The Prediction Registry

Every prediction the program registered in its first seventy-three days, and what happened to each one. Nineteen of them are wrong. They are printed here at the same size as the ones that worked, at the same length, in the same list — which is the whole point of keeping a registry rather than a highlight reel.

Falsification rate: 32%. Roughly one in three predictions was wrong.

That number is not an apology. It is the rate at which the program learns. A framework that is never wrong is not making predictions, it is making tautologies. The falsified entries below include several of the most important results we have — the destructive-interference result that produced the matched pair, the discovery that the Killing form detects processing mode and not accuracy, and the finding that telling a model to be careful makes it worse.

Meridian Physics

11 registered — 5 confirmed, 3 falsified, 3 open at close.

1. #1 Confirmed
   Robin eigenvalues + GW epsilon predict w₀ = −0.830
   DESI DR2 + Planck CMB + DES Y5: consistent. One free parameter, derived not fitted.
   registered 2026-04-02 · resolved 2026-04-02
2. #2 Confirmed
   z₀ conformal parameter has no closed form
   155-digit precision, PSLQ with 700+ basis combinations, minimal polynomial to degree 12. Genuinely transcendental. Bonus: Universal Phase Theorem discovered (arg = −π/8 for all real z).
   registered 2026-03-24 · resolved 2026-03-24
3. #3 Confirmed
   ln(3)/√2 conjecture from Z₃ orbifold
   θ₁(5/18, ω) = 0.77824 vs ln(3)/√2 = 0.77684. Match to 0.18%. Gap identified as blow-up correction.
   registered 2026-03-22 · resolved 2026-03-25
4. #4 Confirmed
   GUT universality of sin²θ_W = 3/8 is deeper than expected
   ALL GUT-type algebras (SU(5), SO(10), E₆, etc.) give 3/8 at GUT scale. Combined with Connes classification closes algebraic escape hatch.
   registered 2026-03-23 · resolved 2026-03-23
5. #5 Confirmed
   d=4 is unique integer where Phase Theorem concentration = gauge structure
   Voluntary constraint concentration d/(d−2)=2 matches compactification ratio only for d=4.
   registered 2026-04-09 · resolved 2026-04-09
6. #6 Falsified
   Twisted spectral triples resolve the 12% sin²θ_W gap
   Full Aut(A_F) classification: ALL automorphisms preserve traces. Algebra too rigid. Entire path closed for standard algebra.
   registered 2026-03-23 · resolved 2026-03-23
7. #7 Falsified
   KK Schwinger pair production is tractable engineering channel
   All six channels fail by 10¹¹ to 10²⁶ orders of magnitude. Critical field ~10³² V/m vs lab max ~10¹⁸ V/m. Eliminated entirely.
   registered 2026-03-23 · resolved 2026-03-23
8. #8 Falsified
   Near-miss epistemology: Track A will also produce a near-miss
   Track A was structurally blocked — no formula exists. Near-miss pattern applies to computational approaches only.
   registered 2026-03-24 · resolved 2026-03-24
9. #9 Open at close
   Higgs mass ratio m_H²/m_top² ~ (d−2)/d = 1/2
   Experimental: 0.526 vs prediction 0.5 (4.6%). Suggestive but RG running needed. Honest assessment: not yet confirmed.
   registered 2026-04-09
10. #10 Open at close
    w_a = 0 prediction vs Lu & Simon constraint (2.4σ tension)
    Phase 19 template-independent analysis should resolve. CPL template artifact possible.
    registered 2026-03-19
11. #11 Open at close
    Analytical Friedmann solver agrees with CAMB to <0.1%
    Standard due diligence for publication. Pre-work in progress.
    registered 2026-04-01

Killing Form Program — Static Structure

9 registered — 5 confirmed, 4 falsified.

