Paper fidelity — AntHocNet v1.0

How this implementation maps to the canonical paper, what is reproduced, and where it deviates — honestly. The reference is [1] = Di Caro, Ducatelle & Gambardella, AntHocNet: an Ant-Based Hybrid Routing Algorithm for Mobile Ad Hoc Networks, PPSN VIII, LNCS 3242, 2004 (digest: publications/papers/2004-ppsn-anthocnet.md). The living compliance ledger is issue #91.

What v1.0 claims. A faithful, simulator-agnostic implementation of the AntHocNet algorithm with NS-2 and NS-3 adapters, whose mechanisms and parameters match [1] and whose delivery-ratio and overhead advantages over AODV reproduce the paper. It does not claim bit-for-bit reproduction of the paper's Qualnet numbers, nor that every headline metric is reproduced on every ns-3 channel model — see Known deviations below.

Mechanisms (feature ↔ paper § ↔ evidence)

Mechanism	[1] §	Implementation	Status
Reactive path setup (forward/backward ants)	3.1	core/ant_router_logic	✅ core-tested
Multipath setup (hops+time acceptance filter; [1] states 1.5×)	3.1	GenerationTracker, enableMultipath (#96/#97)	✅ default on — factors are the thesis's a1 = 0.9 / a2 = 2.0 since #177
MAC-queue-aware per-hop cost (Q+1)·T̂_mac (A2)	3.1	enableMacMetric (#67/#70)	✅ formula matches; default off (gated)
Pheromone deposit τ=((T̂+h·T_hop)/2)⁻¹, running avg γ	3.1	pheromone_engine	✅
Stochastic data routing, pheromone²	3.2	betaData=20 (#179)	⚠️ deliberate deviation — follows the thesis's β₃ = 20 instead of [1]'s squared rule; measured, see below
Proactive ants, unsquared pheromone, broadcast≤2	3.3	betaAnts=20, proactiveMaxBroadcasts=2 (#45/#179)	⚠️ deliberate deviation on the exponent (thesis β₁=β₂=20); broadcast≤2 ✅
Pheromone diffusion via hello ants	3.3	enableDiffusion (ADR-0007)	✅
Reactive ant floods when it has no pheromone	3.1	default behaviour; enableDirectedReactive steers instead (ADR-0016)	✅ faithful by default — the deviation is off
Hello beacons (1 s, 2 missed → remove)	3.3 fn.1	helloInterval, allowedHelloLoss	✅
Link-failure detection + notification	3.4	detectors A/D (#19/#44/#54)	✅
Bounded local repair (≤2 broadcasts, wait 5×)	3.4	repairMaxBroadcasts, repairWaitFactor	✅

Every pinned-down parameter and its match/deviation status is tabulated in the paper digest §3; the corresponding Config defaults, the provenance category of each ([1] §x / thesis / repo choice / unknown) and how to calibrate one are in configuration.md.

Headline results vs AODV (paper regime: 50 nodes, 1500×300 m, RWP 20 m/s, 20 CBR flows)

Paper claim ([1] §4.2)	Reproduced?	Evidence (this repo, ns-3, 5 seeds)
PDR ≥ AODV, gap grows with difficulty	✅	92.1 vs 81.4 (disk); 92.9 vs 83.4 (two-ray). #22 (mobility), taxonomy
Overhead (NRL) below AODV	✅	45 vs 61 (disk); 44 vs 75 (two-ray)
Delay/jitter QoS advantage (bounded tail)	🟡 partial	two-ray (paper PHY): mean-delay parity with AODV (52.8 vs 52.6 ms), jitter within 11 %, after the #21 reconv hold cap (#104). Disk model: tail narrowed (delay99 −37 %, jitter −26 % via ReconvHoldCap, measured at the then-default 1 s; 200 ms since #371) but still above AODV

Correction pending re-measurement (#169, 2026-07-25)

Every benchmark number in this document was produced with reactiveMaxBroadcasts = 2, which — because a reactive forward ant broadcasts at every node lacking pheromone for the destination — was a hop limit on route discovery: destinations more than ~5 hops away were never found at all. Both primary sources bound the reactive flood by duplicate suppression and maxPathLength, never by a broadcast count; the 2-broadcast rule belongs to proactive ants ([1] §3.3). The default is now unbounded.

This most likely explains the sparse/long-path results that were previously read as protocol character — in particular the static-network inversion (delivery 59.9 % vs AODV 98.9 % at pause=900, while AntHocNet led under constant motion) and the decay of its advantage as the field is stretched. Until the sweeps are re-run, treat every result below — including the headline PDR/NRL figures — as pending re-measurement, and do not cite the sparse-static weakness as a property of the algorithm.

