4
primary models, with the anchored predictive one governing the verdict
10,000
Monte Carlo iterations per product, with a percentile band
17
report sections, none below 300 words
P95
a conservative shelf life and a 95% interval by default, not a loose number
01
What it answers
Two questions. Never just one.
A shelf-life study that delivers a single number, without separating safety from quality, is not defensible before an auditor. The declared shelf life is the minimum of the three axes, always naming which one rules and which agent governs it.
Explicit precedence
The shortest axis wins. In a tie, safety over sensory over quality — for the same number, the axis with the greater regulatory consequence is named. The verdict carries every candidate, not only the winner: the question "and what about the other number?" is answered without recomputing anything on the side.
Two things that are not conflated
Growth potential δ — describes the MATRIX
The process does not move it. It is always computed from the EURL reference. A lethal step does not launder a permissive matrix: at 6 °C, pH 7 and aw 0.99 the pathogen crosses 100 cfu/g again whatever happens.
Time to 100 cfu/g — describes the LOAD
Here the process does rule: it sets the level you start from. A product may still permit growth and yet not reach the limit within its shelf life. The report keeps them separate.
02
How it works
A mechanistic model anchored in the data.
The curve is not characterised by regression over three points: a literature-validated model is anchored and corrected with the observations, bounded to its horizon of validity. That is the answer to "few points, badly spread".
03
Suspicion about the reading
Dead DNA is not corrected. It is declared.
After a listericidal step, DNA from inactivated cells persists for weeks: that is precisely where a molecular count most overestimates the viable load. The rule does not touch the number — that would be manufacturing it: it issues a confidence verdict and says what would confirm it.
The sensory↔microbiology correlation is decided on the residuals after removing the temporal trend from both series: the raw correlation is explained by the shared clock, not by a relation between them. It requires the correct physical sign — more load, worse score — and significance p ≤ 0.05.
Three independent lines of evidence of active metabolism
What is not done
Growth in the series is measured by fitted slope, never by maximum minus first, and with a threshold above measurement noise. A positive correlation is not a "better correlation" for having a larger magnitude: it is physically impossible, and it confirms nothing.
04
Verified on a real study
Eight products, four time points, one conclusion per product.
Ready-to-eat seafood dishes treated by high pressure, sampled at 0, 45, 75 and 90 days at 2 °C. The output unit is the product: eight shelf lives, not one average.
The difference between the crumb with and without preservative — 45.9 against 31.8 days — was not assumed: it came out of the data, bounded to the sampled range. When the product diverges from the reference cardinal value, the observation rules. And where no specific spoilage organism has been identified, the executive summary writes a dash, not an invented pathogen.
05
Where it fits
The panel goes in at the top. The shelf life comes out at the bottom.
OmniOta
Antes
Raw reads from any platform through to taxonomy. It is what turns an aggregated label into a list of named species.
reads → taxonomy
DeLorean
Antes
Relative abundances into reportable absolute units, with copy-number correction and food-matrix calibration.
abundancias → CFUeq/g
FoodSpan
Aquí
Per-product kinetics, three decision axes and the dossier. The output unit is the product, in days from manufacture.
CFUeq/g → shelf life
A DeLorean analysis is a snapshot of a single observation time: one condition, several products. FoodSpan stacks N analyses — one per time point — and builds the product × time matrix. Each product carries its own food matrix, and the storage temperature is locked: an inconsistency between analyses blocks the import.
06
The deliverable
A dossier an auditor cannot break open.
The prose of each section is written as chained connective narrative with verified DOIs. The tables, the figures and the arithmetic of the decision do not go through it: they are deterministic, and numbers are not drafted.
An automated audit reviews the emitted HTML — length, thin sections, captioned figures, verified citations — before it goes out.
07
Technical sheet
In
DeLorean analyses in CFUeq/g — one per observation time — the product's physicochemical characteristics (pH, aw, preservatives), declared treatment and packaging, storage temperature and date of manufacture.
Out
Shelf life per product in days from manufacture, with limiting axis and agent named, a curve with observed anchors and retrospective and prospective zones, every decision candidate, and the full report ready for audit.
Output unit
The product, never the study. A shelf-life study is a multi-product container and a comparator: eight products are eight shelf lives.
Escenarios
What would happen if — a weaker hurdle, thermal abuse, a new preservative, a switch to modified atmosphere — over the same mix of observed and predicted.
Limit · lethality
The lethality of the treatment is not modelled here: it is already built into the finished-product count. Modelling heat penetration, cold spot or D/z is thermal process validation — another authority and another module.
Limit · scope
When the safety axis does not apply, the conclusion rests on the validated process and the operator's batch records. Both are verification of the self-control system and fall outside the scope of a shelf-life study. That limit is stated in the report, not hidden.
Limit · nature of the data
CFUeq/g is a calibrated culture equivalent, not a plate count. In silico does not replace a challenge test: it extends it and contextualises it with the real product.
Bring your panels. We return the dossier.
BIOTECNO.org · Vitoria-Gasteiz
EU + Codex