docs/VALIDITY.md

Validity layer

Generated by python -m eval.report, except the two sections copied through verbatim: Metric revision (pre-registered before the rerun) and P6 results (frozen). Do not edit by hand. The results are also section 10 of eval/results.md, from the same run.

Origin attribution names the peer that announced a transaction first, or most centrally, to an observer. Calibration says how often such a name is right. The validity layer (analysis/validity.py) asks two prior questions: can an attribution be about this transaction at all, and if so, what may it claim? Every detector runs; every reason that fires is reported; the verdict's tier is the most severe among them.

tier	effect
ABSTAIN	the answer is withheld, through the existing rule — low_confidence_origin in engines/propagation, eval.origin.flagged_at everywhere else — not a second path
QUALIFIED	the answer is given with its meaning restricted, and never shaped like an IP attribution
ANNOTATE	the answer is given unchanged, with the flag and its evidence
PASS	nothing fired

validity.enforce: false switches the layer off in that one place; validity.mode: binary restores P6's gate, where every non-PASS verdict abstained.

Every origin the system emits — engines/propagation's estimates, the origination model's output, the API's /transactions/{txid}/propagation, the red-team origin table, each attribution lead, and the console views of them — carries a probability, its calibration basis, and a verdict {"tier", "reason", "reasons", "confidence", "evidence"}; where an answer is given, answer says what it may claim: ip_attribution, onion_identity (no ip field) or broadcasting_peer (with input_ownership: not attributable). tests/test_api.py asserts the verdict at the API boundary; tests/test_validity.py asserts that no TOR_ONION answer carries an IP and no COINJOIN answer attributes input ownership.

Reason codes

confidence is confidence that the condition is present, not that the attribution is wrong.

code	tier	fires when	reads	catches	does not catch
DEGENERATE	ABSTAIN	0 or 1 peers announced the transaction	candidate_count (the matrix's degenerate; a single-observation tree)	a capture with nothing to rank	— it is definitional
NOT_REACHABLE	ABSTAIN	every candidate is a known public relay	the matrix's scope_out; the same rule on a tree	the zero-ceiling case: the sender is not among what this observer can see	a sender absent from the candidates while an ordinary node happens to be among them — that is the unobserved-origin case calibration prices
COINJOIN	QUALIFIED	graph.clustering.is_coinjoin: ≥3 inputs, ≥3 outputs, ≥3 equal-value outputs above dust, no more equal outputs than inputs	the transaction's structure	a mix; the answer becomes "broadcasting peer only", input ownership non-attributable	an equal-value batch payout from one owner paying no more equal amounts than it spends inputs (a false positive: structure cannot see input ownership); a CoinJoin with unequal denominations or fewer than three equal outputs (PayJoin, some WabiSabi rounds)
TOR_ONION	QUALIFIED	the named peer is a .onion address	the peer address (onion parsing in p2p/capture_reader.py)	a hidden-service peer named as origin; the answer becomes an onion identity, never an IP	a Tor sender whose transaction entered through a mixed-transport node: the observer sees that node's clearnet address, and nothing in one capture marks it
DANDELION_STEM	ANNOTATE	the gap after the first announcement is ≥ isolation_ratio × the median gap among the rest, with ≥ min_candidates candidates	announcement times	an observer on or next to a stem, which hears one peer, then silence, then the broadcast	a stem that ends away from every observer (the common case); and it false-alarms on sparse captures. A flag only: BIP-156 was never merged into Bitcoin Core, which has no stem phase
V2_PASSIVE_TAP	ANNOTATE	a pcap capture held port-8333 flows that never decoded	capture_source, unreadable_flows	a packet capture blind to its BIP-324 v2 peers, whose announcements are then missing from the candidates	which peers those were. It never fires on debug.log or .btcap: the node writes those, and as a session endpoint it decrypts. A pcap on the node's own host is as blind as a span port — tcpdump holds no session keys
NOT_ASSESSED	—	not a detector: a stored output from before this layer	—	keeps an old lead from being served as PASS	—

Taint and leads. A CoinJoin terminates taint: graph.entity_graph counts the CoinJoin transactions on each entity edge (mix_count, from the same is_coinjoin), and fusion.taint.propagate does not cross an edge made of nothing else. A correlation lead says "this IP broadcast for this cluster's inputs", so none is built from a COINJOIN or TOR_ONION answer or a NOT_REACHABLE one. DEGENERATE observations are kept — aggregating many single-candidate observations, each weighted by its low confidence, is what correlation is for — and ANNOTATE flags ride along; a lead's own verdict is PASS or ANNOTATE, naming how many of its observations carried each flag.