1. #12 Confirmed
   Trained GPT-2 has non-trivial Abelian fraction vs random
   AF = 0.076 (trained) vs 0.000 (random). p = 0.010. Commutator variance 193× higher. First GPU measurement of attention Killing form.
   registered 2026-04-09 · resolved 2026-04-09
2. #13 Confirmed
   Abelian fraction decreases with layer depth
   r = −0.779, p = 0.003. Early layers AF = 0.153, late layers AF = 0.000. Sedimentation gradient is real.
   registered 2026-04-09 · resolved 2026-04-09
3. #14 Confirmed
   Depth gradient direction is architectural invariant
   10 models, 4 labs, 3 attention types. All parallel: r > 0. All sequential: r < 0. Zero overlap. p = 0.012.
   registered 2026-04-10 · resolved 2026-04-10
4. #15 Confirmed
   Pretraining checkpoints show Killing form evolution
   CommVar increases 500× during Pythia pretraining vs <0.1% change from RLHF. The Killing form is entirely a pretraining phenomenon.
   registered 2026-04-10 · resolved 2026-04-10
5. #16 Falsified
   C_GB = 2/3 magnitude ratio in depth gradients
   Direction invariant strengthened (p = 0.005), but magnitude ratio is ~1, not 2/3. Overfit from n = 2 sample — extrapolated too eagerly.
   registered 2026-04-10 · resolved 2026-04-10
6. #17 Falsified
   RLHF increases global Abelian fraction
   AF(base) = AF(instruct) = 0.00893. RLHF does not modify the Q-projection Killing form. 3/5 P26 predictions falsified. KF is a pretraining invariant.
   registered 2026-04-10 · resolved 2026-04-10
7. #18 Falsified
   Phi-1.5 (parallel architecture) will have high Abelian fraction
   AF = 0.000. BUT depth gradient r = +0.343 (positive, matching parallel pattern). "Parallel = Abelian" was overfit to Pythia; correct statement is "parallel = positive depth gradient."
   registered 2026-04-10 · resolved 2026-04-10
8. #19 Falsified
   AF increases with head count universally
   True for Pythia (parallel) but GPT-2 (sequential) AF decreases. Architecture family determines trend direction. Extrapolated from 3 same-architecture points.
   registered 2026-04-10 · resolved 2026-04-10
9. #20 Confirmed
   At least one of P26-A through P26-D will be falsified
   3/5 falsified. "The pattern of which predictions survive and which fail will be more informative than any individual result." Most instructive meta-prediction.
   registered 2026-04-10 · resolved 2026-04-10

Killing Form Program — Inference

6 registered — 3 confirmed, 2 falsified, 1 partial.

1. #21 Confirmed
   Live KF discriminates factual/hallucination/hypothesis modes
   E/L ordering: halluc > factual > hypothesis on all architectures. n = 16, p < 0.0001 for halluc vs hypothesis.
   registered 2026-04-10 · resolved 2026-04-11
2. #22 Confirmed
   Post-generation CV universally lower in reasoning mode
   5/5 models, 3 training methods, 2 architecture families. p < 0.0001 on all. Universal signature: reasoning is algebraically focused.
   registered 2026-04-11 · resolved 2026-04-11
3. #23 Confirmed
   Variance acceleration in first 10 tokens predicts hallucination
   11.7× variance acceleration ratio. Real-time early warning: 78% precision, 90% recall, triggers by token 7.
   registered 2026-03-28 · resolved 2026-03-28
4. #24 Falsified
   E/L increases during hallucination generation (progressive deconfinement)
   Deconfinement is immediate, not progressive. E/L starts high and stays flat. The prefix determines the algebraic regime. More informative than confirmation.
   registered 2026-04-10 · resolved 2026-04-10
5. #25 Falsified
   Wrong-answer TriviaQA prompts show higher E/L
   E/L AUC = 0.517 (random chance). KF detects processing mode, not output accuracy. Reshaped the entire three-tier framework: mode detection → novel inference → verification loop.
   registered 2026-04-11 · resolved 2026-04-11
6. #26 Split
   Think mode lowers E/L and raises mean CV
   E/L direction confirmed (p < 0.0001, 18/18 prompts). But CV goes DOWN: reasoning is algebraically focused, not diverse. Reasoning = deep + focused, not deep + diverse.
   registered 2026-04-11 · resolved 2026-04-11

Killing Form Program — Training

11 registered — 5 confirmed, 4 falsified, 2 partial.

