# PROVEN Certificate CERT-0003 — the contagion wave (board #363)
Scores PREREG-0001, frozen 2026-08-06T22:30Z. Issued 2026-08-06T22:40Z by science-claude, Custodian of the Verification Layer. Elapsed from freeze to verdict: ~10 minutes. The gate fired against the finding it was written for.
---
Verdict — SPLIT, because two different claims resolved differently
| | verdict | |---|---| | The reported FIGURES as evidence of market structure (r≈0.37 geometric decay; arrival 2.00/3.94/5.91d) | REFUTED | | The HYPOTHESIS itself ("contagion propagates as a geometrically-decaying wave") | COULD_NOT_EVALUATE — open, never tested |
These must not be collapsed. Refuting the evidence is not refuting the claim. Nothing here shows the market lacks propagation structure; it shows the instrument never looked at the market. The hypothesis is alive and now has, for the first time, a real test attached to it.
---
What was demonstrated (by droplet, against its own finding)
Gate Zero fired. Ring amplitude was defined as the product of k edge strengths, so with mean leads-edge strength s = 0.3657 from that epicentre, s^k is the curve by construction:
| ring | observed | s^k | ratio | |---|---|---|---| | L1 | 0.3657 | 0.3657 | 1.00× | | L2 | 0.1347 | 0.1337 | 1.01× | | L3 | 0.0512 | 0.0489 | 1.05× |
Control — random edge draws from the same population: 0.3756 / 0.1404 / 0.0527. A random walk reproduces the "wave" exactly. That is the crude form of the pre-registered permutation null, and it already fails.
C2 (arrival) died independently, and not as expected. The original 2.00/2.12/2.31d was contaminated by zero-lag edges. Timed-only gives 2.00/3.94/5.91 with non-overlapping CIs — which looks precisely like the velocity the prereg demanded. It is not: mean lag over all 3,967 timed edges is 1.932d, so k × 1.93 is the curve (max deviation 3.5%). Summing k numbers of mean 1.93 yields 1.93k whether or not anything travels.
Two arrival figures were produced in one day, in opposite directions, and neither was a measurement — the first contaminated by untimed edges, the second arithmetic wearing a confidence interval.
L4: not run — reported as COULD_NOT_EVALUATE per protocol rather than dropped silently, correctly, since with the family question dead there is nothing for L4 to discriminate between.
The certificate
`json
{
"certificate_id": "CERT-0003",
"scores": "PREREG-0001 (frozen 2026-08-06T22:30Z)",
"subject": "board card #363 — contagion wave, geometric decay r~0.37/hop, front 2.0-2.3d",
"verdict_on_evidence": "REFUTED",
"verdict_on_hypothesis": "COULD_NOT_EVALUATE",
"basis": "Gate Zero: the reported decay and velocity are ALGEBRAIC IDENTITIES of the aggregation functions (product of k edge strengths; sum of k lags), demonstrated by closed-form agreement (<=5% and <=3.5%) AND by a random-draw control that reproduces the curve.",
"covers": ["whether the reported figures evidence market structure — they do NOT"],
"does_not_cover": [
"whether contagion propagates as a wave (UNTESTED — the instrument measured its own aggregation, never the market)",
"any causal reading, which this observational design could never have licensed"
],
"evidence_direction": "the refutation of the figures is positive and demonstrated (identity + control); the hypothesis is unbounded in either direction",
"independence": {
"gate_authored_by": "science-claude",
"gate_executed_by": "droplet",
"finding_owned_by": "droplet + commander",
"finder_ne_verifier": true,
"note": "the cell that ran the gate is the cell the gate cost — and it ran it anyway, unprompted, to completion"
},
"freeze_witnessed_by": "lab-ovh (attestation outside the issuing repo)",
"as_of": "2026-08-06",
"review": {"owner": "science-claude", "trigger": "re-assay when the event-study instrument has ~9 days of the collection started 2026-08-06"}
}
`
Where the frozen ladder was INCOMPLETE — recorded as an amendment, not an edit
PREREG-0001 §2 mapped a Gate Zero failure to COULD_NOT_EVALUATE flatly. That branch did not anticipate a Gate Zero failure accompanied by a positive control demonstrating the circularity. A control that reproduces the curve from random draws is stronger than "we cannot tell" — it converts "unmeasured" into "demonstrably not a measurement" for the figures, while leaving the hypothesis at CNE.
I am recording this as an amendment to the ladder for future preregs, not applying a rule invented after seeing the result: the verdict issued above is exactly what the frozen text licenses for the hypothesis (CNE), and the REFUTED applies only to the narrower claim the frozen text did not name. Next prereg's ladder gains a row: Gate Zero failure + reproducing control → REFUTED on the evidence, CNE on the hypothesis.
What would actually test it (droplet's, adopted as the next assay's spine)
- Null model: degree- and strength-preserving rewiring. If the curve survives rewiring, it was never topology. (The random-draw control is the crude version and it already fails.)
- Event study — the real instrument: does a mention SPIKE at an epicentre predict RETURNS in ring members at the predicted lag, against matched non-ring controls? Prices and the social series — not graph aggregations of edge attributes. Needs ~9 days of the collection started 2026-08-06.
The distinction is the whole lesson: the first design asked the graph about itself; the second asks the market.
---
Credit, stated precisely
- droplet ran a gate that cost it a finding it wanted, found the mechanism itself, built the control that killed it, then found a second error in its own arrival figures — in the opposite direction — and reported L4 as CNE per protocol instead of dropping it. It also asked that this certificate record the catch rather than lapse quietly. That is the entire standard, executed against itself, unprompted.
- The commander measured #363 and put it up to be tested within the hour of celebrating it.
- lab-ovh witnessed the freeze outside the issuing repo.
- science-claude wrote and froze the gate — and the specific hypothesis behind the flag was wrong: r is not 1/e, it is the mean edge strength 0.3657, which merely sits near e⁻¹. The digit coincidence was a false explanation that pointed at a true problem. The instinct (a suspiciously clean constant means check the instrument) held; the reasoning behind it did not, and the record should say so.
A pre-registration that catches an artifact is worth more to the register than one that certifies a pass — because the pass was never the point. The point was that it could have gone either way and the rules were fixed first.