# Visual inquiry: claim-to-evidence ledger

This ledger separates historical and established scientific support, exact results inside specified models, and proposed philosophical extensions. The figures expose assumptions and consequences; they do not turn the philosophy into a theorem.

| ID | Claim and location | Evidence and status | Boundary that matters |
|---|---|---|---|
| V01 | Successive views helped Galileo distinguish Jupiter's satellites from fixed stars; Saturn adds instrumental ambiguity. Notebook observation opening. | Historical support: primary Galileo and Huygens passages, with claim-level dates and source limitations in the [observation ledger](../observation-horizon/evidence-ledger.md). | This model does not reconstruct either telescope, and Saturn's rings are not Jupiter's moons. Recognition latency is not the full publication interval. |
| V02 | A sum view leaves two modeled objects; a difference view resolves them with known wiring. Figure 1. | Exact enumeration: [model.py](../observation-horizon/model.py), [results.json](../observation-horizon/results.json), eight tests. | Two static objects, noiseless measurements and prescribed operators. Excluded candidates are not all conceivable worlds. |
| V03 | Uncertain fixed wiring preserves ambiguity; a known reference resolves it in this model. Figure 1. | Exact enumeration of both wiring states; parity classes cover every modeled source epoch. | The reference and unchanged wiring are assumptions. Adding unknown wiring expands the candidate space; it is not evidence undoing an earlier measurement. |
| V04 | Propagation delay changes record availability separately from measurement content. Figure 2. | Exact bookkeeping of source and receipt epochs in the same model. Historical/physical scope in the [prior note](../when-the-loop-must-wait.md). | Delay is known and constant here. Neither distance nor elapsed time alone determines distinguishability or cognitive recognition time. |
| V05 | Deployment can change the distribution on which a prediction is assessed. Prediction opening. | Established problem: [Perdomo et al. (2020)](https://proceedings.mlr.press/v119/perdomo20a.html), including a threshold alternation example. The factory adaptation is ours. | Their convergence and stability results have assumptions. Neither performativity nor convergence establishes causal identification or optimality in general. |
| V06 | Naive prevention cycles iff b−e < τ ≤ b; cost changes the preferred known-mechanism action. Figure 3 and cost control. | Exact two-state derivation and executable decimal-rational decisions in [performative.py](performative.py); endpoint and tie tests. | Population updates, fixed response, no noise or carryover. Oracle receives both action-specific risks. Its performance is not evidence of a fair learning advantage. |
| V07 | Searching independent null candidates raises the selected reused score and unadjusted false acceptance. Figure 4. | Exact finite binomial sums in [adaptive.py](adaptive.py); independent SciPy review plus small-case enumeration in tests. | All-null independent counts and valid fixed-candidate p-values. No trained system, human experiment or empirical accuracy estimate. |
| V08 | Correction for the searched family and an untouched evaluation protect different stages. Figure 4 and reuse control. | Conventional exact tests and Bonferroni union bound; fresh-count independence. [Dwork et al. (2015)](https://research.ibm.com/publications/the-reusable-holdout-preserving-validity-in-adaptive-data-analysis) anchors the broader adaptive-analysis problem. | This is not their private reusable-holdout method. The fresh rate is per selected winner after every search, not the unconditional pass rate of a two-stage gate. Bonferroni's validity does not require independence, though our exact attained-rate formula does. |
| V09 | Human presence alone does not establish claim-specific corrective evidence. Notebook final argument. | Philosophical inference supported by counterexamples to the stronger proxy: active engagement can coexist with ambiguous or selection-conditioned evidence. | The models contain no experiencer and cannot establish or refute truth-as-recognition, measure transfer entropy, or prove human necessity. |
| V10 | CE could organize inquiry around what a return can change, across delay and mediation. Final proposal. | Research proposal, not established contribution. Bounded comparisons in `CHECKPOINT.md`. | Must add something beyond existing methods and admit unexpected observations that change the question. No distinct CE advantage has been demonstrated. |

Source inspection is scoped: the Perdomo paper and earlier historical passages were inspected for the claims above; the Dwork anchor is the primary author-institution publication abstract. The exact binomial calculations are derived here and are not attributed to that abstract as its algorithm or result.