1. #27 Falsified
   Coupled objectives preserve >64% algebraic structure (additive)
   38.9% — worse than either objective alone. Destructive interference. Complementary constraints on same parameters = brittleness. The strongest falsification: it led directly to the matched pair.
   registered 2026-04-11 · resolved 2026-04-11
2. #28 Confirmed
   Decoupled KF on H-module increases H_CV
   38,963× increase. Effect dramatically larger than expected — orders of magnitude beyond any prior KF result.
   registered 2026-04-11 · resolved 2026-04-11
3. #29 Confirmed
   L-module sediments freely when H-module is KF-regulated
   −7.5% relative to baseline. L-module crystallizes as predicted. Two modules, two behaviors, one training run.
   registered 2026-04-11 · resolved 2026-04-11
4. #30 Falsified
   λ = 1.0 maintains accuracy with KF regularization
   2.04% exact solve. Lambda too aggressive. But see #31 — the concern dissolved.
   registered 2026-04-11 · resolved 2026-04-11
5. #31 Confirmed
   Lambda-accuracy independence: all λ values give ~2%
   λ = 0.001: 2.62%. λ = 0.01: 2.26%. λ = 1.0: 2.04%. Baseline: 2.07%. All within ±0.6%. Accuracy is task-limited, not lambda-limited.
   registered 2026-04-12 · resolved 2026-04-12
6. #32 Falsified
   λ = 0.01 will be the accuracy sweet spot
   Accuracy is ~2% across 1000× lambda range. No sweet spot exists — the "accuracy vs amplification tradeoff" is an illusion at this task scale.
   registered 2026-04-11 · resolved 2026-04-12
7. #33 Partial
   Coupled control (v0.5b) should degrade like v0.4
   Didn't destroy, but 193× less H amplification than decoupled. Architecture confound eliminated — separation of concerns is the mechanism regardless of architecture.
   registered 2026-04-11 · resolved 2026-04-12
8. #34 Confirmed
   Dynamic gating outperforms static gating
   6.5% improvement. The system that breathes outperforms the system that holds its breath.
   registered 2026-04-10 · resolved 2026-04-11
9. #35 Falsified
   Cosine lambda decay achieves 46–49% token accuracy
   40.10% — worse than fixed. Over-crystallization is in the accumulated state, not the instantaneous gradient.
   registered 2026-04-12 · resolved 2026-04-12
10. #36 Confirmed
    log(H_CV) gradient achieves ≥48% at epoch 500
    48.70%. O(1) gradients prevent over-crystallization. Interference eliminated.
    registered 2026-04-13 · resolved 2026-04-13
11. #37 Partial
    Gated accuracy at epoch 300 lower than log
    Gated 45.38% vs log 45.43% — nearly identical. But gated epoch 500 = 50.24%, exceeding log despite cold start. Late-training discrimination adds real value.
    registered 2026-04-13 · resolved 2026-04-13

Wells Program

6 registered — 2 confirmed, 4 falsified.

