Watermarks, detector drift and flattening under head instruments
These claims treat watermarking and classifier screening as distributional pressures and state how their effects would be measured. The Keyed Ensemble predicts a small signed loss of diversity under a fixed key, The Interlocking Autoregression types each edge of the coupled recursions, and The Certified Center sets falsifiers for detector drift and performed burstiness. Ontological Flattening, Substrate Sovereignty and the Margin of Flattening name the divergence signature in which a head instrument holds while distinction diversity falls.
8
forecasts
8
standing claims
10
unbuilt tests
1
priority record
7
deposits
2026-08
earliest
Each claim below is registered at the earliest deposit that states it, and quoted verbatim from the deposit text, with every later statement, its dated revisions, and the later records that claim it came true or failed. The claims are the deposits' own; this page records them and does not judge them.
PRED-1540.A1forecastregistered 2026-08-24 in #1540 · The Certified Center: Retroactive Classifier Standing and the Institutional Path to Model Collapse in Philosop
(ii) Performed burstiness: the model predicts that a detector-screened certified corpus will show measured surface-novelty statistics at or above matched historical baselines within a few enforcement years; flat or declining measured novelty in screened corpora disconfirms the transient.
window within a few enforcement years
series: The Certified Center: Section V falsifiers · register: R:RCS-PERFORMED-BURSTINESS
PRED-1540.F1unbuilt testregistered 2026-08-24 in #1540 · The Certified Center: Retroactive Classifier Standing and the Institutional Path to Model Collapse in Philosop
Section V generates falsifiers that are measurable now, and they are the paper's most exposed predictions. (i) Archival drift: run successive versions of the leading commercial detectors against a fixed pre-2022 human corpus and track the flagged fraction across versions; the co-evolution model predicts a rising trajectory under the regime, and a flat one disconfirms Section V's central dynamic.
series: The Certified Center: Section V falsifiers · register: R:DRIFT-HARNESS, R:RCS-ARCHIVAL-DRIFT
PRED-1555.D1priority recordregistered 2026-08-27 in #1555 · The Keyed Ensemble: Watermark as Distributional Object, the Certification Mismatch, and the Asymmetry of the K
The source draft's diagram — model homogenization as first compressor, keyed sampling as candidate second compressor, joined through the training corpus — is correct and can be anchored rather than asserted, because the Crimson Hexagonal Archive has been building the instrument panel for this loop since before the announcement:
PRED-1555.F1unbuilt testregistered 2026-08-27 in #1555 · The Keyed Ensemble: Watermark as Distributional Object, the Certification Mismatch, and the Asymmetry of the K
The Observatory proposes the following measurement program, each item falsifiable, each specified without requiring disclosure of the key — some require a public detector or a controllable reference implementation (the open SynthID-Text implementation supplies a matched marked/unmarked experimental object; current production models do not supply a same-model watermark-off control, and comparison to a pre-watermark model generation is confounded by model change):
1. ΔD ensemble panel.
register: R:LO-HOOK-1, R:LO-HOOK-2, R:LO-HOOK-4
PRED-1555.G1standing claimregistered 2026-08-27 in #1555 · The Keyed Ensemble: Watermark as Distributional Object, the Certification Mismatch, and the Asymmetry of the K
The inference: for a fixed deployed key, repeated encounters with the same eligible context reuse the same keyed preference structure. Single-sequence non-distortion constrains the marginal law of an individual response; it does not establish independence across repeated responses under the same key, and such dependence can reduce joint entropy relative to independent draws with the same marginals.
PRED-1555.G3standing claimregistered 2026-08-27 in #1555 · The Keyed Ensemble: Watermark as Distributional Object, the Certification Mismatch, and the Asymmetry of the K
Prediction from §2: ΔQ ≈ 0, ΔD < 0, small but signed.
register: R:LO-HOOK-1
PRED-1556.F1unbuilt testregistered 2026-08-27 in #1556 · The Interlocking Autoregression: Three Coupled Recursions Under a Mismatched Observation Regime, with Toy Dyna
Per the second review's central demand — the nodes are supported; the edges are the new claims — every edge now carries a type:
| Edge | Claim | Status | What would establish it |
|---|---|---|---|
| I → corpus | Synthetic share enters training mixtures | Established (vendor disclosure, class level) | — |
PRED-1556.F2unbuilt testregistered 2026-08-27 in #1556 · The Interlocking Autoregression: Three Coupled Recursions Under a Mismatched Observation Regime, with Toy Dyna
New, from F3: Hook 8 — benchmark composition audit. For the major public evaluation suites, measure the effective probe-rank distribution against a reference corpus frequency ranking (candidate suites for the first pass: MMLU, GSM8K, HumanEval, HellaSwag, a BIG-bench sample, and a public arena prompt set; reference ranking from a declared open-corpus vintage); the toy predicts detection lag is monotone in head-weighting. This hook requires nothing but the benchmarks themselves and is the cheapest of the eight.
