CAMS v1.0.1 · Instrument Test Battery · 12 June 2026

Testing the Instrument:
Six-Part Battery for CAMS v1.0.1

Kari Freyr McKern · 12 June 2026 · Corpus: 2,532 society-years · 25 entities · 41 BS-verified CSVs
Instrument version: v1.0.1 (recalibrated from v1.0-Final following formalism audit)
CAMS is a framework for measuring institutional health. Over the week of 9–12 June 2026 the whole instrument was put through a structured test battery: data integrity, cross-run reliability, mathematical formalism, classifier repair, vantage-point bias, and prospective prediction. This report records what was tested, what failed, what was repaired, and what the instrument now honestly claims.
Corpus verified
47/49
Files pass exact formula check. 2 corrected (Poland, Latium).
Bond strength
max |Δ| = 0
Pre-computed vs recomputed Bij: exact match across full corpus.
Classifier anchors
12/12
Historical anchor cases pass after formalism repair.
Cross-run reliability
±1.9 V̄
Year-level drift between independent five-scorer runs. Published bands updated.
Vantage-point bias
~0.5–2 V̄
Proximal end of each series scores harshly; quantified and disclosed.
Predictions frozen
6
USA, Germany, China · 2027–2028 · explicit falsification criteria.

Step 1 Is the data sound?

The project's full archive — 49 datasets covering 25 entities across up to two centuries — was checked file by file. Every number was recomputed from scratch against the canonical v1.0-Final formulas. Forty-seven of forty-nine files passed perfectly.

Two corrections required
Poland and Latium Vetus contained derived Bond Strength figures that no version of the official formula could reproduce — evidently written without running the calculation. Raw scores were intact; the faulty columns were recomputed and replaced. Four societies scored at five-year intervals were removed so the archive speaks one consistent temporal language. All downstream work uses the corrected files.

Separately, four countries scored at five-year intervals rather than yearly were removed from the comparative corpus so that all cross-society analysis rests on consistent annual data.

Step 2 If we measure twice, do we get the same answer?

A fresh, independent five-scorer run of Australia 2020–2026 was compared against the existing series. The two runs broadly agreed on direction but disagreed on magnitude by about twice the published error bars implied.

Reliability finding
Published uncertainty estimates measured agreement within one scoring exercise, not the larger drift between independent exercises. The true inter-run spread at year level reaches ±1.9 V̄; at node level, ±4.9 NV. All later work uses the larger, honest figure. Prediction bands are set at ±2.0 V̄ accordingly.

Step 3 Does the mathematics actually work? — The JUNO formalism audit

The v1.0-Final formalism was tested against all 2,532 society-years. The graph-theoretic core survived. The classifier layer did not.

Error 1 — Fatal: the σi floor clip kills three regime triggers

The published formula uses σi = (A·C/100) · max(K−S, 0.1). Because max(K−S, 0.1) is always positive, σi is also always positive — making negative σi thresholds unreachable:

Dead code
  • Phantom Type II can never fire (requires σmin < −0.7). Observed firings across 2,532 society-years: zero.
  • The σmin ≤ −0.85 arm of Local Node Failure is dead code.
  • The σmin > −0.3 condition in Stable Adaptive is vacuous.

The thresholds were calibrated on the unclipped form (A·C/100)·(K−S); the singularity floor was added afterward without re-deriving them.

Error 2 — Germany 2024 claim fails on the verified corpus

v1.0-Final claimed "Germany 2024 now correctly classifies as Local Node Failure." Observed in the verified germany_cams5 series: V̄ = 7.2, Vmin = 4.4, B̄ = 0.221, unclipped σmin = −0.77 → Strained, not LNF. Vmin misses the 4.0 trigger by 0.4; σmin misses −0.85 by 0.08. The claim rested on a series that is not in the corpus.

Error 3 — USA figures unverifiable: USA not in the corpus at the time

v1.0-Final's "USA 2026: V̄ = 3.5, Vmin = −0.1, B̄ = 0.131 — Confirmed" could not be verified; no USA file existed in the corpus at the time of the claim. The USA five-scorer ensemble has since been migrated (see Step 6 below). The observed figures differ materially from the claimed ones: V̄ = 1.17, Vmin = −0.70, B̄ = 0.120.

