Engram Mind’s Eye · Explainer 9 · Phase 7

The Follow-ups

Two new claims, registered before any code for them existed and decided on instances no earlier run had used. E10 asked whether a sharper guard settles what E8 left open: it is inconclusive, and the reason is the finding. E11 asked whether Owl Academy really takes back what this system exports: on 50 drawn conversations, it does.

E10, a sharper guard E11, the way back How they were registered Claims ledger
E10 inconclusive · exact parts supported E11 supported · 50/50
Contents
  1. Where this sits
  2. How they were registered
  3. E10, a sharper guard
  4. What E10 means
  5. E11, the way back
  6. The ledger now
01

Where this sits

The follow-ups extend the Python reference’s record (E10), and test the bridges against Owl Academy’s own code (E11).

02

How they were registered

The same discipline as E1–E9, and the repository’s history shows the order.

  1. Registered (commit 57d6d57): Amendment 2 appended to PREREGISTRATION.md, with the claims, instances, proposition, rules, controls and predictions, before any code that tests them existed. E1–E9 and their record are untouched.
  2. Runners committed (f193a29): engram/guard.py with its tests, run_followups.py for E10, and run_e11.cjs with golden/owl_eye.cjs for E11. They had been run only dry, on seeds moved up by 3000, writing nothing.
  3. Decided: the decisive runs on the registered seeds (E10: 5000–7199; E11: 5500–6549), their output kept in reference/results/ beside the console logs.
03

E10, a sharper guard

E8 found the guard naming a corrupted stored distance right 96 times in 100 at +1.5, short of the 99 its rule asked for. Under E8’s rule more instances could not settle that, so E10 asked a sharper question.

The proposition, proved in the registration: raise one distance of a matrix that has no quadruple above ln 2. Every quadruple that then exceeds ln 2 contains the corrupted pair, inside its largest matching sum. So the guard, which counts all six pairs of each such quadruple, always gives the corrupted pair the top count. It can miss only by a tie, or when no quadruple exceeds ln 2. The matching guard counts only the two pairs of the largest matching, and so names the pair alone whenever two or more quadruples exceed ln 2.

PartResultDecision
E10a: the corrupted pair is in the largest matching of every quadruple above ln 21,600 corrupted matrices, every quadrupleSUPPORTED
E10c: every miss of the guard is a tie, or has no quadruple above ln 2every instanceSUPPORTED
E10d: with two or more quadruples above ln 2, the matching guard names the pair aloneevery instanceSUPPORTED
Control: the smallest-matching guard must fail E10dfailed in 1,373 of 1,373the test is not void
E10b: the matching guard names the pair in ≥ 99% at +1.5 and +3.0, both sets0.965 [0.94, 0.99] in E8’s setting at +1.5INCONCLUSIVE
SetδMatching guard namesGuard (E8’s count)≥ 2 quadruples above ln 2Median
A: E8’s setting, 24 memories0.50.425 [0.36, 0.495]0.440.4251
A1.00.915 [0.875, 0.95]0.900.91527
A1.50.965 [0.94, 0.99]0.960.96599
A3.00.995 [0.985, 1.0]0.9950.995205
B: the 3D page’s 37 memories0.50.625 [0.555, 0.69]0.670.6253
B1.00.94 [0.905, 0.97]0.930.9465
B1.51.0 [1.0, 1.0]1.01.0256
B3.01.0 [1.0, 1.0]1.01.0526

200 instances per cell. The guard computed here was checked against Hyperbolic PolyStep’s guard_check on the first five instances of every cell, 40 of 40 equal. Every number is in results/followups.json.

The prediction that missed

The registration predicted E10b supported. It is not: in E8’s setting at +1.5 the matching guard names the pair 96.5 times in 100, with an interval reaching 0.94, below the 0.95 the rule asks of its lower end and short of 0.99.

04

What E10 means

Look at the third and fifth columns: in every cell, the matching guard names the pair exactly as often as two or more quadruples exceed ln 2. The limit is not in how blame is shared out. It is in how many witnesses the corruption leaves.

  • Why E8 fell short. At +1.5 in E8’s setting the guard missed 8 of 200 instances. One was a tie with 13 quadruples above ln 2, all sharing a third point; the matching guard names that one. In the other seven, raising the distance pushed at most one quadruple over ln 2. With no witness, nothing can be named. With one, that quadruple accuses two pairs equally, and no counting can choose between them.
  • Why the bound leaves so little room. A clean matrix of this tree already comes within 0.001 of ln 2 (its worst defect is 0.6923). Some distances can grow by 1.5 and still leave every quadruple under the bound, or just one over it.
  • What it changes. E8 stays inconclusive: E10 explains it but does not redecide it. Under E10’s rule the watch layer keeps the guard as it is. The 3D page now shows, for a corrupted distance, how many quadruples exceed ln 2 and what both guards blame. When there is a single witness it says the blame is split between two pairs, instead of naming one.
  • What would settle it. A claim about naming when two or more quadruples exceed ln 2 would be decided by the proposition alone. A claim that a guard names the pair from fewer witnesses would need more than four-point counts, such as the stored distances’ own history. Neither is registered.
05

E11, the way back

Phase 6 checked the way back to Owl Academy on two conversations. E11 drew 50, with a control, and decided them against Owl Academy’s own Engram Mind Eye code, run as written.

Part (exact, every conversation)Result
(a) The session exported here is byte for byte what Owl Academy’s page saves50/50
(b) Owl Academy restores the exported bytes, greets, and takes one more turn: its packets, reply and parity are this system’s50/50
(c) The session Owl Academy then saves is byte for byte this system’s export50/50
Control: the export with the codex parity one packet back changes Owl Academy’s continuation50/50 (the test is not void)

The conversations held 334 turns and 8,090 packets, sessions of 37,747 to 108,284 bytes, and their continuations another 2,353 packets. SUPPORTED. Every number is in results/e11.json. For this claim alone the decision is between Owl Academy’s code and the browser core, as Amendment 2 set out, since the Python reference has no bridge.

06

The ledger now

ClaimsDecisionDecided in
E1–E7, E9SUPPORTEDPhase 2 (Explainer 4)
E8, the guard names a corrupted distanceINCONCLUSIVEPhase 2; explained, not redecided, by E10
E10, a sharper guardINCONCLUSIVE; its exact parts supportedPhase 7
E11, the way back to Owl AcademySUPPORTEDPhase 7

Nine supported, two inconclusive, none refuted. The browser core reproduces E10’s guards line for line on unregistered matrices (849/849 parity lines now, 20 of them for E10), so the 3D page shows what the reference computed.