Understanding Your Results

A finished run lands you on a panel with a curve, a parameter table, confidence intervals, an R², a residual view, and a short interpretation. This page walks through each of those, in the order most people read them.

Read the curve first, the table second

The plot tells you whether the model is plausible before any number can. Look for the four things that go wrong most often:

• No upper plateau. The top asymptote is unconstrained because the highest dose did not yet flatten. The fitted Top is a guess; the IC₅₀ may shift if you add a higher dose.
• No lower plateau. Same problem, mirrored: the assay never fully “turned off.” Common with cytotoxicity at low end.
• One outlier dragging the inflection. A single hot replicate near the IC₅₀ moves it more than the same outlier at the ends.
• Inflection outside the dose range. The IC₅₀ is being extrapolated. The number is real, but the precision is overstated.

The parameter table

Each row is a fitted parameter with a value, a 95% confidence interval, and (depending on assay) units. For a 4-parameter logistic, that means:

Confidence intervals tell you what the fit doesn't know

Gradiance reports 95% CIs derived from the covariance matrix of the fitted parameters. The headline rule: a CI that spans more than one order of magnitude means the parameter is not actually identified by your data, even if the point estimate looks reasonable.

• Tight CI (within ±25% of the value): the data fully constrain the parameter. Quote it.
• Wide CI (factor of 2 or more): the parameter is in the right neighborhood, but the curve has more freedom than the data restrict. Add doses on the limiting side, or constrain the parameter (e.g. fix Bottom to 0) if you have prior information.
• Unbounded CI (reported as ±∞ or empty): the curve never plateaued on that side. The parameter is undefined from this dataset alone.

R² and the quality tier

R² is the fraction of variance the model explains. Gradiance auto-derives a quality tier from it:

R² is necessary, not sufficient. A high R² with a sigmoid that bows the wrong way still tells you the model is wrong. The residual plot is what catches that.

Residuals: where the curve and the data disagree

Switch to the residual view to see (observed minus predicted) plotted against concentration. Healthy residuals are scattered evenly above and below zero with no structure. Three patterns worth recognizing:

• S-shape or U-shape. The model is the wrong shape. For dose-response, try 5PL; for standards, try a polynomial or 4PL instead of linear.
• Funnel (variance grows with concentration). Use 1/Y or 1/Y² weighting so high-concentration points stop dominating the fit.
• One spike with the rest tight. A single outlier. Ask the agent to exclude that point and refit, or check the well in the source plate.

The plain-English interpretation

Below the table is a one-paragraph interpretation written by the agent. It restates the headline parameter, calls out anything that warrants review, and names the model and weighting used. This is the text you can paste into a results section — it is grounded in the same numbers the table shows, not generated separately.

QC warnings you may see

• IC₅₀ extrapolated. The fitted IC₅₀ falls outside your tested concentration range. The model is reaching past your data.
• Low dynamic range. Top minus Bottom is a small fraction of the response axis. The signal window is narrow; small noise looks like a real curve.
• Shallow slope. |Hill| below ~0.5. Either the data are unusually shallow or the model is fitting noise.
• Asymptote at bound. A constrained Top or Bottom stuck at the limit you set. Loosen the constraint or accept that the constrained value is being treated as ground truth.

What to do next

For most runs, the answer is “export.” When something looks off, the fastest fix is conversational: hit Refine in the workspace header and tell the agent what you suspect (“the bottom looks too high”, “exclude the 100 µM point”, “try 5PL”) and it refits and regenerates the panel in place. Refines don't consume an additional run credit.

Provenance and the audit bundle

Every fitted run also has an Audit Bundle (JSON) export that captures the full chain of evidence: input file hashes, the agent's recorded steps, the chosen model, the fitted parameters and CIs, and the version of Gradiance that produced them. If a reviewer asks “how did you arrive at this number?”, the audit bundle plus the exported Python script answer it without you having to remember.