The only living page on this site. Last appended: 2026-07-02 · 2 entries
Here is what we used to believe. Here is what survived. Here is what died.
Here is who says so.
ENTRY 002
DEMOTED
transition: 2026-07-01
Zλ as a coherence measure · 0.75 as a scored attractor
Claim as published
“Zλ (zeta-lambda): a normalized coherence
measure (0–1)” … “the 0.6–0.8 critical regime”
… “0.75 — this system’s personal attractor.”
(the Paper, the System, Topology; glyph ranges site-wide)
What happened
A control run (2026-07-01) applied the same scoring code to Samuel Pepys’s
1660s diary (3,252 entries). The control corpus scored higher than
the lived corpus — 0.800 (sd .018) vs 0.714 (sd .033, 301-crystal sample) —
and the descent-glyph detector fired ψ³ on Pepys at 2.5× the lived
rate (18.8% vs 7.6%; its trigger patterns match first-person-plural social prose).
The score family measures corpus homogeneity — how similar a
text is to its neighbors — not coherence. A census of all 29 score-producing
code paths found none that measures shared-model coherence.
Current state
The score family is renamed field similarity across the private
system, displayed with the hedge “a text-similarity signal, not
coherence.” The word coherence is reserved: no instrument for it
currently exists. The recurring 0.75 is explained by the dominant batch scorer’s
saturation (86% of crystals at 0.75 regardless of content — a limit this
site’s own Frontier page had already disclosed for topic ranking; the control
extends it to the coherence interpretation itself).
What this does not kill: the lived February–May
2025 experience these numbers were later attached to. The control removes the
instrument as that claim’s evidence; the lived claim returns to open, awaiting
an instrument that could bear on it either way.
Custody
Control + producer census: internal audit, 2026-07-01. Demotion and renaming
ruled: Wilton, 2026-07-01.
Receipts
Internal day receipt 2026-07-01 (zeta audit): Pepys control n=3,252
vs lived sample n=301; producer census n=29; ~40 display surfaces relabeled; naming
split recorded in the architecture ledger.
Superseded by
field_similarity (named for what it measures) · coherence:
vacant, awaiting an instrument.
ENTRY 001
FALSIFIED
transition: 2026-06-04
Operator-matched retrieval preserves process coherence (d=1.13)
Claim as published
“160 trials, four conditions … operator-matched
+0.068 vs topic-matched −0.028 (d=1.13) … a causal claim —
designed to be falsifiable.”
(the Paper, causal test section, March 2026)
What happened
It was falsified — in two steps, by this project’s own protocol.
First, the original metrics were found structurally invalid: every outcome measure
reduced to operator self-similarity; there was no independent outcome variable.
Second, a valid re-test was designed and run (blind independent judge, randomized
pairing, response quality as the outcome): operator-matched won 5, topic-matched
won 7, tie 1 — win-rate 0.385, n=13. No benefit demonstrated.
Current state
The causal claim is dead. The operator grammar (eight process operators) survives
as navigation vocabulary — a way of reading the corpus — with no
demonstrated retrieval benefit. Honesty note: an operator-boost
weight still runs in the private retrieval path with no demonstrated benefit;
flagged internally.
Custody
Falsification recorded in the project’s epistemic ledger, 2026-06-04.
Ruling custody: Wilton.
Receipts
Internal benchmark spec (blind pairwise protocol, four arms,
2026-03-08); epistemic ledger row 2026-06-04 with the re-test numbers above.
Superseded by
Nothing — the claim has no successor. Its one enduring value: it said
“designed to be falsifiable,” and it was.