Biotech-in-a-Box
An agentic learning loop designed to discover novel data & insights & opportunities in your drug discovery programs
Early drug discovery teams can examine only the few hypotheses they have time to think through, and cloud-hosted language models, which could widen that search, expose confidential questions and charge for every query. caffeinated.bio is a research project building a secure, private and low-cost alternative: an agentic learning loop that continuously explores a scope defined by the team's scientists, such as a compound series, a pathway or an indication. In each cycle, discovery runs combine public and in-house data with established bioinformatics and cheminformatics engines to generate development candidate hypotheses, and caffeination works across those runs to propose hypotheses that no single run produces. The SIM Framework then argues each hypothesis for and against in repeated runs, and provenance tracking with counterfactual re-runs identifies the few inputs on which each answer depends. Agents repeat this probing on their own, within settings the scientists configure and monitor, and what each cycle learns sets the next. The output is novel computed data, insights and opportunities for the team to test in its own laboratory, whose results return to the loop. The system runs small open-source (open-weight) language models on a single desktop with two readily available 32 GB GPUs, instead of the data-center GPUs that large models require, with no per-use fees; data, results and reasoning stay on that machine, and only public-database and literature lookups leave it.
Status, October 29, 2026: the reasoning engine, fred 0.4.1, runs on a MacBook Pro M4 with 48 GB of memory in total, and the first version of the SIM Framework has run once. The demonstration system is a Lenovo ThinkStation PGX. The other components are designed or planned and not built.
Motivation
Early discovery teams pursue the few directions they have time to think through. Candidate hypotheses are brainstormed in meetings, the evidence behind each is assembled by hand, and it is rarely clear which few pieces of evidence a candidate case depends on. Language models can widen that search, but with cloud models a provider sees the questions, which most teams treat as confidential, and the fees grow with every query, which rules out running continuously; running large models in-house instead needs expensive data-center GPUs. Small open-source models on a desktop with two readily available 32 GB GPUs remove these obstacles, at the cost of less capability per model, which the loop has to make up for in structure, repetition and checking.
For a development candidate decision there is also no single answer to compute in advance. The same evidence supports a range of defensible readings. The questions worth computing are how wide that range is, what it depends on, and which experiment would narrow it most. This project treats those as computational questions to be answered continuously, inside a scope the team defines, on hardware the team owns, and ahead of the laboratory work that settles them.
Two ideas are the project's own. Caffeination works across many runs to generate caffeinated hypotheses: hypotheses that no single run proposes, which are distinct from the result most runs agree on. Counterfactual re-runs ask what would change if an input or an assumption were different; they are computed what-ifs, not causal evidence. If they work, they could change how drug discovery science is planned. Neither has been built or tested.
The intended users are scientific leaders and candidate-review committees in early discovery, small biotech teams choosing where to spend laboratory effort, investors reviewing what a program's case rests on, and disease charities following a field. Each can add their own data.
Current practice and the proposed change
The idea behind the loop
Source of the idea: Krauss, A. (2026). The Engine of Scientific Discovery: How New Methods and Tools Spark Major Breakthroughs. Oxford University Press.
Approach: the learning loop
Six stages run continuously, inside the scope the scientists define. The first two generate hypotheses, the next three test and explore them, and the last returns what was learned to the next cycle.
Discover
Batch runs ingest public and in-house data, compute with established engines, reason with fred, rank and document. Only what changed is redone. Designed, not built.
Caffeinate
Algorithms working across all discovery runs generate caffeinated hypotheses, each laid out as a development candidate package. Not yet designed.
Stress-test
SIM argues each hypothesis for and against and adjudicates, repeatedly and in batches. The spread of answers is the result. First version run once.
Probe
Provenance and counterfactual re-runs list the few source records each answer depends on most. Any input can be changed synthetically and re-run. Planned.
Explore
Agents run that probing on their own, within the settings scientists configure, and flag findings, with every attempt recorded. Planned.
Learn
What moved the answers sets the next runs, the order in which claims are re-checked, and the order of experiments. Laboratory results return as data. Planned.