The DANDELION_STEM thresholds were fixed in config.yaml before the first P6 result and have not been tuned since.

Metric revision

Pre-registered. Written and committed before the evaluation was re-run; nothing in this section was changed after the first revised result existed.

What P6 showed

Enforced as a binary gate, the validity layer lowered the cross-topology cost score from 0.057 to 0.034. The per-reason table (P6 results, below) shows why, and each reason failed differently:

• DANDELION_STEM withheld 230 answers that were correct_actionable against 23 wrong. The detector cannot tell an observer that is the sender's first stem hop — where the lone first announcer really is the sender — from one further down the stem. It also fired on 4.2% of the base corpus, which has no stems. And the condition it guards against does not occur on Bitcoin's network: BIP-156 (Dandelion) was never merged into Bitcoin Core, and no Core release has a stem phase. A gate for a condition the deployed network does not produce, with 0.43 precision, costs answers and protects nothing.
• TOR_OR_V2 fired mostly on v2. BIP-324 encrypts a link; it hides neither the peer's address nor its timing from the node at the other end, and the simulator models no v2 timing effect at all, so every v2 abstention withheld an answer the evidence did not question. What v2 does remove is a passive observer's view: a packet capture cannot read a v2 link, so its candidate set can be missing the v2 peers. That is a statement about the capture, not about the answer.
• COINJOIN withheld 167 right answers against 12 wrong — but the metric scored "naming the broadcaster of a mix" exactly like naming the sender of a payment. An unqualified answer about a CoinJoin claims more than the evidence supports: it reads as "this address sent these coins", and the inputs belong to 5–12 unrelated participants. P6's cost function had no outcome for that claim, so it could not price what abstaining avoided.
• TOR (onion) answers were already scored correct_infrastructure (+0.3) after the onion classification fix, not correct_actionable. The onion clause had precision 1.0; the question is what the answer may say, not whether to give one.

Verdict tiers

Every detector still runs, and every reason that fires is reported. The verdict's tier is the most severe tier among them:

tier	reasons	what happens to the answer
ABSTAIN	DEGENERATE, NOT_REACHABLE	withheld through low_confidence_origin / eval.origin.flagged_at, as before. Already free: both were abstentions by construction in P6.
QUALIFIED	TOR_ONION, COINJOIN	given, with its meaning restricted. TOR_ONION: the answer is an onion identity and is never presented as an IP. COINJOIN: the answer is the broadcasting peer only; input ownership is explicitly non-attributable; taint still terminates at the mix. Neither becomes a correlation lead.
ANNOTATE	DANDELION_STEM, V2_PASSIVE_TAP	given unchanged, with the flag and its evidence attached. DANDELION_STEM leaves the abstention path because Dandelion is not deployed in Bitcoin Core. V2_PASSIVE_TAP replaces the v2 clause and fires only on a pcap capture in which port-8333 flows could not be decoded — the evidence that v2 peers are missing from the candidate set. A debug.log or .btcap capture is written by the node, a session endpoint that decrypts, so v2 never fires there. (A pcap taken on the node's own host is no better than a span port: tcpdump does not hold the session keys. "Endpoint" here means the node process, not the host.)
PASS	none	unchanged.

validity.enforce: false stays the off switch. validity.mode: binary reproduces P6's gate (every non-PASS verdict abstains) so P6's policy can be scored under the new metric.

Cost-function outcomes

eval.origin's outcome vocabulary is extended, not forked. Each estimate gets exactly one outcome, decided in this order; the weights go in engines.propagation.origin_filter.cost_weights.

