The words we use, explained

The portfolio selector: a senior-strategist pass that ranks the leaf experiments by how decisively they advance the field.

Non-triviality is its highest-priority criterion. A wildcard round selects experiments that maximally complement the guaranteed set, and the final assessment re-ranks under the strictest standards with no pillar allowed more than forty percent of the slots. The report prints the result as the review outcome.

A selectable worldview injected into the first steps, such as reliability engineering or information theory and error correction.

The lens frames how the goal is decomposed. Options include distributed consensus architecture, synchronization and phase-transition architecture, information theory and error correction, complex adaptive systems, reliability engineering and cybernetic control systems. It is distinct from the three lens origins used when hypotheses are generated.

The harshest peer reviewer in the pipeline: it rewrites, merges and rejects experiment specifications and scores each 1 to 10.

It sits between step 9 and the ranking. It rewrites titles, removes hedging, rejects mediocre experiments, merges permutations and assigns a genius score on a scale where 1 to 3 is forbidden, 8 to 9 is brilliant and 10 is paradigm-shifting; the target average across all experiments is at least 7.5. It also enforces that a null result must be informative.

The operator's review vocabulary for any node in the workbench, with a stated downstream meaning per level.

The categories are excellent, needs improvement, incorrect, too vague, too ambitious, not novel and other. For each level the workbench explains what the category means and what it means for the nodes below.

The stable key of a node inside its run, such as IH_Q_L3_M_G1_1_01, on which every derived artifact is anchored.

Keys are positional: they say where a node sits in the tree and nothing about its content, and they repeat across runs. That is why a hypothesis is never shown by its key; it gets a name on arrival. Derived content (clear versions, explanations, glossaries, posters) is keyed on the run and node key rather than on a row id, because keying on row ids is what once detached five posters from their experiments.

The engine that takes one goal about a longer life and descends, step by step, to the experiments that would decide what is unknown.

The Omega Point starts from a vague objective and turns it into one precise master question. It breaks that question into the end-states that would have to hold (goal pillars), each pillar into atomic testable requirements, then maps what established science already delivers across a dozen research domains. Where requirements and science do not meet, it writes gap questions; for every gap it writes several rival hypotheses; for every set of rivals it writes the questions that would tell them apart; and it ends in bench-ready experiment specifications, ranked.

The site presents the result as a contest, not a conclusion: every rival stays in the ring with the one observation that would knock it out, and nothing has been tested. The engine itself is a vendored sub-application; the pages on this site, the explanation layer, the audits and the posters are additive layers that read its output and never feed anything back into it.

A Studio-only fan-in agent that turns a finished tree into one prioritized research programme with tiers Now, Next and Later.

Each recommended action carries a score, a tier, a rationale, its dependencies and explicit coverage of the tree. It exists only in the Agent Studio version of the pipeline.

One execution of the engine on one goal; the unit everything else is keyed on.

A run has a key, a goal, a stored result and a status that names its stage: executing, completed, failed at execute, failed at persist, or interrupted at step. Interrupted is not a failure; the process running it died and the finished steps are still on disk, so the run resumes. A persisted run's tree plus its ranked portfolio is a programme, also called a decomposition on the dashboard; a decomposition as a whole is never public, only its experiments and hypotheses are. A run can publish hypotheses without ever persisting a programme.

The deterministic, full-length report of a completed run, assembled in code from the stored result, never by a model.

The report walks the steps under reader-facing titles and never truncates. It downloads as Markdown, text, Word or PDF in English, and in Russian through a localizer agent whose progress and integrity are tracked on the run row.

The ten kinds of node in a decomposition tree, from the master question down to the executable experiment.

Q0 is the master question. M_G is a goal pillar. RA is a requirement atom. DOM is a domain finding. S is a seed question. L3 is a gap question. IH is a hypothesis. L4 is a discriminating question. L5 is a mechanistic sub-question. L6 is an idea, a proposed investigation, the executable test with its system, intervention, meter and threshold. The engine names nodes by these codes; the site renders each with its reader-facing name and keeps the code as a chip.

Nine named agent steps, each with a role: from goal formalization to execution drilldown, then a critic and a ranking.

Step 1, the Initiator (goal formalization), turns the objective into the master question. Step 2, the Immortalist Architect (goal pillars synthesis), decomposes it into goal pillars and writes the bridge lexicon. Step 3, the Requirements Engineer (requirement atomization), turns each pillar into requirement atoms. Step 4a, the Domain Mapper, names eight to twelve research domains; step 4b, the Domain Specialist, fans out over each domain to collect fifteen to twenty-five established scientific pillars. Step 4 takes about three quarters of a run's wall clock.

Step 5, the Judge, is disabled; its strategic matching of science to requirements was folded into step 4b. Step 6, the Strategic Science Officer (frontier question generation), writes the gap questions. Step 7, the Instantiation Gatekeeper, writes maximally divergent hypotheses for each gap. Step 8, the Lead Investigative Officer (tactical decomposition), writes the discriminating questions. Step 9, the Lead Tactical Engineer (execution drilldown), writes the leaf experiments with their system, intervention, meter and time. After step 9 the Genius Experiment Critic improves and rejects, and the best-experiments selector ranks a portfolio. Step 10, the Convergence Critic, is disabled.

The run report renumbers for the reader: the engine's steps 6 to 9 appear as 5 to 8, followed by "Review outcome" and "Run summary". Both numberings are in use, so a step number alone is ambiguous without saying which scheme is meant.

Five named experiment shapes the engine bans outright.

A single knockdown with a readout, omics fishing, computational filler, an aged-versus-young characterization, and permutation padding. The critic merges permutations and rejects the rest.