Claimed bounds: mostly hold

ClaimObserved (full corpus)Verdict
Bij ∈ [0, 1]max 0.657✓ Holds
λ₂raw ∈ [0.27, 3.42]0.379 – 3.668Top exceeded — 8-society calibration artifact; full corpus widens range
λ₂ discriminates regimesStable μ=2.60 · Strained μ=1.79 · LNF μ=1.09 — monotone✓ Holds on full corpus

Step 4 Repair and recalibrate — CAMS v1.0.1

All three errors are attributable to a single root cause: thresholds calibrated on the unclipped σi, then a floor was added without re-deriving them. The repair is minimal and preserves the original intent.

Fix applied in v1.0.1
Sign-preserving clip: σi = (A·C/100) · sign(K−S) · max(|K−S|, 0.1)

This preserves negative σi where K < S (the diagnostically meaningful cases) while avoiding the zero singularity. Corpus σmin now spans −2.76 to +1.68, and all three previously dead triggers fire at sensible rates: 38.2% (σ < −0.3), 13.7% (σ < −0.7), 8.0% (σ < −0.85).

The formerly impossible Phantom Type II category came alive: 41 historical cases, all historically exact — Weimar Germany 1931, late Qing dynasty, partition-era Poland, post-Soviet Ukraine. The category was never wrong; it was unreachable.

Classifier precedence was specified explicitly: Freeze → Systemic Crisis → Phantom Type II → Local Node Failure → Stable Adaptive → Strained. The repaired classifier was validated against twelve historical anchor cases — all twelve passed.

Node Value verification: computed vs pre-computed across full corpus

Fig 1. Node Value verification — computed Vi vs pre-computed values across the full 2,532 society-year corpus. Points lie on the 1:1 line; max |Δ| = 0.00.

JUNO Bond Strength verification: max |Δ| = 0.00e+00

Fig 2. JUNO Bond Strength verification — pre-computed vs recomputed Bij. Max absolute deviation = 0.00e+00: the canonical bond-strength formula is verified exact across the full corpus.

Phase space trajectories under v1.0.1 classifier

Fig 3. Phase-space trajectories (V̄, B̄, λ₂) under the v1.0.1 classifier. Stable Adaptive societies cluster top-right; Systemic Crisis cases cluster bottom-left. Strained societies form a diffuse mid-band.

Regime timeline across the corpus 1800–2026

Fig 4. Regime timeline, selected societies 1800–2026, under the repaired v1.0.1 classifier. Phantom Type II cases (hollow institutions) appear throughout the historical record once the dead code is removed.

Spot checks

CaseVminRegimeSensible?
Hong Kong 1943−3.70.060LNF + Systemic Crisis + Freeze✓ Occupation collapse
Ukraine 20225.80.60.180Local Node Failure
China 202513.510.40.372Stable AdaptivePer corpus
Russia 20268.87.10.250StrainedPer corpus
Australia 20268.96.90.276Strained
Germany 20247.24.40.221StrainedContradicts v1.0-Final claim
Ukraine 2014–2026 detail: node trajectories and regime transitions

Fig 5. Ukraine 2014–2026 detail. Local Node Failure in 2022 is driven by Vmin = 0.6 (Shield absorbing full wartime stress while other nodes degrade). The trajectory is coherent with the corpus anchor cases.

Step 5 Testing the measurer, not just the measured — vantage-point bias control

The deepest worry with historical scoring is hindsight. Do recent years look worse simply because their troubles are vivid, while older crises have been smoothed by time? To test this, the years 2000–2010 in America were scored twice: once from today's vantage (a 2026-scorer) and once as if by an analyst working in early 2012.