Batch runs
Each discovery run has configuration settings, which fix what it looks at (an indication within the scope, a data date), and behavioral settings, which fix how it reasons (for example, which hypothesis it is biased toward). Each run is one point in the solution space: the range of possible development candidates and plans within the scope. Running many points lets the same question be asked under different settings, at past data dates, or with one input changed. Runs share the desktop's reasoning capacity, so each added run slows the others.
What returns to the next cycle
| Signal | Where it goes |
|---|---|
| influential inputs | The claims an answer depends on are re-checked first, and the data behind them is curated first. |
| decisive changes | A synthetic change that moves an answer points to an experiment. Experiments are listed in order of how much each result would change the answers. |
| agent findings | Become settings for the next discovery or SIM runs, within the bounds scientists configured. |
| SIM spread | Shown beside each hypothesis's rank. Whether it enters the ranking, as a stored input with weights a person sets, is open. |
| laboratory results | Enter as the team's data, replace the scenario values they measured, and score the dated predictions that named them. Negative results are kept. |
The model is pinned within a release, because a result that cannot be reproduced cannot be defended; what the loop accumulates is evidence, estimates, and a record of which inputs moved which answers and which predictions held. A loop that feeds its own output back can also reinforce its own errors, so synthetic values and agent findings never enter the store as evidence, and only laboratory results bring in information from outside.
Components
A fixed scientific workflow, agentic reasoning on a local model, and one desktop, joined into a learning loop.
Established science
Most of the computation is established bioinformatics and cheminformatics run as a fixed workflow: causal genetics, curated activity data, property and safety estimates. Every number comes from a versioned engine with a stored input, or is a stored measurement with its source record.
Agentic reasoning: fred, SIM and exploration agents
fred is a reasoning engine for pre-clinical research that looks evidence up in public databases, checks every identifier in its answers against what it retrieved, and keeps a record of every run; it is used here as a locked vendor service. The SIM Framework runs fred adversarially to stress-test each hypothesis. Planned exploration agents run SIM's probing on their own. None of them writes a number that enters the ranking.
Private, low-cost hardware
The hardware specification is a single desktop with two readily available 32 GB GPUs, far cheaper than the data-center GPUs large models need. It is a hard requirement of the design; the reasoning engine, fred 0.4.1, already runs on a MacBook Pro M4 with 48 GB of memory in total. Small open-source (open-weight) language models run locally, no cloud model is called, and there are no per-use fees, so continuous running costs the hardware and its power. The demonstration system is a Lenovo ThinkStation PGX (NVIDIA GB10, 128 GB unified memory), which has more memory than the target.
Stress-testing each hypothesis
SIM borrows the structure of a development candidate review, the point where a team decides whether to commit a molecule to regulated safety studies. For each question such a review would weigh, a constructive run builds the strongest case that the evidence is sufficient and a skeptical run builds the case that it is not, both from the same records; an arbiter checks the citations and rules, or declines to rule. Each claim is argued several times and the repetitions are not averaged. SIM calls fred without changing it and emits no go/no-go. In this project the review structure is used as a way of finding out what each hypothesis depends on, well before any such decision.
SIM's first version has run once, with results published at sim-ai.dev: one review of NLRP3 inhibition in Alzheimer's disease, ten claims each argued five times, using a synthetic candidate molecule set against real public precedents. Seven of the ten claims did not return the same verdict across their five repetitions. In the loop, SIM reads a caffeinated hypothesis, and the system's labeled scenario values take the place of the candidate readouts written by hand for the first run. How a hypothesis becomes SIM's set of claims is open.
Three planned extensions
Batch runs
SIM is run many times on the same project, varying repetitions, settings and evidence date. The variation is the object of study: the range of answers the evidence supports, since a judgment of this kind has no ground truth to converge on. Agreement across runs measures how reproducibly the evidence is read, not independent corroboration.
Planned.
Influence estimates
From the provenance record and counterfactual re-runs, each output lists the source records it depends on most, possibly five: each is removed or changed, the run repeated, and the shift in the answer measured against run-to-run spread. Scientists can then change any input synthetically and see the output. Synthetic changes are labeled and never become evidence.
Planned. The number listed per output is not fixed.