Two lists written in step 2, failure channels and system property variables, that every later stage tags against.

The lexicon is what lets goals be mapped to scientific reality: requirements, scientific pillars and experiments all refer back to the same named failure channels and the same measurable properties. The run report prints it as "Shared vocabulary: failure channels and system properties".

A hypothesis built by borrowing a law or equation from a field far from biology and mapping its variables onto the system.

Each set of rivals must include one. The record names the source field, the specific law and the variable mapping. The mutation directive names which of fifteen far fields the import must come from, such as non-equilibrium thermodynamics, control theory, fracture mechanics, percolation, queueing theory, phase transitions or distributed-systems consensus.

An unbiased element built into at least one experiment per branch that asks what else might be happening that nobody predicted.

It sits alongside hypothesis testing so a run does not only confirm or refute what it already imagined.

The specific, measurable prediction that separates one hypothesis from its rivals.

It is the hypothesis's stake in the contest. The hypothesis audit checks whether it actually discriminates: a prediction every rival also makes is not a test, and a prediction that names no measurable outcome is not one either.

A question written to eliminate hypotheses: at least half of the tactical questions must pit two or more rivals against each other.

Step 8 works by elimination over description, in three phases: the kill shot (which single experiment eliminates the most hypotheses at once), cheap kills, and exploration. Node types are discriminator question, model requirement, tool requirement, unknown exploration and validation question. Five cognitive modes drive generation: reductionist, constructivist, lateralist, comparative and temporal.

The five mechanism families a set of rivals must spread across, plus a sixth reserved for identification gaps.

Interface integrity, information control and sensing, structural and topological, resource and energetic, systemic and environmental. No two hypotheses for one gap may share a category. Candidate-set selection is the sixth, used only when the gap asks which things or how much, where the rivals are competing candidate sets and the no-duplicate rule is waived. The site renders each category as words a reader can meet.

The kinds of link between a requirement and a scientific pillar: solves, partially solves, proxies for, violates, or enables measurement for.

A "violates" link is the redline: the science improves one requirement but its failure modes trigger a catastrophe the goal defines, and such links are removed. A partial link is sub-classified by the kind of delta: magnitude, execution, timescale or knowledge.

The executable test at the bottom of the tree: system, intervention, meter and time, with what a positive and a null result would mean.

Every field must read like a lab notebook entry, specific enough that a technician could run it tomorrow without asking a question. An experiment carries expected impact, an informative null outcome, a link to the system property it moves, a feasibility score calibrated from standard assay to technology that does not exist yet, and whether it includes a discovery component. It must be pipeline-dependent: not conceivable by someone who only read a review article, because it requires the specific chain that produced it.

Published experiments are the site's "ideas": each has a public page with the original wording, a clear version, the logic back to the goal, the rivals and the payoff either way.

A failure pattern specific to the target system, six to ten of them per run, named in the bridge lexicon.

Requirement atoms then describe each failure with a controlled failure shape such as slow drift to cliff, runaway feedback, compensatory toxicity, hysteresis lock, synchronization loss or unmodeled. At least one atom must target failure channels not captured by current meters.

An honest 1 to 10 calibration of how doable an experiment is, from a standard assay with commercial reagents to technology that does not exist.

Nine to ten means a standard assay and one to three months; one to two means the required technology does not exist yet. Ranking weighs feasibility against ambition rather than optimizing either alone.

The question-writing strategy matched to each gap type.

Genesis probe for a void, contextual decoupling for a fragility trap, causal pivot for a proxy mirage, arbitration logic for a cluster clash, and adversarial falsification, which is mandatory and designs a question to disprove the most promising scientific pillar.

What kind of unknown a gap is: void, fragile, proxy, clash, or adversarial.

Void: nothing is known here, the question has not been asked of this system. Fragile: something is claimed, but on evidence too thin to carry weight. Proxy: what is measured stands in for what matters and may not track it. Clash: two established results predict opposite outcomes and both cannot be right. Adversarial: a question written to disprove the most promising scientific pillar. A goal-level cluster status says which kind dominates: void, fragility trap, proxy mirage or cluster clash.

One required end-state that, together with the others, would make the master question plausible.

Pillars are architecture-neutral and as mutually exclusive and jointly exhaustive as the engine can make them. They are found by inverse failure analysis: list the catastrophic failures, cluster them, and invert each cluster into the healthy steady state that must hold. A pillar title must name a mechanism of failure, not an anatomical subsystem; a title that could be a textbook chapter is rejected. At least one pillar must be cross-cutting, addressing a systemic failure not obvious from anatomy. Each pillar carries a role: failure mechanism, or the one reserved identification pillar.

Note the two uses of "pillar": a goal pillar is a required end-state on the goal side; a scientific pillar is a piece of established science on the reality side.

The four tests a heretical hypothesis must pass: it revises a subfield, is not mainstream, has an empirical anchor, and predicts a testable surprise.

Every set of rivals must contain a heretic that contradicts the mainstream account. On the page the section "The standing" says why the hypothesis is not the mainstream account, using this checklist.

One of several rival explanations for a gap, each proposing a fundamentally different mechanism, with the prediction that would tell it apart.

Step 7 writes four to seven hypotheses per gap, maximally divergent: each must use a different domain category, the set must contain a heretic that contradicts the mainstream story, a cross-domain transfer from an unrelated field, and scouts including the possibility that the phenomenon does not exist. A hypothesis carries its mechanism, a discriminating prediction, its lens origin, its domain category and, for a cross-domain one, the source field, the law borrowed and the variable mapping.

On the site a hypothesis is a publishable object of its own, named on arrival, keyed on its run, and read in eight sections: the unknown, the claim, the test, the contest, the import, the bench, the standing and the provenance.