Vantage-point bias: confirmed, directional, quantified
Each vantage scores its own recent past most harshly by ~0.5–2.0 V̄ points. The ΔV̄ gradient across the 2000–2010 window is −0.18/yr (r = −0.83, p = 0.002). The same decade's decline looks 60% steeper viewed from 2012 than from 2026.
YearV̄ (2012 vantage)V̄ (2026 vantage)ΔV̄
200012.2810.29+1.99
20018.397.34+1.05
20048.657.91+0.74
20082.422.68−0.25
20105.205.78−0.58

The sign pattern matches the pre-stated hypothesis exactly: each vantage scores its own recent past more harshly and the distant past more benignly. The blinding was simulated (the scorer knew post-2012 history), which means the measured divergence is a lower bound on the true vantage effect.

Practical implications
  • Long-run trajectories can be trusted in shape.
  • The final 3–5 years of any series carry vantage-proximal stress inflation of ~0.5–1.5 V̄.
  • Cross-society comparisons of current state are safer than within-series recent trends (shared bias partially cancels).
  • Regime assignments within ±2 V̄ of a threshold in the most recent years should not be cited without this caveat.
  • The ±2.0 V̄ prediction bands are vindicated: vantage tilt at the proximal end sits inside the pre-registered band.

Step 6 Predictions on the record — pre-registered 2027–2028

A diagnostic tool earns trust by predicting, not just explaining. Six predictions for 2027 and 2028 were frozen on 12 June 2026, with explicit falsification criteria. No post-hoc adjustment is permitted.

Method: per-node, per-dimension OLS trend over the last five scored years, clamped to [0,10]; FG and regime computed by v1.0.1. Bands set at ±2.0 V̄ to reflect honest inter-run reliability.

SocietyYearVminλ₂σminPredicted regimeResult
USA20270.52−1.770.1110.78−0.79Systemic Crisis
USA2028−0.40−2.540.0990.70−0.77Systemic Crisis
Germany20278.555.860.2571.75−0.41Strained
Germany20288.705.620.2621.75−0.42Strained
China202713.7810.100.3802.65+0.31Stable Adaptive
China202813.8810.030.3832.65+0.24Stable Adaptive
Falsification criteria (frozen)
A prediction hits if: (1) the v1.0.1 regime matches, AND (2) observed V̄ falls within ±2.0 of predicted V̄.

A prediction fails informatively if regime misses in the direction of improvement for USA or deterioration for China — the outcomes recency-salience bias predicts.

Instrument falsified (not just missed forecast): if two or more of the six predictions miss regime by two or more cascade levels, the classifier thresholds do not generalise prospectively and v1.0.1 must be withdrawn rather than patched.

What it all adds up to

The archive is now verified end to end. The mathematics genuinely works under v1.0.1. The diagnostic categories have earned their places against real history. But the week's most important product is humility made precise.

The instrument measures two things at once: the state of a society, and the standpoint of whoever is doing the measuring. That second component is real, runs in a consistent direction, and is now quantified — roughly 0.5–2.0 V̄ of exaggeration in the most recent few years of any series.

Long-run trajectories can be trusted in shape. Dramatic claims about the most recent three years of any country deserve a discount that is now known rather than guessed. The framework no longer merely asserts that it is honest about uncertainty. It has receipts.

Current instrument status
CAMS v1.0.1 · Graph-theoretic core verified on 2,532 society-years · Classifier repaired and validated against 12 historical anchors · Cross-run reliability quantified (±1.9 V̄) · Vantage-point bias quantified (~0.5–2.0 V̄ proximal) · Six prospective predictions frozen for 2027–2028 · Thresholds recalibrated on full corpus (not original 8-society set)

Source files: CAMS_Validation_Report.md · JUNO_Formalism_Test_Report_2026-06-12.md · Blind_Scoring_Control_Report_2026-06-12.md · JUNO_Prospective_Predictions_2027_2028.md · CAMS_Testing_Summary_Plain_Language.md · JUNO_v1.0.1_Recalibration_Patch.md — all in HariSeldon/cam5 working directory. Datasets: JUNO_FG_corpus_v101.csv · JUNO_predictions_2027_2028.csv · USA_blind2012_vs_2026_comparison.csv.