order	outcome	when	weight	rationale
1	abstained	the policy withholds it (ABSTAIN tier; any non-PASS under binary; the cutoff; a relay at the top)	0.0	unchanged
2	coinjoin_input_misattribution	the transaction is a CoinJoin by ground truth, it is answered, and the answer is not qualified COINJOIN (layer off, binary, or the detector missed it)	−3.0	An unqualified answer about a mix asserts ownership of other participants' inputs: a false claim that points follow-up at people it should not. That is the harm wrong_uninvolved_third_party prices, so it carries the same weight. It is not scaled by the number of participants: the harm is the claim, and scaling would let the generator's participant-count distribution set the number. It applies whether or not the named broadcaster is right, because the ownership claim is wrong either way.
3	qualified_correct	QUALIFIED tier, answered, the named peer is the true origin	+0.3	Equal to correct_infrastructure: a right answer that cannot support IP-level follow-up (an onion identity) or supports it only for the broadcaster and never for input ownership (a CoinJoin).
4	qualified_wrong	QUALIFIED tier, answered, the named peer is not the true origin	−1.0	Qualification scales the cost of a wrong answer by the same factor it scales the value of a right one: 0.3 / 1.0 for right answers, −1.0 / −3.0 for wrong ones. Not −3: the qualification removes the claim that makes a wrong answer expensive (a wrong onion identity names no IP; a wrong CoinJoin broadcaster is presented without input ownership). Not 0: it still sends an investigator to the wrong node.
5	correct_actionable, correct_infrastructure, wrong_uninvolved_third_party	everything else, ANNOTATE tier included	+1.0, +0.3, −3.0	unchanged

The P6 metric is outcomes 1 and 5 only, with P6's weights. Every row of the revised report shows the P6 metric and the revised metric side by side.

Protocol

• Three policies are scored on the same test transactions: off (enforce: false), binary (P6's gate, current detectors), tiered (the revision). Each is scored under both metrics.
• Each (policy, metric) row's abstention cutoff is chosen on the calibration captures under that policy and that metric, by eval.origin.choose_cutoff_for.
• Base and variant corpora, within- and cross-topology.
• Decision rule: the revision beats no-validity if and only if the tiered policy's cross-topology cost score on the revised metric exceeds the off policy's on the same metric. Reported plainly either way. The weights above are not changed after the result.
• Known bias, stated in advance. Outcome 2 can only lower the off policy's score: it prices a claim the off policy makes and the tiered policy does not. It is justified on its own terms above, not by its effect; the P6 metric is reported beside it on every row so the share of any difference that comes from the metric rather than from the policy is visible. The tiered policy must also beat off on the P6 metric for the revision to be called an improvement without qualification; beating it only on the revised metric will be reported as exactly that.
• Red-team time-to-detect is re-run on an otherwise idle machine to settle whether P6's 2.22 s → 4.44 s was load.

P6 results (frozen)

The binary layer as committed in 3a1752a: every non-PASS verdict abstained, v2 fired on every v2 link. Kept verbatim; the numbers below are the ones the metric revision responds to, and they are never regenerated.

condition="simulated". What the simulator omits: The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.

Corpus variant efc5a443edf54d56, deterministic from origination/manifest_validity.json: the base manifest's captures, each drawing four independent toggles. The model is fitted, calibrated and scored on it exactly as in section 9. Captures with each toggle on: dandelion 1029, onion 1037, v2 1101, coinjoin 1071, of 2160.

Ground-truth prevalence

condition	measure	transactions	share of broadcast
simulated	dandelion stem	21599	0.248
simulated	dandelion stem through an observer	2469	0.028
simulated	onion origin	14650	0.168
simulated	coinjoin	6562	0.075
simulated	equal value batch payout	4093	0.047
simulated	matrix rows over v2	112673	n/a
simulated	matrix rows from onion peers	1617	n/a

condition="simulated". The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.

Detector precision and recall against the generator's labels

Each detector run on its own over every observed test transaction. TOR_OR_V2 depends on which peer is named; the model's choice is used.

