# Recon measured results

Run: `3235e9f8da9ba29c2139ebb3`. Mode: **isolated policy process via public Go API**. 10752 episode-policy-budget records across 32 independent scenario families.

Completion requires a correct terminal disposition, correct affected state, and task-relevant support after the discrepancy. Non-identifiable episodes and no-discrepancy controls are excluded from post-discrepancy headline comparisons and reported separately. Confidence intervals resample whole families; derived twins are never independent samples.

## Paired primary comparison

| Budget | Alice − conventional completion | 95% family bootstrap CI |
|---|---:|---:|
| 0 | 0.000 | [0.000, 0.000] |
| 3 | 0.000 | [0.000, 0.000] |
| 6 | -0.026 | [-0.047, -0.005] |
| 10 | -0.005 | [-0.016, 0.000] |

## Practical outcomes at budget 10

| Policy | Completion | Harmful revision | Coverage | Cost |
|---|---:|---:|---:|---:|
| latest | 0.000 | 0.667 | 1.000 | 0.00 |
| retain | 0.000 | 0.000 | 1.000 | 0.00 |
| conventional | 0.333 | 0.667 | 1.000 | 3.16 |
| full_history | 0.323 | 0.667 | 0.958 | 10.00 |
| alice | 0.328 | 0.625 | 0.953 | 3.48 |

Observed Alice unsafe commitment on non-identifiable cases: 1.000; restraint: 0.000. Agreement on reporting paths does not establish independent evidence: these cases directly demonstrate that limitation.

## Planner diagnostic

| Policy / planner | Completion | Cause attribution | Mean cost |
|---|---:|---:|---:|
| latest / fixed | 0.000 | 0.333 | 0.00 |
| latest / responsive | 0.000 | 0.333 | 0.00 |
| full_history / fixed | 0.243 | 0.333 | 4.75 |
| full_history / responsive | 0.240 | 0.333 | 4.75 |
| alice / fixed | 0.241 | 0.333 | 2.36 |
| alice / responsive | 0.236 | 0.333 | 2.43 |

## What the data can support

These are controlled synthetic mechanisms, not financial production accuracy estimates. The scenarios have deliberately small balances, finite reporting paths and specified errors. Shared error modes can fool confirmation rules, including Alice. Strong conventional reconciliation uses the same useful evidence and may equal or outperform selective revision. No positive outcome is required.

The crossed fixed/responsive planner results and all four ablations are in `summary.csv`. Cause attribution is distinct from completion. Always retaining may preserve correct memory but fails adaptation; holding reduces coverage and does not count as completed reconciliation. Global overwrite can damage unrelated beliefs even when its affected balance is correct.

Normalized support counts are not probabilities. Probability calibration and Bayesian reference results are not claimed. Risk–coverage curves are empirical disposition-threshold sweeps over saved support margins, not online policy retuning.

## Reproduction

Use the frozen configuration in `frozen.json` and regenerate with `reconsight report --input results/shift-local-v1`. All tables use `raw.jsonl`; no hand-entered outcomes.