September at Inflexa: 14 releases
Reports that trace every claim to the raw data, packages without rebuilds, protein lookups, sturdier long analyses, and control over what each agent costs. Here is what changed and why.
Six changes stand out. Here is each one, and why we made it.
Every claim in a report traces back to the raw data.
“Where did this number come from?” is the first question anyone asks of an AI-written analysis. In an Inflexa report you can now click a claim and follow the chain back to the raw data it rests on: the figure, the step that drew it, and the input it read.
The trace works offline. Every hop in the chain is content-addressed, so each link names the exact bytes it points to.
How provenance works1,306 genes are differentially expressed in NASH vs healthy liver.
- #9c1e4f0a…figurevolcano_NASH_vs_healthy.png
- #51b7d2e9…stepde_analysis.R (DESeq2)
- #e03a6c18…raw dataGEOGSE126848_Gene_counts_raw.txt
Each hop is addressed by the hash of its content, so a file that changed after the fact no longer matches.
Add an R or Python package in seconds.
The sandbox image used to bake in every library. That made it many gigabytes, and adding one package meant rebuilding the whole image.
Packages now live in a local store on your machine, outside the image. inflexa store add <package> installs one more from PyPI, CRAN, or Bioconductor for your analysis. Every other analysis keeps the package set it was working with.
One new package meant rebuilding the whole image.
inflexa store add <package>The agent can look up proteins.
New UniProt and AlphaFold tools return a protein's curated function, its location in the cell, and its predicted structure. In the web UI, the AlphaFold model appears as a 3D structure card, colored by the model's confidence.
- protein
- KRAS · P01116
- function
- GTPase that relays growth signals
- location
- Cell membrane, cytosol
Long analyses keep their work.
Each chat round is now saved as soon as it completes. When a turn fails partway through, the rounds before it are already stored, and the session carries on.
When a conversation grows past what the model can hold, older turns are summarized. Nothing leaves the context silently.
- Round 1saved
- Round 2saved
- Round 3saved
- Round 4failed
The session survives, with rounds 1 to 3 of the turn intact.
A throttled database no longer reads as “no data”.
When a public database rate-limited a lookup, the agent used to receive something that looked like an empty result.
It is now an error. “Nothing was found” and “we could not check” mean different things in a scientific conclusion, and the agent now knows which one it has.
Nothing is known about this target.
The lookup did not run. Try again before concluding anything.
Newest models, with control over cost.
Inflexa works with the newest Claude models.
The model picker now sets the reasoning effort of each agent too. Deeper effort spends more tokens on hard problems; lighter effort is faster and cheaper for routine work. By default the conversation agent runs at high, and the sandbox and utility agents at medium.
OpenTelemetry traces for each agent turn.
Every agent turn now has its own OpenTelemetry trace, with spans for model calls, tool calls, and sandbox runs, so you can see where a turn spent its time. The spans carry no prompt text.
Get the latest release
Already installed? Run inflexa upgrade to bring your install up to date.