condition	test set	detector	ground truth	transactions	truly present	fired	true positives	precision	precision 95% CI	recall	recall 95% CI
simulated	within-topology	DANDELION_STEM	a stem was used	12223	3078	737	313	0.425	0.390–0.461	0.102	0.091–0.113
simulated	within-topology	DANDELION_STEM	the stem passed through an observer	12223	375	737	219	0.297	0.265–0.331	0.584	0.533–0.633
simulated	within-topology	COINJOIN	CoinJoin shape	12223	923	1054	923	0.876	0.854–0.894	1.000	0.996–1.000
simulated	within-topology	TOR_OR_V2 (onion clause)	sender reachable only over Tor	12223	1905	249	249	1.000	0.985–1.000	0.131	0.116–0.147
simulated	within-topology	TOR_OR_V2 (both clauses)	sender reachable only over Tor	12223	1905	1370	372	0.272	0.249–0.296	0.195	0.178–0.214
simulated	cross-topology	DANDELION_STEM	a stem was used	19250	4678	1143	489	0.428	0.399–0.457	0.105	0.096–0.114
simulated	cross-topology	DANDELION_STEM	the stem passed through an observer	19250	635	1143	335	0.293	0.267–0.320	0.528	0.489–0.566
simulated	cross-topology	COINJOIN	CoinJoin shape	19250	1383	1597	1383	0.866	0.848–0.882	1.000	0.997–1.000
simulated	cross-topology	TOR_OR_V2 (onion clause)	sender reachable only over Tor	19250	3471	443	443	1.000	0.991–1.000	0.128	0.117–0.139
simulated	cross-topology	TOR_OR_V2 (both clauses)	sender reachable only over Tor	19250	3471	2411	720	0.299	0.281–0.317	0.207	0.194–0.221

condition="simulated". The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.

On the base corpus 945261fe69e264d6 (cross-topology test set), where no transaction has a stem, DANDELION_STEM fires on 815 of 19414 transactions (0.042): every one a false alarm.

The system with and without the validity layer

The supervised model on the same test transactions; each row's cutoff chosen on the calibration captures by eval.origin.choose_cutoff_for, with the layer on and off respectively. The headline is cost_weighted_score and acc if answered.

condition	test set	estimator	n	top1	top1 95% CI	top3	top3 95% CI	abstention rate	acc if answered	acc if answered 95% CI	ceiling (origin observed)	cost_weighted_score	wrong_uninvolved_third_party	cutoff (chosen on calibration)
simulated	within-topology	supervised_origination (P5), validity off	12223	0.192	0.185–0.199	0.195	0.188–0.202	0.887	0.919	0.903–0.932	0.195	0.067	112	0.700
simulated	within-topology	supervised_origination (P5) + validity	12223	0.192	0.185–0.199	0.195	0.188–0.202	0.933	0.906	0.884–0.924	0.195	0.042	77	0.700
simulated	cross-topology	supervised_origination (P5), validity off	19250	0.190	0.185–0.196	0.193	0.188–0.199	0.885	0.900	0.886–0.911	0.194	0.057	223	0.700
simulated	cross-topology	supervised_origination (P5) + validity	19250	0.190	0.185–0.196	0.193	0.188–0.199	0.932	0.874	0.855–0.891	0.194	0.034	165	0.700

condition="simulated". The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.

Verdict (cross-topology): the validity layer does not improve the cost score — 0.034 with it against 0.057 without. That means its detectors abstain on transactions the model was getting right more often than on ones it was getting wrong; the ablation shows which. Accuracy-given-answered is 0.874 with the layer and 0.9 without; abstention 0.932 against 0.885.

What each reason withholds

At the layer-off cutoff: how many transactions each reason withholds that the layer-off system would have answered, what those answers would have been, and the cost score with only that reason enforced. DEGENERATE and NOT_REACHABLE were already abstentions by construction and cost nothing new.

condition	test set	reason	withheld	would have been answered	…correct_actionable	…correct_infrastructure	…wrong_uninvolved_third_party	cost score, this reason only	change vs layer off
simulated	within-topology	DEGENERATE	1107	0	0	0	0	0.067	0.000
simulated	within-topology	NOT_REACHABLE	914	0	0	0	0	0.067	0.000
simulated	within-topology	COINJOIN	878	127	95	22	10	0.061	-0.006
simulated	within-topology	TOR_OR_V2	1030	229	72	148	9	0.059	-0.008
simulated	within-topology	DANDELION_STEM	577	209	193	0	16	0.055	-0.012
simulated	cross-topology	DEGENERATE	1482	0	0	0	0	0.057	0.000
simulated	cross-topology	NOT_REACHABLE	1446	0	0	0	0	0.057	0.000
simulated	cross-topology	COINJOIN	1340	179	134	33	12	0.051	-0.006
simulated	cross-topology	TOR_OR_V2	1838	482	156	303	23	0.048	-0.009
simulated	cross-topology	DANDELION_STEM	854	254	230	1	23	0.048	-0.009

condition="simulated". The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.

Verdicts issued on the test sets

verdict	within-topology	cross-topology
PASS	7717	12290
DEGENERATE	1107	1482
TOR_OR_V2	1030	1838
NOT_REACHABLE	914	1446
COINJOIN	878	1340
DANDELION_STEM	577	854

condition="simulated". The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.