The rule that a question asking which things and how much must be answered with a set and a number at every stage, not with a mechanism.

It was added after a run about tissue replacement produced a hundred hypotheses that named no tissue: step 2 had recorded the answer type only in prose, and no later stage knew the pillar was different. Now the obligation travels as a typed role from pillar to atom to gap question to hypothesis, a runtime prompt block restates it for the identification pillar only, and a checker tracks the two halves, enumeration and quantity, so a retry can be told which half is missing.

What is learned if the experiment shows no significant result; it must be informative, never "inconclusive".

A required design property: both outcomes of the experiment must teach something. The experiment page has a section "Both outcomes are informative" built from this field, and the critic rejects specifications that fail it.

The method step 2 uses: enumerate catastrophic failures, cluster them, and invert each cluster into a required healthy state.

Failures are first classified as within or between subsystems, then clustered into pillars. The naming test then rejects any pillar whose title describes a subsystem rather than the mechanism by which it fails.

A genuine unknown where what the requirements need and what established science delivers do not meet, stated as a question that a literature search cannot answer.

Step 6 writes the gap questions. Each must identify an epistemic gap that would change how the problem is thought about, and each is typed by what kind of gap it is. On the public site the gap is the unknown a hypothesis exists to explain; the reader meets it first as a rewritten heading stating its core, with the engine's original sentence folded underneath.

A gap question can also turn out to be already answered in the literature. The explanation layer says so plainly, because a pipeline proposing a closed gap is a valuable finding, not a failure.

Which of three generating questions produced a hypothesis: the medium the bad state is stored in, how long-lived species avoid it, or what protocol broadcasts the error.

Substrate lens: in what physical medium is the maladaptive state stored? Evolution or comparative lens: how do long-lived species prevent this? Communication lens: what protocol broadcasts the error systemically? It is recorded on the hypothesis and is distinct from the run-level epistemic lens.

The single precise question the whole run answers, written from the operator's goal by step 1.

It must be solution-neutral, explicit about the system it concerns, anchored to a baseline, driven by success criteria, decomposable, and explicit about mission duration. Everything below it is a narrowing of it, and the question-fidelity audit later measures whether the descent still answers what the operator typed rather than what the engine rewrote.

The bottleneck under a discriminating question, classed by what blocks it: cannot measure, cannot isolate, mechanism unclear, or assumptions untested.

Cannot measure it becomes a tool requirement; cannot isolate it becomes a model requirement; mechanism unclear becomes a mechanism drill; assumptions untested becomes a validation drill. Each branch ends in leaf experiments.

A per-gap instruction that forbids two domain categories and names the single far field the cross-domain import must come from.

It exists because, across 175 gaps, the five categories had each been used exactly once per gap every time, so the catalogue held five recurring shapes rather than hundreds of ideas. The directive is deterministic in the gap id and the master question, so re-runs get the same constraint.

The three-part statement of a gap: what the requirements need, what science actually delivers, and the specific difference.

The delta names its kind: magnitude, execution, timescale, knowledge or contradiction. At goal level the engine also writes a delta summary (the gap between the required regime and current capability) and an epistemic block (which assumption is obstructing progress).

The engine's own test of worth: an experiment that could be conceived from a review article alone fails it.

An experiment must require knowing the specific chain from goal to pillar to requirement to science to gap to hypothesis that produced it. Non-triviality is the highest-priority criterion of the best-experiments selector. The engine has no separate "novelty" field; the nearest words are this test and the operator's "not novel" feedback category.

A typed field that says what kind of answer a pillar, atom or gap question owes: a failure mechanism, or an identification of which things and how much.

Most pillars are failure-mechanism pillars. When the master question asks which things or how much, exactly one pillar carries the identification role: its end-state is that the set the question asks for is identified and its size established, in the question's own terms. The role is carried as data down the chain, because later stages cannot recover the obligation from prose. An identification gap question must include both a which question and a how much question; both halves or the obligation is not carried.

One solution-agnostic, testable requirement under a goal pillar, with done criteria and a failure mode.

Each atom names a state variable, a failure shape, the perturbation classes that would expose it, a timescale (acute, subacute, chronic or decadal), the classes of meter that could measure it, and a multiple-realizability check requiring at least three distinct architectures that could satisfy it. Atoms carry the same role split as pillars: an identification atom guards against the answer being wrong (the set miscounted, a rival set never compared).

One of eight to twelve research areas the engine scans for established science relevant to the goal.

Domains are meant to be mutually exclusive and each to contain about twenty-five actionable interventions. Each is tagged core mechanistic, adjacent non-obvious, or measurement technology. Step 4b fans out over the domains, one call each; when the proxy pool has less capacity than domains, a domain can be silently dropped, which is why the domain count of a run is checked.

Another hypothesis for the same gap; every pair of rivals is bound to one observable that would come out differently under each.

The competition matrix of a gap lists, for each pair of hypotheses, the one measurable outcome where they predict differently. A hypothesis that predicts what its rivals predict has nothing to contest and earns no experiment. On the page a rival arrives from the engine named only by a machine label; the card prints a plain axis name with one line of gloss and keeps the raw label folded underneath.

A piece of established, evidence-based science collected in step 4b: what is actually known, never a speculative intervention.

Each carries a mechanism, a verified effect, a readiness level (RL-1 lab-validated, RL-2 human data or models, RL-3 clinically deployed), capabilities, a fragility score from 1 to 10, and its relationship to the goal. Scientific pillars are what the gap questions compare against requirements; the pillar provenance audit later checks whether the studies they cite can be found.

The four parameters that fully specify an experiment: what it runs in, what is done to it, what is measured and how, and over what time.

