ADR-0018: The proactive emission gate compares virtual and regular pheromone per link, not best-vs-best¶

Status: Accepted — the #180 re-derivation; shipped gate-off by default
Date: 2026-08-02

Context¶

The thesis conditions proactive-ant emission on diffusion having found something worth checking: "only if the best virtual pheromone is significantly better (in our experiments: at least 10% better) than the best regular pheromone, a proactive forward ant is sent out" (Ducatelle 2007, lines 4084–4088). PR #188 shipped that literal scalar form with the thesis's 0.10 margin and measured it harmful — pass rate 0 % (line) to 3.8 % (grid), starved maintenance, +36–68 % NRL and −5 to −6.4 pp PDR — after which the mechanism was kept but the default margin set to 0 (gate off), and #180 owned re-deriving what the thesis actually compares before any non-zero margin could be considered.

Two table-hygiene defects had to fall first, because both corrupted the very quantities the gate compares:

#262 — virtual pheromone aged on an accidental per-hello clock (α^degree per second), depressing it degree-dependently (median v/r 0.49, ~4× monotone fall with degree). Fixed: both tables now age on the identical α^(Δt/interval) clock.
#279 — every reception re-seeded the 1-hop regular entry at the unitless constant 1.0, pinning it ~100× low and corrupting every 1-hop hello advert (the diffusion input). Fixed: receptions refresh at the metric's unloaded 1-hop value.

With both fixed, the uniformity probe (core/tests/exp_uniformity_probe.cpp, PR #278) measured what remains — and what remains is structural, not a calibration error:

The matched per-link v/r ratio is centred 0.97–1.06, degree- and hop-independent, p75 1.20–1.32, p90 1.32–2.19.
The best-vs-best ratio at gate sites has a hard ceiling τ(h−1)/τ(h) = (T̂+h·T_hop)/(T̂′+(h−1)·T_hop) ≈ h/(h−1): the neighbour's advertised best path is one hop shorter than the node's own best path to the same destination, so the scalar ratio never exceeds the one-hop cost ratio. Measured byte-exact: grid 6-hop ceiling 1.2000, line 7-hop 1.1629.

The consequence: a fixed scalar margin m is satisfiable only for destinations within 1 + 1/m hops. m = 0.10 ⇒ h ≤ 11; m = 0.20 ⇒ h ≤ 6 (pass rate measured falling from 35.6 % at k = 1.10 to 0.0 % at k = 1.20 on the 6-hop grid). On the satellite regime — the gate's payoff surface per #248, since MANET benchmark guidance is proactive-off — ISL paths routinely exceed that, so a scalar gate silently disables proactive maintenance for exactly the far destinations that need it most. No single k is right across path lengths: the re-derivation is a shape question.

Decision¶

shouldSendProactive(dest) passes iff some neighbour n satisfies v(dest,n) ≥ (1+m) · r(dest,n), where m is proactiveVirtualMargin, with two qualifications:

A virtual hint on a link whose regular entry sits at/below minPheromone passes trivially — diffusion turned up a link the sampling never priced (unsampled or evicted), which is the genuine "good new virtual pheromone" case and the strongest reason to send an ant. The probe measured this tail reaching 3.5×10⁵ — it is real signal, not noise.
The boundary cases are unchanged from the scalar implementation: no regular route at all → always send (the session needs an ant most); margin ≤ 0 → gate off (the ablation); diffusion off → gate cannot be evaluated, unconditional emission (that ablation keeps meaning "proactive without guidance", not "no proactive").

Comparing v and r on the same link cancels the h/(h−1) systematic — both estimators describe the same path, differing only by the diffusion bootstrap and refresh phase — so the margin finally means what the thesis intended: "the hint is materially better than what I already know", independent of how far the destination is.

The default stays 0 (gate off). The re-derived shape makes a non-zero margin meaningful; whether one becomes a default is a benchmark decision on the satellite regime, made against the A/B record on #180. The probe's distribution puts the useful range at m ≈ 0.2–0.5 (selecting the ~5–15 % genuine-anomaly tail of link-checks).

Alternatives considered¶

Retune the scalar margin. No k works: every k is a hop cutoff, and the cliff is sharp (35.6 % → 0.0 % between k = 1.10 and 1.20 on a 6-hop grid). Rejected — this is the mistake #188's history warns against repeating.
Scalar with m ≤ 1/(h_max−1). Keeps thesis literalism at the cost of a margin so small (~0.05 on a satellite grid) it mostly measures refresh phase, and it still couples the knob to the topology's diameter. Rejected as default; the scalar form remains reachable for fidelity experiments only by reverting this ADR's change, deliberately not by a runtime switch (a knob whose wrong setting silently re-introduces a measured pathology is the #206 anti-pattern).
Gate on the unsampled-link case only (drop the margin entirely). Loses the thesis's efficiency reading — a same-link improvement above the γ-ramp band is real information worth checking too. Rejected.

Consequences¶

The gate's semantics change wherever proactiveVirtualMargin > 0 is set; nothing changes at the shipped default (margin 0 short-circuits before the comparison — byte-identical behaviour).
Core tests pin the new shape (test_proactive.cpp cases 10–11c: same-link block/pass at the margin, the unsampled-link trivial pass, and the hop-ceiling cancellation case the scalar form fails).
The #180 A/B (satellite corridor field, gate m = 0.3 vs gate off vs the old scalar) is the empirical record for any future default change.
docs/fidelity.md deviation 4b and docs/configuration.md's proactiveVirtualMargin row updated: the mechanism is thesis, the shape is a measured correction, the default remains a repo choice.
Wire format untouched (the gate is a local emission decision).

References¶

#180 (the finding, probe measurements and algebra), #188 (the scalar form's measured failure),

PR #278 (exp_uniformity_probe), #248 (regime scoping), #216 (the satellite field the gate serves), core/src/ant_router_logic.cpp (shouldSendProactive), Ducatelle 2007 §4 (lines 4084–4088).