faithful parity first)

Opening the leak — feeding real per-token surprise into the groundedness canary so it continuously damps confident-but-unanchored generation — raises verified-pass-rate and lowers the groundedness-canary firing rate on held-out tasks, versus the same loop with the valve closed (modelUncertainty = 0).

• Pump (Observe): evidence retrieval already feeds the loop; keep it the energy source.
• Leak (Verify): parse the local decode's logprobs → surpriseField → modelUncertainty;

let it down-weight / re-route resonant-but-surprised spans, not just score them.

• Guardrail: the leak must never be the only term — a loss term with no pump just

damps to silence. Resonance is kept; the room is simply not sealed.

This is a Reason/Verify research thread that rides the existing canary + surprise primitives and the owned looped model. It adds no subsystem. Capture the resonance; do not seal the room.

Evidence