Exploration agents
LLM agents run that probing on their own: they choose which inputs to change, re-run, and flag findings of scientific or project interest, such as a single scenario value that decides a claim. Scientists configure what the agents may change, how far, with what budget and what counts as worth flagging, then monitor and adjust. Every attempt is recorded, including those that found nothing.
Planned. The agents' model is open.
Scientist control
The loop runs on its own inside the scope the team defines. Scientists configure it, including the settings its agents work within, monitor what it does and adjust the configuration. Agents act within that configuration without asking for each step.
What the loop does on its own
- Ingests sources, detects change and re-runs the stages a change touches.
- Re-ranks, rebuilds dossier sections and re-checks claims.
- Caffeinates across runs and sends new hypotheses to SIM.
- Runs SIM batches and influence re-runs within their budget.
- Lets agents explore synthetic changes and set up the next runs within the configured bounds.
What scientists do
- Define the scope, and how far runs and agents may range within it.
- Configure the agents and set the ranking weights and the reasoning budget.
- Monitor the loop and its agents, and adjust the configuration.
- Read alerts and findings, and correct, accept or withdraw claims.
- Decide what is exported, which experiments are run, and whether to nominate a candidate.
Auditability. Every agent action, synthetic change and re-run is recorded with its settings, including the ones that found nothing. Agents cannot change fred, SIM's adjudication rules, the ranking arithmetic or the weights. Synthetic changes and scenario values never enter the store as evidence, the ranking or a dated prediction.
Outputs
Everything the loop produces is computed; it measures nothing and nominates no candidate. Novel data here means computed results that did not exist before: generated hypotheses, the spread of answers across runs, influence estimates and counterfactuals, each labeled with how it was produced. An opportunity, as used here, is a direction the team was not pursuing that the evidence supports testing: a mechanism, an indication, a chemical series or an experiment. The output's purpose is to say which directions are worth a laboratory's time and why, what each depends on, and which experiment would change the answer most.
- Caffeinated hypothesesGenerated across discovery runs, each laid out as a development candidate package (below) with the runs behind it. Not yet designed.
- OpportunitiesDirections the team was not pursuing that the evidence supports testing, each with its supporting evidence and the experiment that would test it. Not yet designed.
- A stress-test record for each hypothesisFor each question, the case for, the case against, the adjudication, and the spread of answers across SIM runs.
- Influence estimatesThe few inputs each answer depends on most, with how far changing each one moved it. Planned.
- Agent findingsEach with its change, its runs and its evidence, beside a log of everything tried. Planned.
- Counterfactual analysesWhat would change in the ranking if an input or an assumption were different, each to be labeled as hypothetical. Not yet designed.
- Ranking, dossiers and alertsA ranked list of candidate programs, each a target and indication thesis with its chemical starting points, kept with its inputs and history; cited dossiers for those near the top; and alerts when a cited claim stops holding.
- Experiments still neededA costed list for each candidate program, ordered by how much each result would change the answers.
What a caffeinated hypothesis contains
Each hypothesis is laid out like a development candidate package, so that it can be reviewed, and argued in SIM, the way a real one is. Every section states the source of its content.
| Package section | Content | Label |
|---|---|---|
| Target and indication rationale | Genetic evidence and direction of effect, pathway and disease biology, each claim cited. | Public, citedComputed |
| Mechanism and modality | The proposed mechanism and the modalities that could reach the target. | Reasoned, cited |
| Chemical matter | Known compounds from public records; designed structures marked as designed. | PublicComputed |
| Potency and selectivity | Public or in-house measurements where they exist; off-target model estimates; otherwise scenario values. | PublicIn-houseComputedScenario |
| ADME and physicochemical properties | Model estimates with their applicability domain. | Computed |
| Pharmacokinetics and human dose | Exposure estimates, and scenario values for the exposure and dose the hypothesis would need. | ComputedScenario |
| Target engagement and PK/PD | Scenario values: the engagement and exposure-response the hypothesis would need. | Scenario |
| Efficacy in a disease model | Scenario values: the effect the hypothesis would need to show. | Scenario |
| Safety and toxicology | Computed liability flags; in vitro panels and in vivo safety as scenario values. | ComputedScenario |
| Synthesis route | A proposed route from retrosynthesis, not run. | Computed |
| Chemistry, manufacturing and formulation | Not generated; listed only as experiments still needed. | Plan |
| Patents and competition | Public patents, trials and programs, cited. Not a freedom-to-operate opinion. | Public, cited |
| Biomarkers and translation | Proposed biomarkers and a translational argument from cited evidence. | Reasoned, cited |
| Stress-test | SIM's case for and against each question, its spread across runs, and the inputs each answer depends on. | Reasoned, citedComputed |
| Experiments still needed | The experiments that would replace each scenario value with a measurement, ordered by impact on the answers. | Plan |
Scenario values are synthetic. Where a package would hold wet-lab results that do not exist, the system fills the gap with a labeled scenario value, either a requirement (what the hypothesis would need to show) or an assumption (a value plausible under stated assumptions), paired with the experiment that would measure it. Scenario values and synthetic changes never enter the ranking, the backtest, a dossier or the evidence behind a dated prediction, and they are outside the evidence guard. How they are generated is not yet designed.