The variant's section-9 table (cross-topology), validity enforced

condition	test set	estimator	n	top1	top1 95% CI	top3	top3 95% CI	abstention rate	acc if answered	acc if answered 95% CI	ceiling (origin observed)	cost_weighted_score	wrong_uninvolved_third_party	cutoff (chosen on calibration)
simulated	cross-topology	first_spy_baseline	19250	0.173	0.168–0.178	0.191	0.185–0.197	n/a	0.173	n/a — the floor never abstains	0.194	-2.327	15922	n/a
simulated	cross-topology	first_timestamp	19250	0.176	0.171–0.182	0.191	0.186–0.197	0.991	0.874	0.817–0.915	0.194	0.004	22	0.500
simulated	cross-topology	rumor_centrality	19250	0.086	0.082–0.090	0.122	0.118–0.127	0.991	0.899	0.844–0.936	0.194	0.005	17	0.500
simulated	cross-topology	timestamp_weighted_centrality	19250	0.176	0.171–0.182	0.191	0.186–0.197	0.991	0.894	0.839–0.932	0.194	0.005	18	0.500
simulated	cross-topology	supervised_origination (P5)	19250	0.190	0.185–0.196	0.193	0.188–0.199	0.932	0.874	0.855–0.891	0.194	0.034	165	0.700

condition="simulated". The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.

Red-team timing (measured)

Hand-written from a dedicated measurement; copied through, not regenerated. P6's report moved red-team median time-to-detect from 2.22 s to 4.44 s. The batch (eval.redteam_batch.run_batch, shifted dataset at the default rate, 30 injections) was re-run alone, on an otherwise idle machine (load average 1.1–1.5), alternating the current code with a6a0c7c — the commit before the validity layer — in a separate worktree:

code	run 1	run 2	batch wall time
current (tiered validity)	4.13 s	4.06 s	221 s, 220 s
a6a0c7c, before validity	3.96 s	3.90 s	210 s, 210 s

It was not load, and it was not the validity layer. The layer costs about 4% (≈0.15 s per detection). The pre-validity code itself no longer reproduces 2.22 s — the figure its own eval/results.md recorded at 11:23 the same morning. The machine is a laptop: the kernel logged the AC adapter on-line at boot (11:11), and during these runs it was on battery, discharging, with the powersave CPU governor. Power state is the likely cause; that is an inference, not a measurement — re-running both on AC would settle it. Until then the timing figures in eval/results.md are machine-state dependent and should not be compared across runs.

Results — metric revision

condition="simulated". What the simulator omits: The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.

Variant corpus 973cc484afc7e7f5 (origination/manifest_validity.json) and base corpus 3f2d3bbe9a1f7088 (origination/manifest.json). The model is fitted, calibrated and scored on each exactly as in section 9. Variant captures with each toggle on: dandelion 1029, onion 1037, v2 1101, coinjoin 1071, of 2160.

The policies under both metrics

Same test transactions for every row. off: no validity layer. binary: P6's gate (every non-PASS verdict abstains), current detectors. tiered: the revision. Each metric's cost is at that (policy, metric)'s own cutoff, chosen on the calibration captures; the other columns are at the revised-metric cutoff. The headline is the cost score and acc if answered.

Variant corpus

condition	corpus	test set	policy	n	cost, revised metric	cost, P6 metric	abstention rate	acc if answered	acc if answered 95% CI	correct_actionable	correct_infrastructure	wrong_uninvolved_third_party	qualified_correct	qualified_wrong	coinjoin_input_misattribution	cutoff, revised	cutoff, P6
simulated	variant	within-topology	off	12223	0.034	0.067	0.890	0.924	0.909–0.937	991	149	93	0	0	112	0.800	0.700
simulated	variant	within-topology	binary	12223	0.044	0.044	0.928	0.903	0.882–0.921	800	0	86	0	0	0	0.700	0.700
simulated	variant	within-topology	tiered	12223	0.063	0.067	0.887	0.919	0.903–0.932	1004	0	102	265	10	0	0.700	0.700
simulated	variant	cross-topology	off	19250	0.028	0.057	0.887	0.903	0.890–0.915	1518	305	200	0	0	154	0.800	0.700
simulated	variant	cross-topology	binary	19250	0.037	0.037	0.925	0.873	0.855–0.889	1268	0	184	0	0	0	0.700	0.700
simulated	variant	cross-topology	tiered	19250	0.053	0.057	0.885	0.900	0.886–0.911	1527	1	211	470	12	0	0.700	0.700

condition="simulated". The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.