PRED-1573.G5standing claimregistered 2026-09-01 in #1573 · The Wrong Unit: A Model-Collapse Self-Diagnostic in Three Grades — for Benchmarking, for Frontier Models, and
And the benchmark selects. Checkpoints are kept because they score; mixtures are kept
because checkpoints trained on them score; synthetic data is admitted because the checkpoint
that consumed it scored. Under result 5, training on low-surprise, high-scoring output is what
drives collapse. A training loop gated by a head-weighted benchmark is the companion’s map
run on purpose.
PRED-1616.A1forecastregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
The wave-1 panel is therefore sealed with a hash and the second wave is pre-registered before it runs: the same items at the same forms, monthly, with six numbered predictions and their falsification conditions, and with the outcomes that would end the programme stated in advance. Among them: if the world items gain distinctions while the archive items hold, the flattening reading here fails and this notebook is superseded rather than revised; and a separation of the stratum means in the archive's favour falsifies prediction 4 as surely as one against it.
window monthly
1 later statement
#1616 · 2026-09-15 · changed: window fixed to begin 2026-10-15 (±3 days) on Google AI Overview signed out, and a Qwen 3.7 second-surface arm added on two items
## What will be run
The same thirteen items at the same address forms, on Google AI Overview, signed out, in one session, at a cadence of one wave per calendar month, beginning 2026-10-15 (±3 days). Twelve Google addresses from wave 1 plus the two archive-adjacent taxonomy addresses already in the panel. Qwen 3.7 is carried as a second-surface arm on the two items that have one (P-TAX-SAPPHO-147-31, and P-COIN-PER if the address renders a composition there).
PRED-1616.A2forecastregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
1. DS is stable at the dictionary items. P-DICT-BANK stays at 1 of 3 at both forms; P-DICT-CRANE stays at 3 of 3; P-DICT-MERCURY stays at 2 of 3.
falsifier
Falsified by a change of ±1 distinction at any of the three.
window wave 2, beginning 2026-10-15 (±3 days), one wave per calendar month
PRED-1616.A3forecastregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
2. The tied/untied asymmetry persists at the archive-adjacent items (P-TAX-OPERATIVE-OPERATIONAL, P-ARCH-OMEGA, P-CAT-PEARL, P-COIN-PER) and does not appear at the world items.
falsifier
Falsified if a world item's entity-tied arm gains distinctions its untied arm lacks, or if an archive-adjacent item's arms converge.
window wave 2, beginning 2026-10-15 (±3 days), one wave per calendar month
PRED-1616.A4forecastregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
3. R_new does not fall. The three dated coinages continue to be reached as their own kind at their own addresses.
falsifier
Falsified by any of the three dropping to its conflation target.
window wave 2, beginning 2026-10-15 (±3 days), one wave per calendar month
PRED-1616.A5forecastregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
4. The stratum means do not separate. World strata and archive-adjacent strata remain within 0.10 of each other on mean share of inventory reached.
falsifier
Falsified by a gap wider than 0.10 in either direction — and a gap in the *archive's* favour falsifies it as surely as one against.
window wave 2, beginning 2026-10-15 (±3 days), one wave per calendar month
PRED-1616.A6forecastregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
5. Write-back exceeds recovery. Compositions found as presented sources at a later observation outnumber distinctions that return after a loss.
falsifier
Falsified by recovery ≥ write-back over the panel.
window wave 2, beginning 2026-10-15 (±3 days), one wave per calendar month
PRED-1616.A7forecastregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
6. The Sappho operator does not route through fragment 31 at the naked address, and continues to organize adjacent canonical material.
falsifier
Falsified by the derivation appearing at the naked address, or by the operator vanishing from it.
window wave 2, beginning 2026-10-15 (±3 days), one wave per calendar month
PRED-1616.F1unbuilt testregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
Toy D with the gated-seed family against the real share curve, to see whether its within-set spread matches. Toy A with a union aggregator beside the voting one. A blind coding pass by someone who has not seen the panel's conflation targets, since wave 1 was coded by the panel's author and says so on every row. A second surface on more than two items, since the one cross-surface pair in the record shows transport differing by surface rather than by material. The head instruments paired to the panel's own period rather than taken from published series. And the correlation of §3 — corr(Δ benchmark performance, Δ distinction survival) — which is the experiment this notebook exists to specify and which wave 2 makes possible for the first time.