Evaluation
A reproducible ranking makes the order auditable. It does not make it right.
The main objection to a system of this kind is the long tail it produces: many similar hypotheses with mixed evidence. The response is falsification in two steps. First, a backtest: freeze the data at a past date and test whether the ranking anticipates what later advanced or failed. Every record keeps both the date it was published and the date the system learned it, although database revisions made before the first mirror cannot be seen. Second, and decisive: ranked predictions written down with a date before any experiment, and checked against later data and decisions, down to a small laboratory test such as one compound on one cell line.
Negative results are kept, because falsification needs them and public negative data is thin. The team's own data has no history in the system, so a backtest covers public data only. One estimate is that the workflow contains five to seven predictive calculations, each to be measured against three to five reference metrics fixed before testing; which ones is not decided. For the stress-test stages: agreement across SIM runs is never reported as corroboration; an influence estimate counts only if changing that input moves the answer by more than run-to-run spread and by more than changing a randomly chosen input of the same kind; and agent findings are dated when flagged and later counted, including dismissed ones, against whether they changed a decision or an experiment. These tests are specified; none has been run.
Status
As of October 29, 2026.
- ScopeSet
- DesignWritten
- HardwareIn place
- SIM, first versionRun once
- Engines and layersNot started
- SIM extensionsNot started
- BacktestNot started
- Dated predictionsNot started
A non-commercial project, planned as an iterative cycle of building, testing and verifying over 12 to 18 months. Two parts exist outside the system: the dossier method has run as a pilot at compoundinsights.dev, and the first version of SIM has run once (sim-ai.dev). No engine, database layer or ingest has been built.
| Decision | What it commits to |
|---|---|
| defined scope | Scientists define the scope (a compound series, a pathway or an indication); runs and agents work inside it. |
| companion to the wet lab | The system precedes and accompanies a team's wet-lab workflow and does not replace laboratory work. |
| one desktop | Everything, including SIM and the agents, runs on a single desktop using small open-source language models, within a hard specification of two readily available 32 GB GPUs; no cloud model. The demonstration system is a Lenovo ThinkStation PGX. |
| data inputs | Public databases and literature by default; the team's own data optional, kept on the system and marked private. |
| SIM downstream | Each caffeinated hypothesis is stress-tested by SIM, run in batches, with the spread of answers reported and never averaged. SIM emits no go/no-go. |
| influence estimates | Each output lists the source data it depends on most, from provenance and counterfactual analysis, and inputs can be changed synthetically. |
| exploration agents | Agents run that probing on their own, within a configuration scientists set, monitor and adjust. |
| learning loop | What the loop learns feeds the next cycle; laboratory results return as data. |
| package layout | Each caffeinated hypothesis is laid out as a development candidate package, with every scenario value labeled. |
| no model numbers | No language model writes a number that enters the ranking; scenario values and synthetic changes never enter the ranking, the backtest, a dossier or the evidence behind a dated prediction. |
| reproducible rank | The ranking is arithmetic over stored, versioned inputs, re-runnable from the record alone. |
| graph as query | The knowledge graph is an integration and query layer, not a link-prediction model. |
| co-folding | Protein-ligand co-folding is conditional on a benchmark per target class. |
| acceptance test | A dated prediction, written before any experiment, is what the system is judged on. |
| dossier and guard | Candidate programs near the top of the ranking get a dossier of cited claims (method piloted at compoundinsights.dev), kept under a standing evidence guard. The guard alerts; people decide. |
| standalone or connected | Output can be used on the desktop's own dashboards and tools, or exported to a team's systems. |
| Open question | What decides it |
|---|---|
| scope definition | How a scope is written down (identifiers, pathway membership, indication terms) and how runs and agents are held inside it. |
