Dependency upgrades and reference regressions
Retain the old environment and check changed mechanics before trusting a familiar curve.
A version change creates a comparison task
A library upgrade can alter warmup, order semantics, fee treatment, finalisation or numerical representation without changing the user strategy source. Pin the original environment and preserve its output before testing the replacement. The fixture uses Backtesting.py 0.6.6 with recorded numerical dependencies and a downloadable requirements file. It does not automatically adopt the globally newest package. A version string identifies software but cannot prove identical behaviour, so the event ledger, cash and failure-state checks remain necessary. No live production dependency is changed by this teaching example.
Reconstruct exact inputs and bounded outputs
The public map records raw CSV, source and settings identities. Repeated replay under the stated environment must reproduce the same object. Independent Decimal checks reconcile available rolling values, strict crossover intentions, next-open prices, completed holding distance and fees. Future-suffix checks and malformed input fixtures exercise other contracts. An upgraded environment can produce a different identity or result that needs explanation rather than a silent overwrite. A two-decimal display may hide a small numerical residue, so retain the original output and tolerance while distinguishing harmless formatting from a material accounting difference.
Correctness fixes and economic revisions differ
A parsing bug or unintended future transform may invalidate a prior financial result. Changing a selected exit, data universe or hypothesis after viewing a reserved period is a research revision with exposure, not simply a library maintenance pass. Record the actual reason, affected run and effective source/environment. The synthetic fixture has no economic evaluation, so its tests cannot certify a real model after an upgrade. A production task would also need actual cloud execution, route identity and certification gates. A local successful install or test suite cannot replace those receipts.
Keep the negative reference and its rights boundary
The original six-position result remains negative 1.20 under its fixed fees. It is not replaced by a setting chosen to flatter an upgraded engine. The downloadable source and package licence support a transparent teaching reconstruction; no rights are asserted over proprietary model code or a paid market dataset. The native site build and local refresh hooks regenerate byte-identical public files from frozen original pages and the expanded edition. Those publication checks are different from a strategy regression suite. Keep the exact object and limitation attached when explaining what changed, what stayed stable and what still needs independent financial or operational evidence.
Filled invented system case
The fictional upgrade review first preserves the original CSV, source, settings and receipt under engine 0.6.6. A proposed newer runtime must be compared against the same event and cash contracts rather than overwriting the negative reference with a new screenshot. The page claims no tested result for that proposed upgrade. Economic evaluation and production route remain absent. The completed original software checks are reproducible without being relabelled as strategy certification.
| Object | Record | Bounded finding |
|---|---|---|
| Original runtime | Engine 0.6.6; declared dependencies | Frozen reference |
| Inputs | Raw data/source/settings IDs | Retained |
| Upgrade proposal | No tested replacement here | Separate comparison |
| Financial/production gates | Independent evidence required | Not supplied by install |

Primary references and scope
Expanded source review: 13 October 2026. Documentation supports the stated mechanism, not a tested account, hypothetical fill or provider outcome.