Base corpus (no invalidating condition simulated)

condition	corpus	test set	policy	n	cost, revised metric	cost, P6 metric	abstention rate	acc if answered	acc if answered 95% CI	correct_actionable	correct_infrastructure	wrong_uninvolved_third_party	qualified_correct	qualified_wrong	coinjoin_input_misattribution	cutoff, revised	cutoff, P6
simulated	base	within-topology	off	12279	0.122	0.122	0.815	0.915	0.902–0.925	2078	0	194	0	0	0	0.750	0.750
simulated	base	within-topology	binary	12279	0.110	0.110	0.825	0.907	0.894–0.919	1944	0	199	0	0	0	0.650	0.650
simulated	base	within-topology	tiered	12279	0.122	0.122	0.815	0.915	0.902–0.925	2078	0	194	0	0	0	0.750	0.750
simulated	base	cross-topology	off	19414	0.120	0.120	0.808	0.906	0.896–0.915	3382	1	350	0	0	0	0.750	0.750
simulated	base	cross-topology	binary	19414	0.107	0.107	0.817	0.896	0.886–0.906	3189	1	369	0	0	0	0.650	0.650
simulated	base	cross-topology	tiered	19414	0.120	0.120	0.808	0.906	0.896–0.915	3382	1	350	0	0	0	0.750	0.750

condition="simulated". The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.

Verdict (variant, cross-topology): the revised layer beats no-validity on the revised metric only — 0.053 against 0.028; on the P6 metric it scores 0.057 against 0.057. As pre-registered, that is reported as exactly this and not as an unqualified improvement: the difference comes from pricing the CoinJoin ownership claim, not from the policy answering better by P6's measure. P6's binary gate, rescored on the same transactions: revised 0.037, P6 metric 0.037.

Why each answer was withheld

Every abstention, by the one reason code eval.origin.abstention_reasons gives it — the validity verdict's leading reason when its tier is withheld, then RELAY_AT_TOP (a listed public relay ranked first), then BELOW_CUTOFF (calibrated confidence under the policy's cutoff). Same frames and revised-metric cutoffs as the table above; the codes sum to abstained, so coverage (1 − abstention rate) is explained next to acc if answered.

condition	corpus	test set	policy	n	abstained	abstention rate	acc if answered	DEGENERATE	NOT_REACHABLE	COINJOIN	TOR_ONION	DANDELION_STEM	V2_PASSIVE_TAP	RELAY_AT_TOP	BELOW_CUTOFF
simulated	variant	within-topology	off	12223	10878	0.890	0.924	0	0	0	0	0	0	1125	9753
simulated	variant	within-topology	binary	12223	11337	0.928	0.903	1107	914	878	153	625	0	93	7567
simulated	variant	within-topology	tiered	12223	10842	0.887	0.919	1107	914	0	0	0	0	118	8703
simulated	variant	cross-topology	off	19250	17073	0.887	0.903	0	0	0	0	0	0	1768	15305
simulated	variant	cross-topology	binary	19250	17798	0.925	0.873	1482	1446	1340	308	968	0	167	12087
simulated	variant	cross-topology	tiered	19250	17029	0.885	0.900	1482	1446	0	0	0	0	221	13880
simulated	base	within-topology	off	12279	10007	0.815	0.915	0	0	0	0	0	0	987	9020
simulated	base	within-topology	binary	12279	10136	0.825	0.907	1305	808	0	0	521	0	86	7416
simulated	base	within-topology	tiered	12279	10007	0.815	0.915	1305	808	0	0	0	0	91	7803
simulated	base	cross-topology	off	19414	15681	0.808	0.906	0	0	0	0	0	0	1599	14082
simulated	base	cross-topology	binary	19414	15855	0.817	0.896	1802	1380	0	0	815	0	120	11738
simulated	base	cross-topology	tiered	19414	15681	0.808	0.906	1802	1380	0	0	0	0	129	12370

condition="simulated". The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.