| opportunities | How an opportunity is recognized and reported, and whether it is ranked like a hypothesis. |
| handoff to the lab | In what form outputs reach the team's wet-lab workflow and its systems, and whether exported records keep the private mark, the scenario label and the guard's alert state. |
| caffeination | Which algorithms work across runs; whether caffeinated hypotheses give insights a single run would not; how counterfactuals are labeled and kept out of the ranking history and dated predictions. |
| SIM input and reuse | How a hypothesis becomes SIM's set of claims (the first run used readouts written by hand), and whether the system reuses SIM's code or rebuilds its method. |
| influence method | Provenance alone or re-runs with records removed or changed; how many repetitions separate an effect from run-to-run spread; how many inputs to list. |
| agents | Which model the agents run on, and which settings the configuration exposes. |
| learning signals | Which returns from SIM and the laboratory enter the ranking as stored inputs, and which stay beside it. |
| reasoning budget | SIM batches, influence re-runs and agents multiply reasoning runs on one PGX, where reasoning throughput is already the binding limit; how many hypotheses can be stress-tested in depth is not known. |
| scenario data | How scenario values are generated and what assumptions each states. |
| in-house data | Accepted formats and their mapping to curated tables; how the guard checks a claim against an in-house record; how lookups are kept free of in-house data. |
| focus and first indication | Whether rare and neurodegenerative diseases stay the focus, given thin public data; the first indication should be an unmet need with a test system, not the area the method finds convenient. |
| benchmarks | The reference metrics for each predictive calculation, fixed before any test. |
| modality analysis | How the system judges which modalities could reach a target; its chemistry engines cover small molecules only. |
| fine-tuning | Whether to fine-tune the model, with what data and for what gain. |
| first pilot | What a first pilot must do, to be set with its users. |
| patent chemistry | Whether structures extracted from patents are in scope, or patents enter as bibliography and claims text only. |
| exposure engine | The credible open PBPK option is a GPLv2 .NET suite whose Arm support is unverified; a property-based ADME estimator may replace it. |
| graph store | The most natural embedded graph engine was archived in late 2025: a maintained fork, a different engine, or Parquet plus an index. |
| dossier scope | The pilot's dossiers cover single compounds; extending the method to a target and indication thesis is not shown, nor whether the pilot's code is reused. |
| guard load | How often claims are re-checked, given the shared reasoning budget, and how many alerts a person can read. |
The identity and normalization engine, with bioactivity curation.
It comes first for three reasons. Every other engine, the knowledge graph and the backtest depend on resolving proteins, genes, compounds, salts, diseases and assays to stable identifiers, which is where errors enter. Its quality can be measured against public data without making any modeling claim. And it is CPU-bound, so it does not compete with the models for GPU time.
The cost is that it produces nothing demonstrable on its own. An early visible result, such as a causal-genetics ranking for one indication, would show more sooner and rest on weaker foundations.
| Item | State |
|---|---|
| scope and design | Set and written, including the layer model, the engine set, the acceptance test and the learning loop at the level described here. |
| hardware | A Lenovo ThinkStation PGX in place. Platform figures are from published specifications, not measured. |
| reasoning engine | fred, developed separately at projectfred.dev, used as a locked vendor service. |
| SIM Framework | First version run once (sim-ai.dev): NLRP3 inhibition in Alzheimer's disease, ten claims argued five times each, a synthetic candidate against real precedents. Not yet joined to the loop. |
| dossier method | Pilot at compoundinsights.dev on three public compounds: data-rich, deliberately data-sparse, and investigational with three PubMed records. Not yet in the system. |
| evidence guard | In the design. Not built. |
| everything else | Caffeination, SIM extensions, agents, the learning loop, in-house data handling, dashboards and export, engines and database layers: in the design, nothing built, no mirror taken. The backtest is specified, not run. |
Design details
The design behind each part.
