Reading guide

280 pages, 118 MB, ~330,000 leads. This page says what is worth reading, in what order, and what to trust. The honest summary: the value is in about twenty results and fourteen corrections, not in the volume.

1. Read these six, in this order

#	Page	Why	Length
1	Best 1,000 — overview	The whole programme ranked by kind of evidence	short
2	The ten proofs	The only results not dependent on a measurement	short
3	The integral-identifiability law	Unifies four fields into one counting rule	medium
4	b-value, three years of real data	Real data; a claim corrected and a bug found	medium
5	Ten extended	Where more evidence is obtainable now	medium
6	This guide	Orientation	you are here

2. What the programme found, in four lines

Identifiability is counting, not precision. When measurements are fewer than unknowns, no instrument quality helps. Five fields converged on this independently.
The number is a convention. 5σ is 49% or 100% detection; a tail index is 0.32 or 0.78; an age is 15.2 or 21.3 Myr; a b-value is 0.70 or 2.23 — depending on an unstated choice.
Contrast beats repetition. A second measurement weighted differently is worth more than ten that agree; a third similar marker makes conditioning 56× worse.
The mundane error is the fatal one. Broken symmetry and photon noise are benign; histogram binning, a soft detection roll-off, and a floating-point comparison do the damage.

3. How the folder is organised

Prefix	What it is	Pages	Trust level
`Incoprea_BEST_`	The sorted 1,000	8	ranked
`Incoprea_ORIG_`, `R2_`, `R3_`	Original studies and follow-ups	3	highest — read first
`Incoprea_REAL_`	Real-data analyses	1	real data, superseded by R3
`Incoprea_1K_`, `Mirror_1K_`	2,000 verified claims	18	checked arithmetic
`Incoprea_HYB_`, `HYBRIDS2`	Hybrid vectors and eight deep studies	14	illustrative constants
`Incoprea_Leads_`	Nine agent-written catalogues, 200 leads each, each with a proof or computed tests	10	leads + tests
`Incoprea_P1..P15`	The first day's narrative studies	17	mixed
`Mirror_`	Ten mirrored studies in new domains	107	composition only
`Incoprea_2K_, 25K_, RAND_, FOR_, EXT_, REC_, IMP_`	Composed lead grids, 20,000 each	~100	lowest — questions, not results

4. What to trust, and how much

Evidence	Count	What it means
Proofs	10	true given stated assumptions; reusable
Real-data results	~12	78,493 earthquakes; 602 fireballs; both public
Computed results	~300	numbers at stated parameters; reproducible from code
Verified claims	1,983	arithmetic that held; 1,822 passed
Composed leads	~300,000	questions. Not results.

Do not read the composed grids as findings. They are systematically generated questions; roughly one in a thousand is worth pursuing, and the pages say so on their first line.

5. Every correction the programme made to itself

Fourteen, listed because they are the most trustworthy output — a claim that survived its own test is worth more than one that was never tested.

#	Original claim	What happened
1	Fitted exponents over narrow windows are meaningless	withdrawn — sd 0.015 over half a decade
2	Reversal rates rise with parameter range	withdrawn — 22/23/13% over 1/2/4 decades
3	At most two selectivity crossings	refuted — three occur; parity theorem replaced it
4	Derived-column detector works	failed — 0/200 sensitivity; repaired to 1.00 at 0.10 dex
5	Three channels per side minimum for a double switch	false — 11.7% of two-channel draws
6	Peak ratio set by amplitude ratio	weak — r = 0.267
7	Estimator spread falls as 1/√n	false — flat; systematic not statistical
8	Estimators disagree by 4–14× universally	narrowed — 13% for GR with MLE vs LS
9	Closure dating: half the bias from parameter count	wrong — the error is the term's shape
10	Even harmonics identify a quadratic nonlinearity	fails — contact gives 12× the quadratic's H₂
11	Soft-matter lead: modes inseparable	falsified by its own test (ΔAIC 36)
12	Zero of separation is irreducibly uncertain	wrong — 0.011 nm when the power-law term is fitted
13	Unnormalised condition numbers	inverted a conclusion until columns were normalised
14	Completeness threshold comparison	28% error from `4.4+0.05 ≠ 4.4`

6. What is worth doing next, ranked

Report b per magnitude scale in any seismological use of this folder; the pooled value means nothing.
Prove or refute the crossing bound — small, clean, and it would upgrade a screening rule to a theorem.
Field audit of tail indices — twenty published values, three estimators each.
Apply the derived-column detector to five real catalogues. The tool works; it needs a corpus.
Run the split test on a real recovery stream — ten minutes, decides chemistry versus sample size.

7. What this folder is not

Not peer reviewed. Nothing here has been read by anyone but me.
Not literature-checked. No search was performed at any point; several results are certainly known under other names.
Not experimental. Two public datasets and a great deal of computation.
Not a proof of anything except where it says "proved". The ten items in Tier A are the only results that do not depend on a measurement or a parameter choice.

If one thing is worth your evening: read the ten proofs and the fourteen corrections. The 300,000 leads are questions; those twenty-four items are what survived.

Incoprea Research · reading guide · 3 November 2026 · ~/Desktop/Nov3/ · code and JSON in ~/Desktop/incoprea_nov3/