Confirmed, and it exposed a second defect. The first taxonomy run after the fix bears the diagnosis out: sparse-static rose 83.8 → 93.0 %, overtaking AODV (81.9 %), and the inversion is gone. But large-scale fell 75.8 → 21.7 % and heavy-load 85.0 → 51.4 %, with NRL rising 99.9 → 3071. Cause (#173): with enableMultipath on, a reactive forward ant is admitted by the acceptance band instead of (src,seq) duplicate suppression, so removing the broadcast budget left the reactive flood unbounded in dense graphs. reactiveMaxBroadcasts became a per-(node, generation) broadcast count (default 2 at the time) rather than a per-path budget — bounding the flood without reintroducing the

169 hop limit. Numbers measured between those two fixes carry the flood and

are not representative either. (#177 later retired the count — default now -1, unbounded — because the thesis's a1 = 0.9 acceptance band bounds the flood on its own; see configuration.md §4.)

Where we ship [1]'s algorithm and the 2007 thesis superseded it

The thesis is this repo's designated primary source for parameters that [1] leaves unspecified (#58/#70). On three mechanisms the two sources disagree because the thesis's authors changed the algorithm, and in all three we ship the [1] version. That is defensible — [1] is a real published algorithm and the one this document's claims are written against — but it had not been stated anywhere, so a reader checking us against the thesis would read three deliberate choices as three defects. Recorded here (audit: #182):

1. Multipath reactive route setup. [1] §3.1 admits later same-generation ants within a 1.5× band (enableMultipath, antAcceptanceFactor). The thesis parameterises the same band differently (a1 = 0.9, plus a2 = 2 for first-hop-disjoint ants) and then records dropping the design: reactive setup is restricted to a single route, with multiple routes obtained through proactive maintenance instead. #177 measured the options and adopted the thesis's factors (a1 = 0.9, a2 = 2.0) while keeping [1]'s multipath mechanism — so this item is now a hybrid, not a pure [1] choice; #178 records the citation trap.
2. Proactive broadcast probability (proactiveBroadcastProb = 0.1). Our exact value appears in the thesis, but in §4.3.4, "Older versions of AntHocNet". In the shipped thesis algorithm proactive forward ants are "never broadcast" — on reaching a node with no routing information for the destination they are simply discarded.
3. Proactive broadcast budget (proactiveMaxBroadcasts = 2). Same section, same status: the thesis states the budget as nb = 2 for the older algorithm only. With no proactive broadcasting in the current version, there is nothing for the budget to bound there.

So the accurate claim is "faithful to [1]", not "as close to the 2007 thesis as possible". Provenance for each parameter is in configuration.md §3.1, where these rows are marked thesis §4.3.4 (superseded version) rather than a bare thesis.

The exception: the β exponents now follow the thesis (#179)

The three items above are places where the two sources disagree and we keep [1]. The routing exponents are the one place we deliberately went the other way, so the "faithful to [1]" claim needs this carve-out.

The sources disagree on a ratio, not just a value:

	data exponent	ant exponent	relationship
[1] PPSN 2004	pheromone squared (§3.2, "to be more greedy with respect to the better paths")	unsquared (§3.3, "not squared, so that they sample the paths more evenly")	asymmetric, 2 : 1 — ants explore wider than data
2007 thesis	β₃ = 20 (eq 4.6, "we normally keep β₃ on 20")	β₁ = 20 (eq 4.1), β₂ = 20 (eq 4.5)	symmetric, 1 : 1

[1] gives no absolute values, only the squared/unsquared form; the thesis gives numbers. That is the same situation as hopTimeSec (#88), where the thesis's 3 ms replaced a value [1] defines but never states — and the same resolution.

What we give up by adopting 20/20 is [1]'s reason for the asymmetry: ants no longer sample more evenly than data, because they use the identical exponent. The thesis does not lose exploration, it relocates it — proactive ants there route on max(regular, virtual) pheromone (eq 4.5) and carry a per-node broadcast probability, so exploration comes from the virtual pheromone rather than from a flatter distribution.

Unlike the three items above, this one is measured, not just sourced. A 20-seed × 900 s paired A/B across all six discrete scenarios (#179) moved five of six to PAIRED-IMPROVED with none regressing: PDR +0.50 to +3.23 pp and routing load −9.7 % to −16.7 % where significant, with the baselines byte-identical across arms. It also narrowed the #21 delay-tail deficit against AODV on five of six scenarios (heavy-load +125.4 % → +82.4 %) without closing it. The one adverse signal is dense-small's tail at +9.55 ms, where the bootstrap CI and the Wilcoxon test disagree.

betaAnts conflates the thesis's β₁ and β₂. They coincide at 20, so nothing is lost today — but the thesis sweeps β₂ separately in chapter 5, so the field must be split before anyone deviates per ant type.