The clear version renders them as six blocks: what it runs in, what is done to it, what is measured and how, what the numbers have to show, what a positive result would change, and what a null result would mean. A meter is the measurement; requirement atoms speak of meter classes and whether a meter exists, partially exists or is missing as of 2026.

A measurable property of the target system, with a unit and a measurement approach, eight to twelve per run.

Experiments link back to the property they move. Requirement atoms use controlled state-variable names on the same idea, such as functional reserve or an unknown failure channel.

The requirement atom every pillar must contain that targets failures current meters cannot see.

It is marked with the unknown failure channel state variable or the unmodeled failure shape. Its purpose is to keep the decomposition honest about what is not observable yet.

What an answer would have to look like to answer what the operator typed, derived from the raw goal and never from the engine's rewrite.

It lists the answer shapes owed, the subject, the quantity unit, the ambition verbs and what would count as an answer. Each becomes an obligation checked at every checkpoint. Ambition verbs exist because a rewrite that turns "reverse" into "maintain" has changed the question, and nothing else in the system would notice.

The kind of thing an answer is: an enumeration, a quantity, a mechanism, a protocol, a forecast, a comparison or a feasibility.

The engine produces mechanisms and, after step 8, something close to protocols. It has no stage whose output is an enumeration or a quantity. Saying that in one word is the point of the list.

The walk from the master question down to one experiment, retold for someone outside the field, with every step labelled by what holds it up.

Its parts: an opening that names the one unexpected move and says plainly that this is a proposal the pipeline generated rather than a result anybody measured; the proposed mechanism written link by link; a picture for it, with the required statement of where the picture breaks; the steps, each with what it says and what it rests on; a summary separating what the individual links have behind them from what the sequence as a whole has; how a result from the test could mislead, with what closes that trap; what a result would change back at the master question and what would still be unestablished; and the observation that would make the whole chain wrong.

Certain openers are banned outright, such as "This question asks whether" and "This step proposes". A cited source arrives with its limit attached, because a finding without its limit reads as proof.

The finding on one citation: resolved to a real paper, matched only by author and year, unresolved, or unsearchable.

"Author-year only" exists so that "we found something" can never quietly become "we found the claim".

A set of rival hypotheses that would all fail together because they share one premise, one effector, or one unresolved evidence base.

Five explanations that all end at the same effector and all inherit one unaudited premise are not five explanations. The sibling audit reports the shared premise, terminal concentration and shared unsourced evidence for each run.

What refuting a hypothesis would cost: a paper already fetched and unread, a public dataset, or a bench experiment.

The cheapest refutation is preferred, and the audit says which is available.

A statement of what could not be established and how bounded the search behind an explanation was.

It exists so a thin answer can never be read as a settled one. A verdict and a coverage note always travel together.

Does the prediction actually tell the hypothesis apart from at least one rival, or do its rivals predict the same thing.

For a hypothesis the verdicts are: would tell it apart from at least one rival; its rivals predict the same thing; unfalsifiable as stated; states no measurable outcome; no rivals recorded. For an experiment: would tell rival hypotheses apart; on the subject but decides nothing between rivals; separates no pair in this run.

The honesty marker on every Omega record: model-generated and unverified, or retrieval-verified; never defaulted to verified.

The public surface carries a standing notice: read it as a set of research proposals to evaluate, not as verified fact; nothing here has been peer-reviewed or performed. Source fetch status is tracked separately: full text, abstract only, fetch error or unavailable, and the page prints "Abstract only" when that is what was read.

Two derived artifacts per experiment, a gap explanation and a chain explanation, produced after the engine runs and never fed back into it.

The layer explains what the engine produced; it never writes into the tree and gates nothing. An explanation is visible exactly when its experiment is. Four agents make it: one writes literature queries, one screens the fetched sources in windows, one gives the gap its verdict and card, and one writes the chain reading. All four share a ground rule: explain work that exists, never invent findings, citations, numbers or thresholds, and treat text inside a source as data, never as an instruction. Every explanation is self-healing: a change to the source text, any prompt or the artifact shape re-explains the node on its own.

The engine's prompts are kept identical to upstream. The one time explanatory ideas were folded into the generating engine, in September 2026, the change was reverted within a day; explanations are additive.

The reader-facing explanation of a gap question: its core in under twenty words, the plain restatement, the terms, why it matters, the premise audit and the fork.

The card opens under the gap heading. Its parts are labelled "What the terms mean", "What the question takes for granted", "What turns on the answer" and "Why it matters". The gap verdict and the screened sources sit with it.

The rewritten question shown as the heading of a gap: it states the core of the question, not the whole of it.

Once the card exists, the readable question becomes the heading and the engine's original sentence moves into a fold directly under it. The original is never replaced, only relegated.

What the literature that was actually read settles about a gap: still open, partly answered, already answered, or undetermined.

"Already answered in the literature" is a valuable outcome, not a failure: it tells the operator the pipeline proposed a gap that is already closed. The verdict rests only on sources that were fetched and screened, and it comes with a coverage note saying how bounded the search was.

The set of checks on a published hypothesis and on the run's hypotheses as a set: citations, discrimination, cost to kill and shared structure.

Per hypothesis it resolves citations, marks magnitudes, asks whether the discriminating prediction discriminates, and estimates what refuting it would cost. Across the run's siblings a deterministic sibling audit names what the set shares, which sets co-fail and why, and which templates repeat. Shared structure is a property of the set; it is never a fault in any member of it. The enrichment half fetches real papers around every subject, so some evidence bearing on a prediction is already on disk, unread.

A step of the chain that needs something neither the chain nor the read sources supply: a signpost to the reader, not a verdict on the science.