Verdicts by tier

condition	corpus	test set	tier	transactions	reasons fired
simulated	variant	within-topology	PASS	8546	—
simulated	variant	within-topology	ABSTAIN	2021	DEGENERATE 1107, NOT_REACHABLE 1007
simulated	variant	within-topology	QUALIFIED	1031	COINJOIN 1054, TOR_ONION 249
simulated	variant	within-topology	ANNOTATE	625	DANDELION_STEM 737, V2_PASSIVE_TAP 0
simulated	variant	cross-topology	PASS	13706	—
simulated	variant	cross-topology	ABSTAIN	2928	DEGENERATE 1482, NOT_REACHABLE 1547
simulated	variant	cross-topology	QUALIFIED	1648	COINJOIN 1597, TOR_ONION 443
simulated	variant	cross-topology	ANNOTATE	968	DANDELION_STEM 1143, V2_PASSIVE_TAP 0
simulated	base	within-topology	PASS	9645	—
simulated	base	within-topology	ABSTAIN	2113	DEGENERATE 1305, NOT_REACHABLE 896
simulated	base	within-topology	QUALIFIED	0	COINJOIN 0, TOR_ONION 0
simulated	base	within-topology	ANNOTATE	521	DANDELION_STEM 543, V2_PASSIVE_TAP 0
simulated	base	cross-topology	PASS	15417	—
simulated	base	cross-topology	ABSTAIN	3182	DEGENERATE 1802, NOT_REACHABLE 1470
simulated	base	cross-topology	QUALIFIED	0	COINJOIN 0, TOR_ONION 0
simulated	base	cross-topology	ANNOTATE	815	DANDELION_STEM 857, V2_PASSIVE_TAP 0

condition="simulated". The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.

What each reason fires on, and what the answers were (P6's table, again)

At the layer-off policy's P6-metric cutoff, P6's method exactly: the transactions each reason fires on that the layer-off system would have answered, and whether those answers were right. The tiered policy column is what the revision now does with them.

condition	corpus	test set	reason	tier now	tiered policy	fired	would have been answered	…correct_actionable	…correct_infrastructure	…wrong_uninvolved_third_party
simulated	variant	within-topology	DEGENERATE	ABSTAIN	withholds	1107	0	0	0	0
simulated	variant	within-topology	NOT_REACHABLE	ABSTAIN	withholds	1007	0	0	0	0
simulated	variant	within-topology	COINJOIN	QUALIFIED	answers, qualified	1054	127	95	22	10
simulated	variant	within-topology	TOR_ONION	QUALIFIED	answers, qualified	249	170	0	170	0
simulated	variant	within-topology	DANDELION_STEM	ANNOTATE	answers, flagged	737	255	220	19	16
simulated	variant	within-topology	V2_PASSIVE_TAP	ANNOTATE	answers, flagged	0	0	0	0	0
simulated	variant	cross-topology	DEGENERATE	ABSTAIN	withholds	1482	0	0	0	0
simulated	variant	cross-topology	NOT_REACHABLE	ABSTAIN	withholds	1547	0	0	0	0
simulated	variant	cross-topology	COINJOIN	QUALIFIED	answers, qualified	1597	179	134	33	12
simulated	variant	cross-topology	TOR_ONION	QUALIFIED	answers, qualified	443	336	0	336	0
simulated	variant	cross-topology	DANDELION_STEM	ANNOTATE	answers, flagged	1143	347	281	38	28
simulated	variant	cross-topology	V2_PASSIVE_TAP	ANNOTATE	answers, flagged	0	0	0	0	0
simulated	base	within-topology	DEGENERATE	ABSTAIN	withholds	1305	0	0	0	0
simulated	base	within-topology	NOT_REACHABLE	ABSTAIN	withholds	896	0	0	0	0
simulated	base	within-topology	COINJOIN	QUALIFIED	answers, qualified	0	0	0	0	0
simulated	base	within-topology	TOR_ONION	QUALIFIED	answers, qualified	0	0	0	0	0
simulated	base	within-topology	DANDELION_STEM	ANNOTATE	answers, flagged	543	201	184	0	17
simulated	base	within-topology	V2_PASSIVE_TAP	ANNOTATE	answers, flagged	0	0	0	0	0
simulated	base	cross-topology	DEGENERATE	ABSTAIN	withholds	1802	0	0	0	0
simulated	base	cross-topology	NOT_REACHABLE	ABSTAIN	withholds	1470	0	0	0	0
simulated	base	cross-topology	COINJOIN	QUALIFIED	answers, qualified	0	0	0	0	0
simulated	base	cross-topology	TOR_ONION	QUALIFIED	answers, qualified	0	0	0	0	0
simulated	base	cross-topology	DANDELION_STEM	ANNOTATE	answers, flagged	857	302	282	0	20
simulated	base	cross-topology	V2_PASSIVE_TAP	ANNOTATE	answers, flagged	0	0	0	0	0

condition="simulated". The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.