Known deviations (honest list)

1. Delay tail on the ns-3 disk model (#21). On the contention-dominated range/disk propagation model AntHocNet's delay99/jitter run above AODV even after the multipath (#96), backward-ant flush (#101) and reconv hold-cap (#104) work. On two-ray — the paper's actual PHY — the gap closes to parity on mean delay. The residual is a channel-model artefact (CONTEXT.md §8), not an algorithmic gap. Mitigated by ReconvHoldCap (measured at the then-default 1.0 s; the shipped default is 200 ms since the #371 flip for the v1.5.0 re-baseline). Note (2026-07-25): the T_hop co-lever named in item 2 has now been corrected to the thesis value; the benchmark impact on this tail is pending a re-run (#88).
2. T_hop — resolved from the primary source (#88, 2026-07-25). [1] defines the constant but states no number. The 2007 Ducatelle thesis does: "we kept thop on 0.003 sec". The repo's provisional 50 ms was 16.7× too large and is now corrected to 3 ms. Because T_hop weights hop count against measured delay in every pheromone deposit, this changes all routing goodness values — every benchmark number in this document predates the fix and must be re-measured before being cited as current. A prior sweep suggested a few-ms value cuts delay/jitter ~12 %, so the #21 delay tail may improve.
3. Evaporation (enableEvaporation, default on; ADR-0012) — a time-decay safety net not present in [1] (whose reinforcement is a pure running average). Config-gated so the paper-faithful ablation is available.
4. Proactive ant clocking — [1] emits one proactive ant per n data packets; the repo clocks it on a timer (proactiveInterval). #26 item 04. Note the mechanism is not a deviation against the thesis, which uses a plain periodic timer as we do; what differs there is only the interval. The #248 sweep measured it (7 points, paper field, 900 s, paired seeds): the thesis's §5.3.3 "2 s almost always best" inverts on our field — the curve is strictly monotone toward less proactive, with EnableProactive=false the measured MANET optimum (97.2 PDR / 31.73 NRL vs 94.9/40.51 at the 10 s default; 2 s costs −15.8 pp PDR and ×5.3 NRL). The 10 s default is retained cross-regime — the satellite track (#192) is the configuration that needs proactive sampling; MANET benchmark guidance is proactive off. 4b. The thesis's proactive emission gate is implemented per-link and shipped OFF (proactiveVirtualMargin, default 0; the thesis states 0.10 best-vs-best). Two deviations by measurement, not by omission: the value and the shape. At the thesis value the literal scalar gate suppressed essentially all proactive maintenance (pass rate 0–3.8 %; PR #188 measured the damage as NRL +36-68 %, PDR −5 to −6.4 pp) — originally because two table-hygiene defects depressed virtual pheromone (#262 aging clock, #279 advert magnitude, both since fixed), but the post-fix uniformity probe (exp_uniformity_probe, PR #278) showed the scalar form is structurally broken regardless: best-vs-best compares estimators of different paths, and its ratio has a hard ceiling τ(h−1)/τ(h) ≈ h/(h−1) (measured byte-exact), so any fixed margin silently disables the gate beyond 1 + 1/m hops. ADR-0018 re-derives the comparison per-link — v vs r on the same neighbour, an unsampled-link hint passing trivially — which cancels the systematic and makes the margin hop-independent (probe: per-link ratio centred 0.97–1.06). The mechanism is the thesis's; the shape is a measured correction; the default stays 0 pending the satellite-regime A/B on #180. core/tests/exp_gate_cascade.cpp reproduces the historical failure.
5. Multipath link-failure suppression — with multipath on, a link break that leaves a usable alternate next-hop is absorbed rather than always notified (paper §3.4 always notifies). Benchmark-justified (#96): notification floods cost PDR on the ns-3 channel.
6. Directed reactive discovery (enableDirectedReactive, default off; ADR-0016) — a mechanism in neither source: a reactive forward ant may be unicast along the diffused virtual gradient instead of broadcast. Listed here for completeness rather than as an active deviation — with the default off the shipped protocol matches [1] §3.1 exactly, and the gate exists to make the deviation runnable (the inverse of item 3, where the gate makes the fidelity runnable). #244 measures it; #245 tracks the open hazard that a stale gradient can strand the single steered ant where a flood would not have been.
7. Cross-simulator parity is not guaranteed — NS-2 and NS-3 have different MAC/PHY; the NS-2 adapter also has no pending-queue hold cap yet. Treat cross-sim comparison as behaviour re-validation, not a bit-for-bit port.

Verification status

• Algorithm mechanisms: ✅ covered by core/tests (unit + randomized property/invariant sweeps), NS-2/NS-3 e2e delivery smokes in CI.
• Parameters vs [1] (2004 paper): ✅ verified — see the digest table.
• Parameters vs 2007 thesis: 🟡 partial — thesis obtained 2026-07-25. T_hop ✅ adopted (3 ms). Jitter estimator ✅ defined in the thesis (eq. 5.1, Σ|(tᵢ−tᵢ₋₁)−(tᵢ₋₁−tᵢ₋₂)|) but not yet reconciled with our FlowMonitor jitterSum metric (#89). Full field table ✅ read from §5.1.3 and adopted into --scenario=thesis (#58) — five constants corrected (2400×800 m, 10 m/s, 2048 bps, 250 m range, 20 repetitions); the thesis's two-ray propagation still has to be asked for with --propagation=tworay.
• Numbers on the paper's own 900 s / large-scale field: ⏳ --scenario=thesis preset exists; the multi-hour run is future work.

v1.0 is the milestone where the mechanisms and 2004-paper parameters are faithful and the delivery/overhead claims reproduce; the thesis parameter cross-check and the full disk-model delay-tail closure are the roadmap beyond it (#91).