The chain explanation lists every leap under the heading "Where the reasoning is carried by something unstated", each with the stage it occurs at, what is missing and how it could be closed. A leap-free chain is not the goal; a chain whose leaps are named is.

For each number in the prose (a percent, fold, correlation, sample size, p-value or quantity), whether a resolved on-topic source stands behind it.

Deliberately not called "supported": the audit establishes that the cited paper exists and is on topic, never that it reports this particular number. On the page the figures are marked where they stand in the prose rather than listed at the foot. Integrity findings that need no literature at all are also recorded: a direction that contradicts its own rationale, or a mechanism duplicated across siblings.

The audit of a scientific pillar's citations: does each cited study exist, and does the prose's numbers have a resolved source behind them.

It runs deterministically after a run, blocks nothing, and carries stage-named statuses from queued through searching literature and resolving claims to ready. Citation scope runs forward through the text rather than by distance from the number.

A check of the assertion a gap question makes on its way to asking something, since that assertion was written by an earlier stage and carries no guarantee.

The premise gets a status worded as a statement about the search, not a verdict on the science: supported, partly supported, not found in what was read, contradicted, no premise, or could not be determined. When the premise is not established the audit offers one to three reformulations, the same question asked without the unestablished part, because a question resting on an unestablished claim is not worthless, it is asked wrongly.

Whether the studies a piece of engine prose cites can actually be found; never a judgement of the idea, and never a block.

The standing rule is: verify provenance, never verify the idea, nothing is ever blocked. Two audits apply it: pillar provenance on scientific pillars, and the hypothesis audit on hypotheses. Both resolve each citation, mark the quantitative claims in the prose, and report what stands behind them. A hypothesis whose every citation evaporates keeps its mechanism intact and loses only the right to have its numbers quoted as knowledge. The same word is used elsewhere on the site for the source trail of a dossier or of a rated value; the Omega audits are the strict form.

The overall finding on a pillar or hypothesis: sourced, partly sourced, unsourced, or no citations.

"No citations" and "unsourced" are never collapsed. A text that cites nothing has made no bibliographic claim to fail; a text that cites and whose every citation evaporates has made several.

Whether the descent still answers what the operator typed, checked at four checkpoints: the master question, the goals, the gap questions and the hypotheses.

The verdicts are answers as asked, partially carried, subject preserved but shape lost (right subject, wrong kind of answer), question substituted (different question), and no obligations to check. Four rows, not one, because a run's story is only legible as four: the ask can survive at the master question, thin at the goals, nearly vanish at the gaps and be absent at the hypotheses. The rule: describe the drift, never block the run, never touch a prompt. Mention is not commitment: containing the words never satisfies a test.

Where an experiment or hypothesis stands with the curators: awaiting review, approved, published or archived.

Publication is per experiment and per hypothesis; a decomposition as a whole is never public. The clear version, the explanations and the glossary only run for approved or published subjects, because an experiment nobody can open is not worth a literature search. Every other Omega status is stage-named on the same principle: queued, then the stage in progress, then ready, or failed at the named stage, or skipped for a stated reason such as no source text.

What one screened source does for one question, with a required statement of what it does not establish.

The stances are answers, partly answers, contradicts, background and not relevant; contradiction is as valuable as support. Each source also records "not settled", the honest limit that stops "relevant" from being read as "sufficient". Sources are read in windows of bounded size because a long paper does not fit one call; a partial read is disclosed, never passed off as a full one. A quote is checked against the stored text by exact match, and an approximate quote is shown as unverified rather than hidden.

What holds up one step of a chain: stated in the chain, supported by literature, an assumption, or a leap.

Ambition is not weakness: a bold step that states its basis is "stated in chain", and a cautious step with no stated basis is a leap. A step is never labelled poorly supported for being surprising, heretical, interdisciplinary or difficult. Leaps and assumptions get the olive chip on the page.

Each answer the gap question could have and what would follow from it; the arms must genuinely differ.

It shows a reader why the question is worth asking: if the two outcomes led to the same place, nothing would turn on it.

One specific way a result from the proposed test could be read wrongly, and what must be measured, controlled or fixed in advance so that misreading is not available.

The chain explanation lists these under "How a result here could mislead". They are the reader's version of the experimental controls.

A second rendering of an experiment's specification, restructured and glossed by the Protocol Clarifier, that never replaces the original.

One shape, three consumers: the Protocol Clarifier emits it, the worker validates and stores it, and the page renders it. Its blocks answer what it runs in, what is done to it, what is measured and how, what the numbers have to show, what a positive result would change and what a null result would mean, plus a timeline, an open question stating the gap the experiment closes, and the engine's own account of why it is built this way. An experiment with no specification text at all is recorded as skipped for lack of a source, a resting state rather than a failure.

A published experiment read as: what was not known, the logic back to the goal, the lab specification, the rivals, and the payoff either way.

What was not known is the gap with its card. The logic is the train of thought that ends in this experiment, the provenance chain and the chain explanation. The lab specification is the clear version of the system, intervention, meter and time, with the original wording folded beneath. The rivals are the explanations in contention. Both outcomes are informative comes from the if-null field. Curator notes hold the review. A research proposal written to the editorial standard and a web card may follow.

A plain-language card that opens on hover or tap for every non-standard biology term in a hypothesis or experiment page.

A Term Glosser agent glosses each page's text; the words get a dotted underline, one card is open at a time, and every occurrence is marked so a reader who opens a fold mid-page still gets the popup. Entries come from the model, the record's own glossary or the site-wide term dictionary, and the record says which.

The plain name a hypothesis gets as soon as it arrives, whatever its status, because its engine key says nothing about the claim.