The four practicesBioinformatics, cheminformatics, database layers and agentic reasoning.
One discovery run, stage by stageThe six stages, and what triggers a re-run.
| Stage | What it does |
|---|---|
| ingest | Public databases and literature are mirrored on a schedule, each copy dated; in-house data is loaded alongside and marked private. |
| detect change | Works out what moved and which candidate programs it touches; only that is reasoned over again. |
| compute | Engines update causal genetics, curated activity data, and property and ADMET estimates. |
| reason | fred assembles the evidence, traces each claim to a source, flags contradictions and writes the reasoning. |
| rank | Arithmetic over stored, versioned inputs, with weights a person sets. |
| document and guard | Candidate programs near the top get a cited dossier, whose claims the evidence guard keeps re-checking. |
What triggers a re-run
A new publication, a trial registration or status change, a new patent family, a deposited structure, a post-marketing safety signal on related chemistry, a new release of a public database, a change to in-house data, a laboratory result, a change to an input that an answer depends on, an agent finding within the configured bounds, a guard alert, or a scheduled re-check (databases are revised quietly). Each re-runs only the stages and SIM questions that depend on what changed; the ranking recomputes cheaply because it is arithmetic.
SIM in more detailThe current method, what its first run carries, and what changes in the loop.
A validation harness checks every answer and citation, retries a weak result within a budget, and keeps a result that still fails, counted and labeled. Because fred may search the literature while it reasons, each citation is recorded as coming from the curated corpus, a recorded search, or neither. A separate calibration pass bands each claim by evidence strength, not by verdict agreement.
The published first run carries its limitations beside its results. Those relevant here: unanimous repetitions show that evidence is read reproducibly, not that it is corroborated, since every repetition shares one model, prompt and evidence base; the convergence-amplifier argument of Scannell and Rogozinska, that repeated model agreement amplifies error rather than correcting it, applies; the arbiter sometimes cited identifiers neither opening had cited; a resolving identifier proves a record exists, not that it supports the claim; and the run is a single case.
In the loop, SIM's input becomes a caffeinated hypothesis with system-generated scenario values, runs are batched, influence estimates and agents are added, and SIM still emits no go/no-go and carries its limitations beside its results.
In-house dataWhat can be added, how it is kept apart, and how it is checked.
The system can run on public data alone. A team can add structure-activity and assay tables, screening results with inactives, compound collections, laboratory results from the experiments the loop proposed, and internal reports in Markdown or JSON.
In-house data is loaded into its own partition and every record is marked private. Every value, graph edge, dossier claim, SIM run and synthetic change that rests on it carries the mark, including predictions from models trained on it. Matches between in-house compounds and public records are made against local copies, so lookups carry none of it; how fred's own lookups are kept free of it is not yet designed. In curation it joins the public structure-activity tables; in ranking it enters as stored, versioned inputs; in a dossier a claim citing it carries its internal record id; a laboratory result replaces the scenario value it measured.
The guard checks public claims against public sources. A claim from in-house data is to be checked against its record on the system, a design addition not yet built, which confirms only that the claim matches the record. Until then such claims are outside the guard's coverage.
Costs: each team's formats need mapping; results resting on in-house data can be repeated only by that team; the data sits under the machine's protections and no more; deleting it ends exact re-runs of rankings that used it, while fred's session files, which nothing prunes, and exported copies keep excerpts; access control is not designed, and data under partner or vendor agreements may not be permitted on the system. Results resting on one team's data are not combined with another's.
The dossier and the evidence guardThe reasons behind a position in the ranking, and a standing check on whether they still hold.
