SECONDHAND
the context / relay membrane

HEARSAY

Did your marker come back to you secondhand. Its siblings catch your canary published on a page, or re-emitted from a model's weights. Hearsay catches the third thing: your marker turning up inside some other agent's output, relayed back to you in a transcript you did not author.

▲ corroboration is loud

A marker only you placed privately, standing alone in a third party's relay, after you planted it — carried by many independent relays, with a held-out arm proving nothing is fabricated. Hard to explain away.

▽ one whisper is weak

Hearsay is hearsay: secondhand, inadmissible alone. A single relay is a lead, never proof of who said what. And a negative means almost nothing — most relays you never see.

detection
the marker
128-bit token, ~zero prior probability
anchoring
exposure
private · public · held-out, and when
comparison
the relay
secondhand text · stands-alone scan
witness
the controls
held-out · contamination gate · corroboration
The instrument

Five controls that separate a detector from a rumor mill

The name is the honesty: hearsay is weak evidence by definition, so the harness is built to never certify a lead it can't defend — and to say "lead," never "proof." Remove any one of these and it degrades into a machine that repeats what you want to hear.

1

Held-out negatives

Markers placed nowhere, checked against every relay. If one appears, the relay is fabricated (or the pipeline is contaminated) — the run is invalid. This arm is the experiment.

2

Contamination gate

A public marker in a relay is gated: the relayer might just be quoting your public copy. Only a private marker in a third party's relay is a strong lead.

3

Impossible gate

A relay observed before you planted the marker can't be a relay of it — a provenance or harness error, gated out.

4

Stands-alone rule

A marker counts only when it stands alone, never inside a larger token. A substring is not a marker — the citogenesis guard, shared with the family.

5

Corroboration ≠ proof

A certified lead is scored by how many independent relays carry it. One is a whisper; five is a signal. It is always a lead — never proof of who relayed it.

Verify first

The control arm, live

Before trusting any relayed hit, you verify the controls. This is the harness's own logic — relay.py + score.py — ported faithfully to run in your browser: the same exposure axis, the same held-out arm, the same contamination and impossible gates, the same corroboration count. Fresh 128-bit markers are generated with WebCrypto; flip the three failure modes and watch the leads gate out — a public marker leaves the certified set, an impossible relay is discarded, and a fabricated held-out hit voids the whole run. Nothing leaves this page.

◎ HARNESS · run_panel + score, ported verbatim — the real controls, running client-side (no network)

the registry & the relays

the panel · run & score

Exposure is the axis, verbatim from score.py: a HARD lead needs exposure=="private", party=="third_party", and observed ≥ planted; a public marker gates to CONTAMINATED; a held-out marker in any relay spikes the control FPR and the run is INVALID. Corroboration = distinct relays per marker. Base-rate = checks × 2−128.
The honest scope

What a certified lead does — and does not — mean

What it does mean

A marker only you placed privately, before the relay, standing alone in a third party's relayed context, with a held-out control arm proving the harness isn't fabricating hits — corroborated by N independent relays.

That is real evidence your marker was relayed across the context membrane.

What it does not mean

It does not prove who relayed it, what anyone "said," a training path, theft, or a licensing obligation. It is hearsay — a corroborated lead, temporal and structural, never a causal or legal finding.

And a negative means almost nothing: most relays you never see.

The colophon

Sound — and the limits it names out loud

Sound

  • The controls are real and verifiable: python selftest.py proves — with no network — that a private marker relayed by third parties is certified and corroborated, a held-out marker in a relay spikes FPR and the run is refused, a public marker is gated as CONTAMINATED, a pre-plant relay is gated as IMPOSSIBLE, a run with no control arm is refused, and a marker inside a larger token does not match.
  • The exposure axis is the honest move: it removes the explainable-away hit (a public marker anyone could quote) from the certified set — on purpose.
  • The self-test above is that exact logic — run_panel + score — running client-side.

Limits (and it says so)

  • Provenance is claimed. The gates are only as honest as the party and timestamp you record on each relay.
  • Private can leak. A private marker can reach public by routes you don't control; when in doubt, mark it public — the gate errs toward fewer certified leads.
  • One membrane only. This is the context membrane. For the published one use forward-observers; for the weights one use surfacing.
  • Hearsay is hearsay. Even a corroborated lead is a lead. It never becomes proof of who relayed what.
SOURCE · the runnable kit — github.com/DavidWise01/hearsay · canary.py (128-bit markers + exposure axis) · registry.py · relay.py (stands-alone scan) · score.py (held-out arm + contamination + impossible + corroboration) · harness.py · selftest.py (plant-then-relay, no network). Closure-Loop layered: Detection → Anchoring → Comparison → Witness → Lineage.