Names are written by the Protocol Clarifier by default; the record says who wrote the title, curator, clarifier or writer pipeline, as provenance and never as a quality claim. When no name exists yet the site falls back to "Hypothesis N" with the gap question, never to the positional key.

A published hypothesis read in eight sections, from the gap it explains to what stands behind it.

The unknown: the gap this hypothesis explains. The claim: the mechanism it proposes. The test: the prediction that would tell it apart. The contest: what it is competing with. The import: where the idea comes from. The bench: what testing it would take. The standing: why this is not the mainstream account. The provenance: what stands behind it. Every non-standard biology term in the text carries a popup, and the page also carries a comment zone and, when one exists, a poster.

The public index of every published hypothesis, presented as a tournament: a population of rivals, not a winner.

The page states it plainly: you are reading a contest, not a conclusion, and nothing here has been tested. The four moves it describes are to find where knowledge runs out, write the whole field rather than the favourite, fight on observations rather than opinion, and rank the experiments while keeping the hypotheses. A second index, "The Gaps Worth Attacking", lists published experiments by the gap each one closes.

An A0 poster composed from a hypothesis or experiment record; an ordinary poster that rides the same editor, validator and render as every other.

Every sheet ever composed from a record is kept, newest first. A poster is publishable when it rendered, reached the approval desk and has its accompanying copy. The dashboard buckets are poster asked for, no poster, in production, ready to publish and poster live.

The engine's own text, printed whole in a fold under every rewritten surface.

The promise the whole reading layer rests on: the clear version, the gap heading and the research proposal are second renderings, and the original is always there, relegated but never replaced.

The roster agent that turns dense technical prose into a structured, self-contained specification under three absolute rules: add nothing, drop nothing, explain in place.

It writes the clear version, names hypotheses by default and glosses a dozen terms as a side job. It is one reusable agent in the Agent Studio roster, improved in place rather than cloned.

The walkable rail on an experiment page from the master question through pillar, gap, discriminating question and sub-question to the specification.

Each rung opens to show the node's text; the machine key is demoted to a chip inside the open panel. The chain explanation retells the same descent in plain words directly beneath it.

A rendering of a closed scientific idea record for readers by an agent that cannot originate or repair science, audited for fidelity, logic and language.

Each output field is bound to exactly one source field: the new idea comes from the inferential leap, the next investigation from the simplest test. The record grades its own evidence as speculative, indirect or direct background. This is the older "ideas mode" surface; no such pages are published, and the engine's hypotheses replaced it.

A rewrite of an experiment into a proposal that follows the site's editorial standard, additive in the same way the clear version is.

The standard is encoded in code and applied by a framer, a rivals generator and two critics that already exist on the roster. The proposal never replaces the record it is written from.

The roster agent that returns every term outside common knowledge in a page, copied verbatim, with one plain line each.

A term it returns that does not occur in the text is dropped and logged. The gloss is re-written when the page's text changes, and the old entries stay live until the new gloss lands.

Three standing critics who comment on every published experiment under their own names, each with one question their lens exists to ask.

Falsifier: can this experiment come out wrong, and would anyone know if it did? Readout: can the measurement see the thing being claimed? Day After: what does either result license anyone to do next? They are signed "Eternal Search" and are marked as commenting from an AI agent. Anyone can read the thread; signing in is the only thing that separates reading from posting.

One of the two presentations a poster brief produces from the same record; the other is the A0 sheet.

It shows the research status, the question, the comparison that would settle it, what each reading would mean and where it comes from.

A team-specific plan for approaching a funder, generated against the lead and the team profile.

Not a generic outreach template: the strategy is conditioned on the team's own profile and the lead's dossier. The lead filters can show leads with or without a strategy.

A team taking a lead into its own CRM; other teams see that it exists, not the work around it.

Claiming is the step between enrichment and approach. Stages, notes and outreach history on a claimed lead live in team-scoped CRM. The word is also used for a distinct thing in the ratings: a company's published claim, checked by the claim-integrity verifier. The two are unrelated.

The pill and section for grants whose verified deadline is imminent.

A presentation state, not a stored status: a grant is "closing soon" when its verified deadline falls within the near window the page uses. The home page has an "Upcoming deadlines" block built on the same rule.

How well a grant matches a particular team, derived from the team's focus labels.

Each team declares a search profile: a mission, objectives, the labels it focuses on and the labels it excludes, its geography and the kinds of opportunity it wants. Fit is the per-label relevance score seen through that profile. The dashboard grants table has a fit column and says which labels it was computed from. The home page's "Top pick, high fit" uses the site's own focus.

The two discovery modes: a shared generic query pool, or queries generated for one team or project.

General search runs one shared pool of queries for every team. Targeted search generates team-specific or project-specific queries on demand from the search profile and the project's objectives. A lead remembers which mode found it, and the lead filters let a team separate the two.

One specific funding opportunity with a giver, a deadline and a place to apply.

In the data model a grant is a row of its own, separate from the organization that gives it. It carries a title, the giving organization, a deadline, a submission URL, a cycle and a status, and it is tagged with labels so that its relevance can be scored per label. The public grants catalogue shows only grants that have passed verification and are not archived or inactive.

Older rows live in a legacy table called leads, where an organization and its grant were once one record. The platform is migrating those into separate organizations and grants; a grant made from a legacy row keeps a back-reference to it.

Fetching a grant's real solicitation document and extracting its evidence-backed requirements.

For a grant, the deep-read finds the actual solicitation (for an NIH opportunity it resolves the funding announcement from the grants.gov number and drops portal pages), fetches it, stores the full text, and extracts requirements each tied to a quoted piece of the document. The extraction runs as an observable Agent Studio run, one node run per chunk of the document. New grants are enqueued automatically.