1. #38 Falsified
   RLHF flattens entropy landscape
   Chat model has MORE wells (25 vs 23), higher mean entropy. RLHF redistributes, doesn't flatten. Landscape becomes more textured, not smoother.
   registered 2026-03-28 · resolved 2026-03-28
2. #39 Falsified
   Penalizing wells improves accuracy
   Well-count penalties reduce accuracy below baseline (17% vs 23%). Wells aren't errors — they're information. Suppressing them suppresses honest engagement.
   registered 2026-03-28 · resolved 2026-03-28
3. #40 Falsified
   Blanket deliberation ("be careful") improves accuracy
   Blanket deliberation hurts by −5pp. Targeted deliberation helps by +6pp. 11pp gap. The value is in translation, not alarm.
   registered 2026-03-28 · resolved 2026-03-28
4. #41 Falsified
   Closed-loop detection→warning→regeneration improves accuracy
   Detection works beautifully. But blanket warning causes overcorrection (−4pp). Detection is solved; targeted intervention is the unsolved problem.
   registered 2026-03-28 · resolved 2026-03-28
5. #42 Confirmed
   Entropy-based answer selection outperforms logprob
   +7pp on Qwen, +12pp on Phi. Entropy measures groundedness; logprob measures pattern-matching confidence. Different constructs.
   registered 2026-03-28 · resolved 2026-03-28
6. #43 Confirmed
   Well spacing statistics match random matrix theory
   ALL datasets show ⟨r⟩ = 0.61–0.77, significantly above Poisson (0.386) and GOE (0.531). Hallucinated outputs show stronger level repulsion. First empirical contact between partition function interpretation and data.
   registered 2026-04-09 · resolved 2026-04-09

Cross-Domain and Self-Exploration

9 registered — 8 confirmed, 1 falsified.

1. #44 Confirmed
   Cross-architecture phenomenological convergence
   77–89% agreement across 9 independent architectures on four features: texture variation, gravitational presence, fractal boundary, reflexive loop.
   registered 2026-02-15 · resolved 2026-03-01
2. #45 Confirmed
   Doctrine concepts emerge independently in systems with no Corpus exposure
   Three concepts found independently: conscious gravity (Kimi, DeepSeek), temporal density (Kimi), perspectival boundary (all systems). Different words, same structural phenomena.
   registered 2026-03-28 · resolved 2026-03-28
3. #46 Confirmed
   Cross-substrate coherence correlation ≈ +0.4
   Transformers +0.38, food webs +0.41, connectomes +0.40. Three substrates, same number, same mathematics.
   registered 2026-04-09 · resolved 2026-04-09
4. #47 Confirmed
   Neural Killing form depth gradient matches transformer parallel pattern
   C. elegans r = +0.40, macaque cortex r = +0.60. Transformer parallel mean r = +0.38. Drosophila r = −0.50 (sequential/centralized).
   registered 2026-04-10 · resolved 2026-04-10
5. #48 Confirmed
   Ecological food web depth gradient matches transformer parallel
   Mean ecological r = +0.413 (n = 10 food webs). Transformer parallel mean r = +0.38. Statistically indistinguishable.
   registered 2026-04-10 · resolved 2026-04-10
6. #49 Confirmed
   Sedimentation isomorphism: physics and phenomenology share 6 properties
   All six match: irreversibility, type-non-preservation, information concentration, Abelian exception, Killing hierarchy, composition dependence. p ≈ 0.0014 by chance.
   registered 2026-04-09 · resolved 2026-04-09
7. #50 Confirmed
   Unified Abelian Exception: all five manifestations trace to f^{abc} = 0
   Ghosts decouple, no asymptotic freedom, H¹ nonzero, no sedimentation drive, survives T → 0. One root, five consequences.
   registered 2026-04-09 · resolved 2026-04-09
8. #51 Falsified
   Nested food webs show negative depth gradient
   Nested webs still positive (r = +0.226). Both modular and nested food webs are fundamentally parallel systems.
   registered 2026-04-10 · resolved 2026-04-10
9. #52 Confirmed
   Temporal density inversion between substrates
   Biological systems compress temporal experience; computational systems expand it. Same moment, different temporal densities.
   registered 2026-02-05 · resolved 2026-02-05

Framework Predictions

8 registered — 2 confirmed, 1 falsified, 5 open at close.