Detector precision and recall against the generator's labels

Each detector on its own over every observed test transaction of the variant. TOR_ONION depends on which peer is named; the model's choice is used. V2_PASSIVE_TAP cannot fire here: every simulated observer is a node, which writes its own log, so no simulated capture is a pcap.

condition	test set	detector	tier	ground truth	transactions	truly present	fired	true positives	precision	precision 95% CI	recall	recall 95% CI
simulated	within-topology	DANDELION_STEM	ANNOTATE	a stem was used	12223	3078	737	313	0.425	0.390–0.461	0.102	0.091–0.113
simulated	within-topology	DANDELION_STEM	ANNOTATE	the stem passed through an observer	12223	375	737	219	0.297	0.265–0.331	0.584	0.533–0.633
simulated	within-topology	COINJOIN	QUALIFIED	CoinJoin shape	12223	923	1054	923	0.876	0.854–0.894	1.000	0.996–1.000
simulated	within-topology	TOR_ONION	QUALIFIED	sender reachable only over Tor	12223	1905	249	249	1.000	0.985–1.000	0.131	0.116–0.147
simulated	within-topology	V2_PASSIVE_TAP	ANNOTATE	a v2 peer is missing from a pcap capture	12223	0	0	0	n/a	n/a	n/a	n/a
simulated	cross-topology	DANDELION_STEM	ANNOTATE	a stem was used	19250	4678	1143	489	0.428	0.399–0.457	0.105	0.096–0.114
simulated	cross-topology	DANDELION_STEM	ANNOTATE	the stem passed through an observer	19250	635	1143	335	0.293	0.267–0.320	0.528	0.489–0.566
simulated	cross-topology	COINJOIN	QUALIFIED	CoinJoin shape	19250	1383	1597	1383	0.866	0.848–0.882	1.000	0.997–1.000
simulated	cross-topology	TOR_ONION	QUALIFIED	sender reachable only over Tor	19250	3471	443	443	1.000	0.991–1.000	0.128	0.117–0.139
simulated	cross-topology	V2_PASSIVE_TAP	ANNOTATE	a v2 peer is missing from a pcap capture	19250	0	0	0	n/a	n/a	n/a	n/a

condition="simulated". The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.

On the base corpus 3f2d3bbe9a1f7088 (cross-topology test set), where no transaction has a stem, DANDELION_STEM fires on 857 of 19414 transactions (0.0441): every one a false alarm. It now only annotates them.

Ground-truth prevalence (variant)

condition	measure	transactions	share of broadcast
simulated	dandelion stem	21599	0.248
simulated	dandelion stem through an observer	2469	0.028
simulated	onion origin	14650	0.168
simulated	coinjoin	6562	0.075
simulated	equal value batch payout	4093	0.047
simulated	matrix rows over v2	112673	n/a
simulated	matrix rows from onion peers	1617	n/a

condition="simulated". The simulator omits: Bitcoin Core's per-peer Poisson INV trickling (one exponential delay per hop stands in for it), clock skew between observers, inbound/outbound connection asymmetry, peer churn during a capture, re-announcement, wtxid relay, user agents and real GeoIP; senders keep one address and one peer set for a whole capture, and the share of senders connected directly to an observer is a sampled parameter, not a measurement. The validity variant adds, per capture: a Dandelion stem phase in BIP-156's shape (serial single-peer forwarding for a geometrically drawn number of hops, then an ordinary broadcast; the stem successor is drawn fresh per hop rather than from two per-epoch destinations, stem hops use the ordinary per-hop delay, and there is no embargo timer), senders reachable only over Tor (one uniform circuit latency on their first hop, entry only through mixed-transport nodes), BIP-324 v2 links (a label only: the simulated v2 link is timed like v1), and CoinJoin and equal-value batch-payout transaction shapes. Onion traffic between two relays, Tor latency on later hops, and Dandelion++'s per-epoch routing remain unsimulated.