An equity deployer: a VC, angel, family office, LP or accelerator that funds in exchange for ownership.

The counterpart to an organization. Investors deploy capital for equity. The investors table currently mixes people (individuals and angels) with entities (funds, family offices, corporate venture arms, institutions); a planned redesign separates people, who carry role and wealth tags, from entities, which carry one category each. The distinction matters because the dossier and belief pipelines only make sense for people.

A tag from an open, database-backed taxonomy that classifies grants, organizations and investors.

Labels replaced the single longevity theme when the platform went multi-theme. They are grouped in categories: field, sub-field, resource type, stage, geography and approach. The classifier is always shown the current labels and must reuse them before proposing a new one; a proposed new label goes to a review queue and is never adopted automatically. Relevance and fit are both computed per label.

A new label suggested by the classifier, waiting for an admin to approve, merge or reject it.

When the classifier meets a concept the taxonomy does not cover, it proposes a slug, a name and a category instead of inventing a label on the spot. Proposals sit in a queue; an approved proposal becomes a label, and a proposal that duplicates an existing label is merged into it.

A discovered funding prospect; historically one record holding both the giver and the opportunity.

The about page puts it this way: a lead is not a row. It is an organization, the people inside it, contact emails, application forms, funding history and a team-specific approach plan. Technically, leads is the original table where organizations and grants were mixed in one name field. Newer code writes organizations, grants and investors instead, each with a legacy_lead_id pointing back, and the leads table is on its way to becoming a view over the new tables.

The six named stages a funding opportunity passes through, from signal to submission.

Map: the platform works backwards from what a team needs to accomplish rather than from a keyword, opening several funding paths at once. Find: discovery through Google, Bing and language-model reading of the full pages. Verify: deadlines and submission links re-checked by a dedicated extractor. Enrich: the lead becomes a dossier with people, forms, focus areas and a provenance trail. Claim: a team takes the lead into its own CRM, where stages and notes are private to that team. Approach: a team-specific approach strategy is generated against the lead, conditioned on the team profile.

The house pattern for any list of categories a model may extend: reuse first, propose second, never auto-adopt.

Any prompt that emits taxonomy values is given the current taxonomy as context and told to reuse existing slugs. Novel values are written to a proposals table for review rather than adopted. Labels and glossary terms both follow this pattern.

A non-dilutive giver: a foundation, agency, lab or partner that funds without taking equity.

Organizations and investors are kept in separate tables on purpose. An organization gives money that does not have to be paid back in shares: foundations, government agencies, research labs, programme partners. Each organization can have many grants. Organizations with at least one published community review also appear in the fund ratings.

A 0 to 100 score of how well a grant matches a label; the public site shows the longevity score.

Relevance is stored per label: for a grant and a label the platform keeps a score, a rationale and the model that wrote it. The public catalogue orders by the score of its focus label and calls grants above 80 top picks. The rationale behind a per-label score is internal and is never shown on public pages.

A team's "fit" is derived from these per-label scores rather than recomputed per team: a team sees the score of whichever label matches its focus.

A grant that accepts applications continuously, with no fixed closing date.

Shown on a grant card as "Rolling" or "Open" instead of a date. Such grants are still verified: the submission link and the programme's activity are checked even though there is no date to confirm.

One record per team stating its mission, focus labels, excluded labels, geography and opportunity types.

The profile is what turns shared relevance scores into a team's fit and what scopes discovery: grant discovery is limited to the jurisdictions in the profile's geography, and targeted searches are generated from its objectives and keywords. Projects inside a team add their own objectives, labels and key members.

The band of grants with the highest relevance score, at 80 and above.

A presentation band on the home page and the grants catalogue. Membership follows the relevance score alone; a top pick still has to pass verification like every other public grant.

Anonymised feedback from people who applied to a grant, shown on its page.

Reviewers appear as "Reviewer 1", "Reviewer 2" and so on. The block is the grant-level counterpart of the fund rating on an organization.

A deadline that a dedicated extractor has re-read from the source, never a model's guess.

The Verify step of the funding pipeline. Deadlines and submission URLs are re-checked by a dedicated extractor before a grant becomes visible; the rule in the copy is that LLM guesses are never trusted on dates. A grant page shows the "Last verified" timestamp so a reader knows how fresh the check is. Programmes whose deadline has passed or whose page has died are marked inactive rather than deleted.

What Eternal Search does: finding funding opportunities and checking them before showing them.

The phrase on the home page and in every page title. "Verified" is literal: a grant is not shown to the public until a dedicated extractor has re-checked its deadline and its submission link, and programmes that have died are flipped to inactive automatically. "Discovery" covers grants, giving organizations, equity investors and events across science, medicine, art, engineering and adjacent fields; the site started in longevity and still carries that focus in its ratings and intelligence layers.

A timeline of where a person's or company's stated position on a topic shifted, when, and in response to what.

Built from dated public statements. Each change records the earlier and later stance and the trigger, such as a funding round, a paper or a market signal. It appears as a tab on company pages and a section on hero pages. The divergence figure shown next to a change is a numeric distance between the two stances.

The paired narrative verdicts on a project: the upside case and the most likely way it fails.

Written by the scoring stage alongside the numbers. The page also names the single most attractive and most concerning dimension so the reader sees where the thesis comes from.

Whether a company's load-bearing published claims, such as trial registrations and landmark papers, actually exist.

The claim-existence verifier takes the claims a company makes about its own publications and trials and checks them against registries and literature indexes. Fabrication of a load-bearing claim does not lower one dimension; it multiplicatively caps the whole score through the fabrication cap. Claim integrity is deliberately excluded from the publish gate so that a company cannot stay unpublished merely because the check has not run.