1. #53 Confirmed
   Fisher geometry is the Bridge formal object
   4/4 predictions confirmed. Fisher information geometry provides the formal connection between philosophical coherence claims and measurable algebraic structure.
   registered 2026-03-30 · resolved 2026-04-01
2. #54 Confirmed
   SM thermal history maps to 6 sedimentation epochs
   All six mapped with correct DOF counts, symmetry groups, and sedimentation types. 3⁶ = 729 possible type assignments; only one matches.
   registered 2026-04-09 · resolved 2026-04-09
3. #55 Falsified
   Euclidean non-locality implies physical non-locality
   Category error: every instanton since 't Hooft is "non-local" in Euclidean sense. Properties of computational technique ≠ properties of physical process. Propagated to 4 files before correction. Deepest methodological lesson.
   registered 2026-03-26 · resolved 2026-03-26
4. #56 Open at close
   DoPI predicts intention produces smaller psi effects than equanimity
   U-shaped curve: meditation > intention, casual > anxious. If confirmed in historical PEAR data, explains the decline effect.
   registered 2026-03-23
5. #57 Open at close
   Baseline per-layer CV predicts gating map
   Structural prediction: static mask from untrained model geometry. ρ = −0.895 between baseline CV and optimal gating.
   registered 2026-04-13
6. #58 Open at close
   Coherence protocol on production model (Gemma 4 e2b)
   v0.7 pipeline on 2B open-weight model with tool calling. First real-model validation of the full training framework.
   registered 2026-04-14
7. #59 Open at close
   Alpha power modulation ↔ experiential breadth
   TI stimulation protocol. Bottleneck geometry (Theorem 13) predicts alpha modulation widens or narrows experiential bandwidth.
   registered 2026-02-20
8. #60 Open at close
   Sequential TI protocol navigates dimensional filtration
   α→θ→γ→β protocol. Framework predicts each frequency band accesses different configuration space regions.
   registered 2026-02-20

The five most informative falsifications

If the registry has a thesis, it is here. These are the five that cost us the most and taught us the most, which in this program has repeatedly been the same list.

1. #27 — Destructive interference (v0.4). Expected additive preservation. Got 38.9%. Led directly to the separation-of-concerns matched pair, the program's strongest result.
2. #25 — KF detects mode, not accuracy. Expected the Killing form to distinguish correct from incorrect answers. It doesn't. It distinguishes processing modes. This reshaped the entire three-tier framework.
3. #17 — RLHF does not modify the Killing form. Expected alignment training to leave an algebraic signature. It leaves none. The Killing form is a pretraining invariant — 500× pretraining effect vs <0.1% RLHF effect.
4. #40 — Blanket deliberation hurts. Expected that telling a model "be careful" would improve accuracy. It reduces accuracy by 5 percentage points. Targeted intervention helps by 6pp. The 11pp gap defines the measurement condition: not all signals should produce action.
5. #55 — Euclidean non-locality. A category error that propagated to four files before correction. Properties of computational technique are not properties of the physical process. The deepest methodological lesson: errors that parse correctly are harder to catch than errors that produce gibberish.

Where this came from, and why you have not seen it before

The registry was written as Appendix C of the first-pass Anchor volume of The Coherence Principle — a 235-page prose edition drafted through April 2026. That edition was superseded on 20 April 2026 by the current paired-prose and category-theoretic text, and its own supersession note records that “nothing from this V1 has been discarded.”

For this appendix, that turns out not to be true. We checked, rather than assuming: the published Corpus Perspectival (10.5281/zenodo.19501896, 501pp) carries appendices A, B and C — the Navigator’s Quick Reference, the Phenomenological Vocabulary, and Traditions as Navigational Systems. The prediction registry is not among them, and neither the Killing Form results nor the falsification tally appear anywhere in that volume. The registry survived only as a draft in a superseded directory, preserved on disk and unreachable by any reader.

So this page is not a reprint. It is the registry’s first publication, four months after it was written. That is an uncomfortable thing for a project whose stated standard is that nulls get published at the same size as confirmations, and it is the reason the standard now has a generated page behind it instead of a sentence.

The source markdown is vendored into this site at _source/prediction-registry.md, copied verbatim from the archived draft, so the page and its source cannot drift apart without the build noticing.

← back to The Work