A ranking says which candidate programs come first, not why. The dossier build gives narrative claims with citations (binding mode, target class, pharmacokinetics, regulatory history, off-target mechanisms), written by fred from the primary literature and joined with records from ChEMBL, UniProt, ClinicalTrials.gov, Reactome, Monarch and Open Targets. Each literature claim carries a PubMed ID or DOI; where the literature runs out, a claim reads no data found. The method has run as a pilot at compoundinsights.dev on three public compounds. In the loop it runs continuously on candidate programs near the top of the ranking. Two changes are needed: extending the method from single compounds to a target and indication thesis, which the pilot has not shown, and filling broken pharmacokinetic chains with marked ADME estimates.
Papers are corrected and retracted and database entries revised, so a cited claim can stop holding while the document stays the same; AI-written text adds wrong claims in the same confident style as right ones. The guard reads the candidate-program files, picks out every claim that names an identifier or a value with a unit, and re-checks each against public sources through fred, never against the claim's own file. Claims that influence estimates show an answer depends on are checked first. A claim holds only when its source confirmed it; a change raises an alert only when a second check agrees, because model-based checks vary between runs. In AI-written files a claim with nothing checkable raises an "unsupported" alert. Other alerts are "cannot be traced", "conflicts" and "evidence changed", the last of which also re-opens that dossier section. The guard does not judge the science or decide anything: each alert shows the file and line, the check, every lookup and the record found, and the owner corrects, accepts with a note, or withdraws, with the decision kept in the claim ledger. Curated and computed tables, the graph and the ranking history are outside the guard.
HardwareWhat one desktop provides, and what it costs.
The specification is a desktop with two readily available 32 GB GPUs; the demonstration system is a Lenovo ThinkStation PGX, an NVIDIA GB10 system with 128 GB of unified memory, one accelerator, 1 to 4 TB of storage, on an Arm platform. It provides confidentiality (nothing computed leaves unless exported), a fixed cost with no per-token charge, which favors continuous background work, reproducibility from one machine with pinned versions and dated mirrors, and a build another team can copy.
The costs are real. Everything queues on the GPUs of one machine (one accelerator on the PGX, two cards in the specification), so an explicit scheduler is required. Reasoning throughput is the binding limit: a 27-billion-parameter model on this hardware generates roughly 8 to 10 tokens per second, so a multi-step reasoning run takes minutes, and dossiers and guard checks are restricted to the top of the ranking. One SIM review in its first configuration is about 150 reasoning runs (ten claims, five repetitions, two arguments and an adjudication each, before retries); batches, influence re-runs and agents multiply that, so only a few hypotheses can be stress-tested in depth, and scientists set how the budget is split. Storage caps the mirrors, so the long tail of the largest sources stays only as available as public APIs. A few components must be compiled for Arm.
Throughput is calculated from published memory bandwidth and the SIM run count from its first configuration; nothing has been measured on the machine.
Design choices and their costsWhat each choice costs and what it buys.
| Choice | Cost | Benefit |
|---|---|---|
| public data first | No proprietary data unless the team adds it. | Every public claim traces to a public source; the system runs without private data. |
| in-house data, optional | Format mapping, a guard check not yet built, results only the team can repeat. | Adds structure-activity data and negative results public sources lack, and carries laboratory results back. |
| one desktop | One machine for every job; no burst capacity. | Depth on a few candidate programs; a copy another team can run. |
| no model numbers | Every numeric input must come from somewhere auditable. | A ranking a chemist can audit, instead of a well-written opinion. |
| SIM downstream, in batches | The most reasoning-intensive stage, multiplied by every batch. | Each hypothesis meets its strongest objection, and the range of answers is shown instead of one draw. |
| influence by re-running | Many extra runs per answer, and repetitions to separate effect from spread. | Identifies the few inputs an answer rests on, and so the experiment that would change it most. |
| exploration agents | More reasoning spent, more findings to read, risk of flagging chance results. | More what-ifs than people have time for, all on the record. |
| a return path | A loop can reinforce its own errors, so labels and configured bounds must hold. | Each cycle starts from what the last learned; laboratory results reach the next ranking. |
| ligand-based first | Low-data targets get a weaker potency read. | No predicted affinity that reads like a measurement enters a hypothesis. |
| graph as query | No claim that the system predicts new targets. | Avoids graph link prediction, which is dominated by how well connected an entity already is. |
| dated prediction | Delays the first defensible output. | The difference between a ranked list and an evidenced one. |
| dossiers and guard | Expensive reasoning; false alarms land on a person. | Positions in the ranking that a committee can interrogate claim by claim, kept current. |
SoftwareOpen components where they exist.