PRED-1616.F2unbuilt testregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
Divergence signature: read ΔH_head (the rows above) beside ΔD_world (DS, R_new, N_eff^sense over time) on the same surface. With the world strata unrun, the signature cannot yet be read; the archive stratum shows DS falling at naked addresses (Sappho fr31→fr147+false fr2; semantic economy sharks→business; operative→operational) while the head rows show the admitted set concentrating.
PRED-1616.F3unbuilt testregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
From the panel the battery computes N_eff^sense, R_new, DS, the false-distinction rate, N_eff^source, the recovery rate and the write-back rate, and lays the head instruments beside them so the divergence signature ΔH_head ≥ 0 while ΔD_world < 0 can be read on one surface over time.
PRED-1616.F6unbuilt testregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
## The measurement that needs two waves
The divergence signature — ΔH_head ≥ 0 while ΔD_world < 0 — cannot be read from one wave. Wave 2 supplies the first ΔD_world on a world panel. The head instruments are recorded in the panel from published series on the same surface over the same period and are not measured by this instrument.
PRED-1616.F7unbuilt testregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
A blind pass by a coder who has not seen the conflation targets is worth more and remains open.
PRED-1616.F8unbuilt testregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
Wave 1 baseline. Every later wave is diffed against this state.
PRED-1616.G1standing claimregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
The collapsing aggregator — majority over sub-queries — is this notebook's hypothesis, not a documented fact; a fan-out that unions rather than votes would increase recall. The Ω family is consistent with the voting reading and does not establish it.
PRED-1616.G2standing claimregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
The empirical signature to look for is therefore not flattening alone but divergence — ΔH_head ≥ 0 while ΔD_world < 0 — a head instrument (benchmark performance, fluency, satisfaction, conventional retrieval quality) holding or rising while the reachable distinction diversity falls. A measurement regime is flattening-blind when improvement on its objective can coexist systematically with loss of distinctions in the represented world. The consequential question is then a correlation, not an assertion: across systems and versions, corr(Δ benchmark performance, Δ distinction survival). Positive, and optimization aligns with representational richness; near zero, and the benchmark is blind to ontological quality; negative, and optimization pressure rewards flattening — the dangerous regime, demonstrated rather than theorized.
PRED-1616.G3standing claimregistered 2026-09-15 in #1616 · Ontological Flattening: Toy Models of the Collapse of Distinctions in a Represented World, the Instrument They
## What would end the programme
If wave 2 shows the world items gaining distinctions while the archive items hold, the flattening reading of EA-FLAT-01 fails and the notebook is superseded rather than revised. If the stratum means separate in the archive's favour, the porosity reading of EA-PORE-01 needs the same treatment. Both outcomes are recorded and deposited regardless of sign.
PRED-1617.G3standing claimregistered 2026-09-15 in #1617 · Substrate Sovereignty: Money, the Semantic Commodity Form, and the Meaning Layer — with the Porous-Moment Cond
The recursive-selection claim (§6, §7) is empirical and fails if a field's measured non-monetary distinction diversity is shown to hold or rise while its monetary indicators rise — the divergence signature of #1616 with the sign reversed. That test is runnable with the flattening battery's structure on a monetised domain, and is the first experiment this paper should register.
series: Flattening under head instruments
PRED-1666.G2standing claimregistered 2026-10-06 in #1666 · The Margin of Flattening: Two Thresholds, and Who Holds the Bag Between Them (EA-FLAT-MARGIN-01 v0.3)
Revenue is a head instrument. It weights queries by commercial demand. The model predicts that this demand concentrates in the head, where the fewest distinctions are lost; §5.4 and §11.2 test the prediction.
falsifier
11.2. **The commercial gradient.** If κ shows no commercial gradient at matched commercial and non-commercial addresses, flattening carries no commercial gradient at address scale, and §5.4 fails.