A failed experiment is still part of the evidence.
A localisation run that returns Vienna instead of Padua is still a result.
A run that returns null is still a result.
A run that later turns out to have a bad weight table is still part of the methodological history.
The worst outcome would be to improve the model, overwrite the old run, and publish only the successful version.
The Versioned Receipt is a permanent, re-openable record of exactly what method ran, what evidence it saw, what result it produced and how that result was interpreted at the time.
Later versions may be better.
They do not get to erase their ancestors.
Why an Audit Trail Matters
Reproducible computational research increasingly emphasises provenance: the ability to trace which data, code, parameters and process steps produced an output. Recent reproducibility frameworks use version control, immutable environments, structured documentation and methodological fingerprints such as commit hashes to preserve an audit trail.
The same logic is valuable here even if parts of the Voynich localisation framework are human-coded rather than fully computational.
The question is simple:
Can a future reader reconstruct the exact path from evidence packet to localisation claim?
What the Receipt Must Contain
| Receipt field | Purpose |
|---|---|
| Run ID | Unique identifier for this localisation attempt. |
| Method version | Identifies the frozen rule system. |
| Timestamp | Shows when the run occurred relative to later discoveries. |
| Input packet hash / manifest | Identifies exactly which Voynich observations entered. |
| Candidate regions | Shows which alternatives were allowed to compete. |
| Coordinate target | Prevents confusion between text ancestry, physical production, ownership and other levels. |
| Weights and penalties | Preserves the numerical or ordinal combination rules. |
| Raw candidate scores | Prevents later reporting from hiding close rivals. |
| Assignment / null outcome | Records the actual decision under the frozen threshold. |
| Sensitivity envelope | Shows robustness to plausible weight changes. |
| Ablation matrix | Shows which evidence families drove the result. |
| Expected-evidence failures | Records predictions that did not materialise. |
| Interpretation note | Records what researchers said before later hindsight changed the story. |
The Receipt Begins Before the Final Score
The audit trail should begin with the Preregistration Freeze.
The freeze record says what we intended to do.
The run receipt says what actually happened.
The difference between the two is itself evidence.
If a planned feature was unavailable, record it.
If a rule could not be executed, record it.
If a deviation was necessary, record it before interpreting the regional result.
A Failed Run Is Not Deleted
Suppose version 1 returns:
C5 physical production: null. Padua 0.61, Vienna 0.59, Germany 0.52.
Researchers later discover that one feature was double-counted.
Version 2 corrects the dependence structure and returns:
C5 physical production: Padua assigned. Padua 0.83, Vienna 0.50, Germany 0.47.
Version 2 may be better.
Version 1 remains visible.
Readers need to know that the conclusion changed because the method changed.
Never Rewrite the Old Version in Place
There are two very different statements:
Our validated method identified Padua.
and:
Version 1 returned null. After identifying a double-counting error using known-manuscript validation, version 2 was developed and independently revalidated; version 2 identified Padua.
The second statement has a provenance chain.
It is therefore much more trustworthy.
Versioning Separates Correction From Retrofitting
Science should correct errors.
The Versioned Receipt is not meant to freeze mistakes forever.
It creates the opposite possibility: correction without historical amnesia.
Each new version should say:
- what changed;
- why it changed;
- which previous failure motivated the change;
- which new known-manuscript validation was performed;
- whether the change improved false-positive performance;
- whether fresh holdout data were used.
That makes methodological evolution visible rather than suspicious.
The Receipt Must Include the Bad News
Do not preserve only the winning candidate score.
Preserve:
- near-ties;
- failed expected-evidence predictions;
- false-positive warnings;
- weight cliffs;
- ablation collapses;
- null transitions;
- contradictory coordinate signals;
- unavailable measurements.
A receipt that contains only success is marketing.
A receipt that preserves failure is research infrastructure.
The Receipt Is a Chain of Custody for the Claim
The wider Voynich project now has a separate Custody Through Time lane for the physical manuscript.
The Versioned Receipt is not another custody article.
It applies the same discipline to the research claim.
Who generated the claim?
From which evidence?
Under which method version?
With which failures?
What changed before the next claim was generated?
The physical object has a custody chain. The inference should have one too.
Immutable Does Not Mean Uncorrectable
An immutable receipt does not mean the database can never be updated.
It means old records are appended to, superseded or deprecated rather than silently altered.
A new record can say:
Run VMS-LOC-001 superseded by VMS-LOC-002 because dependency audit revealed correlated image features.
The old result remains reconstructable.
A Methodological Fingerprint Makes the Receipt Stronger
Computational workflows often preserve a commit hash, software version or container identifier that fingerprints the exact method used.
A future Voynich localisation runtime could preserve equivalent identifiers:
- model/rule-set hash;
- evidence-packet hash;
- feature-registry version;
- comparator-corpus version;
- parameter-table version;
- timestamped output hash.
Even a partly manual research system can preserve a strong audit trail if those artefacts are explicit.
The Receipt Must Precede Narrative
The human story should be written after the raw output is sealed.
Otherwise narrative can influence what gets preserved.
The order should be:
run → raw output → receipt → interpretation → article.
Not:
story → adjusted run → selected output → article.
The Versioned Receipt Connects Every Control
- Freeze receipt: what rules existed before the run.
- Sensitivity receipt: how rankings change across allowed weights.
- Ablation receipt: which evidence families are load-bearing.
- Validation receipt: how often the same method misclassified known controls.
- Holdout receipt: how the version performed before Voynich.
The eduKate Versioned-Receipt Protocol
- Assign a unique method-version ID before the run.
- Archive the preregistration and complete method manifest.
- Hash or otherwise fingerprint the input evidence packet.
- Execute the frozen localisation run.
- Save raw candidate scores before interpretation.
- Save null/mixed assignment state and confidence.
- Run the preregistered sensitivity and ablation analyses.
- Record expected-evidence failures and path penalties.
- Create a timestamped immutable receipt.
- Write the narrative interpretation only after the receipt exists.
- If the method changes, issue a new version and link it to the previous receipt.
- Never overwrite a failed run.
What Would Count as a Receipt Failure?
- no record of the exact weights used;
- a previous result disappears after model revision;
- only the winning candidate score is retained;
- the input evidence packet cannot be reconstructed;
- changes between versions are undocumented;
- the narrative predates the sealed raw output;
- a failed result is retrospectively described as if the later version had produced it.
Research Sources
- ENCORE — structured and versioned computational research for transparency and reproducibility.
- Continuous analysis — reproducible workflows and audit trails.
- Jovian — audit trails using software versions, immutable containers and Git hashes.
- FAIRly big — machine-actionable computational provenance records.
- Provenance grounds trust in autonomous science — complete, re-openable records as a basis for audit and correction.
- Voynich and Padua — The Preregistration Freeze.
- Voynich Research Library.
The Final Idea
A failed localisation run should remain capable of embarrassing the next version.
That is healthy.
If the method improves, the history of its mistakes should become clearer—not disappear.