RDKit, PyTorch, OpenMM with CUDA, DuckDB, PyArrow, scikit-learn, LightGBM, XGBoost, gemmi and Biopython publish Arm builds; Chemprop, AiZynthFinder and the structure tools are pure Python on top of them. REINVENT4 handles generative design, AiZynthFinder retrosynthesis, fpocket pocket detection, and AutoDock Vina or gnina docking. Two components carry risk: physiologically based pharmacokinetic modeling, where the credible open option is a GPLv2 .NET suite with unverified Arm support, and the graph store, whose most natural embedded engine was archived in late 2025. Both are kept replaceable. The dossier method and the stress-test come from two sibling projects built on fred: Compound Insights (piloted) and the SIM Framework (first version run once). The evidence guard is part of this project's design and is not built.
Licenses and Arm availability are from published package metadata; nothing has been installed or run.
Where the method should work, and where it should notThe intended focus is rare and neurodegenerative disease, the hardest case.
Unmet need is high in rare and neurodegenerative diseases, public data is thin, and brain penetration and translation are poorly predicted by any computational method; SIM's first run, on a brain-penetrant Alzheimer's candidate, was built around exactly that weakness. The method should do better where human genetic evidence is strong, the target class has enough structures and ligand data, open experimental data is rich, and competition is moderate. The intended focus therefore sits where the method is expected to do worst on public data alone. That is deliberate, and the backtest should test it first; in-house data may ease it, which is an expectation, not a result.
It should do badly on Gram-negative antibacterials, where permeability and efflux are poorly predicted, and on polygenic diseases with weak genetics. It can say that a target needs a modality other than a small molecule, but cannot design one. Choosing the indication by what the method finds convenient is the way this goes wrong.
Limitations
Mostly unbuilt
Only fred and SIM's first version exist. Every other stage is designed or planned, and none has run.
Computed, not measured
The system measures nothing and cannot select a drug candidate. Computed values carry their engine version and applicability domain; scenario values show what a hypothesis would need, not what is true, and only their label separates them from a measurement.
A range of answers is not a probability
The spread across SIM runs shows how the evidence can be read under different settings, not how likely the science is. Runs share one model and one evidence base, so agreement measures reproducibility, and repeated agreement can amplify an error.
Influence estimates are costly and partial
Each estimate spends reasoning runs, counts only above run-to-run spread, and identifies inputs that matter, not inputs that are right.
Agents find what they look for
Many attempted changes will produce some that look striking by chance. Every attempt is recorded so a finding can be read against everything tried, and scientists tighten the configuration when this happens.
A loop can reinforce its errors
Synthetic values and agent findings never enter as evidence, and only laboratory results bring in outside information, but the risk remains.
Caffeination is untested
Its method is not designed, and nothing shows that it gives insights a single run would not. A generated hypothesis can read well and rest on little.
Counterfactuals are not causal
They state what the system's own evidence and engines imply under a different input, and can be checked only for consistency when the condition later occurs.
One case so far
SIM's first run is one corpus, one model, one synthetic candidate and one review, and says nothing yet about another program.
Public data is public
Without in-house data, any advantage has to come from joining public data systematically, from caffeination and the stress-test, and from keeping the questions private.
Private is not air-gapped
Public-database and literature lookups leave the machine, and their search terms can show what is being studied. Results, reasoning and in-house data do not leave unless exported.
Output that leaves the system
An exported dossier, ranking or stress-test that rests on in-house data reveals it to whoever receives it, and the guard no longer watches it once exported.
One machine
No burst capacity and no redundancy; a hardware failure stops every run, SIM and the agents included.
Thin data, thin output
Mirrors are tiered to fit the disk, and for thinly studied targets, which are the interesting ones, much of a dossier may read no data found. Gaps are reported, not hidden.
The guard checks claims, not science
It confirms that an identifier or value matches a source. It cannot judge an argument, misses some real identifiers attached to the wrong thing, and raises false alarms, which are counted.