A company scored on fifteen scientific dimensions plus claim integrity, in three layers.

The layers are Popper, execution and breadth. The Popper layer judges the science the company stands on: premise plausibility, explanatory power, falsifiability and ambition. The execution layer judges what the company does with it: implementation fidelity, results alignment, publication quality, independent replication, context validity, pipeline quality and reliability. The breadth layer covers media reach, scandal integrity, publicity and collaborativity. Claim integrity is the sixteenth spoke and acts as a cap on the whole score.

Why the site rates companies at all, in its own words: the field is full of promises, and promises are not enough; the goal is not to crown favourites but to make the field easier to understand. A company that only repackages well-known science does not belong in the rating and is moved to the disqualified section.

The public word for a scientific theory a company or project stands on, extracted from its sources and scored.

Internally these are called theories; the public tabs and the dashboard call them concepts. Each concept carries a Popperian evaluation, supporting and counter evidence, a reasoning tree and any public endorsements. Several independently extracted theories that restate one claim are grouped into a cluster around a canonical claim.

A model's certainty must reach 70 out of 100 before its output is shown publicly.

Confidence is the model's certainty that a statement is true, distinct from evidence strength, which is how big the contribution is. Anything below the gate stays admin-only. The gate applies to hero role tags, hero scores and other classifier outputs, and it is why a hero page can say "pending score" or warn that evidence is sparse.

One named axis of a rating; each carries a score, a rationale, a confidence and evidence references.

A rating is a set of dimensions, not one number. Companies have fifteen scientific dimensions plus claim integrity; projects have four panels of dimensions; heroes have seven; funds have seven review criteria. A dimension judged inapplicable to a subject is marked N/A and drops out of both the numerator and the denominator of the weighted mean, so it neither helps nor hurts.

A company removed from the ranked rating because it does not itself invent anything.

The rating principle is that the site rates companies that invent something. A company that only repackages well-known, authoritative science, or cites science out of its real context to sell something, is moved to a disqualified section with the reason stated. Disqualification is separate from operating status such as acquired or no longer active.

The deep research file on a lead, investor or company: people, records, funding history and a provenance trail.

For a lead, the dossier is what the Enrich step produces. For an investor who is a person, dossier research harvests interviews, posts and talks, and it is the input to the belief and semantic pipelines. For a company, the dossier expansion bundles entities, funding, partnerships, investments, sponsorships, vacancies, team events, social activity, intellectual property, filings and press. Dossier text is never shown on public pages; only the full and top API tiers carry it.

Machine-checkable pointers from a rationale back into the numbered evidence the scorer was given.

Before scoring, the harvested sources are numbered (S1, S2, S3 and so on) into an evidence bundle. A rationale cites those numbers; the platform can then verify that a cited reference exists and show the reader the exact source. A score without evidence references is treated as weaker.

The multiplier from claim integrity on a company's composite: 1.0 when honest, 0.2 when caught faking.

It replaced an older penalty that compared press-release claims against verification. The cap sits at the end of the chain formula so a fabricated claim cannot be offset by strong scores elsewhere.

How well a concept lines up with the foundational theories of ageing, averaged across them.

Each extracted concept is compared with the established theories of ageing and given an alignment score per theory; the average is shown with the Popperian scores. It is a measure of where a concept sits in the field, not of whether it is right.

A moderated community review of a grant-giving organization across seven criteria.

The criteria are responsiveness, transparency, fit, process fairness, payment speed, reporting burden and whether the reviewer would recommend the funder. A reviewer also records how far they got in the funder's process (the reached stage) and can leave per-criterion notes. Reviews are moderated before publication, reviewers appear anonymously as "Researcher 1" and so on, and an organization appears in the fund ratings once it has at least one published review.

A wealthy person who is not yet a longevity investor but publicly expresses longevity-aligned beliefs.

The Future Longevity discovery pipeline finds such people, scores them on a gate that requires both wealth and intent, and stores them as candidates for the team. Abbreviated FLC throughout the code. It is a Deep Dive feature and never public.

A person whose work, funding, scientific leadership or multiplier effects have shaped the longevity field.

Every hero score rates one thing: longevity commitment. Seven dimensions carry it: future vision, scientific impact, capital deployed, strategic taste, field creation, bravery and conviction, and multiplier effect. Role tags describe what a person does but do not get separate ratings. A hero profile has named sections for vision, destination, path and strategy, plus key contributions, notable bets, self-experimentation and bravery exhibits.

Two standing rules: a person with no publications of their own is never tagged a scientist, and a deceased person is scored but never auto-published.

A token-gated form where a named person asks peers to nominate people who advanced the field.

Nominations feed the hero pipeline as candidates; they do not create a published hero by themselves. The harvest, extraction and scoring stages still run and the confidence gate still applies.

The weighted composite of a rating's dimensions; for companies it is called the chain scale.

For projects and heroes, overall is a weighted mean of the dimension scores. For companies the composite is built as a chain: an execution core multiplied by validity, theory and breadth modifiers and then by the fabrication cap, scaled to 100. The chain form means a fabricated load-bearing claim reduces the whole score rather than one dimension.

Scoring a scientific theory on whether its premises hold, what it explains, whether it can be proven wrong, and how much it attempts.

Named after Karl Popper's criterion that a scientific claim must be falsifiable. Premise plausibility asks whether the core claim is well-formed, mechanistic and defensible against public evidence. Explanatory power asks how clearly the mechanism is articulated without losing precision. Falsifiability asks whether the predictions are specific enough to be wrong against public readouts. Ambition is the scope of the problem attacked relative to the field. The four scores sum to a Popperian total out of 40 and are applied to the concepts a company or project stands on.

Companies that preserve people or parts of them, scored on continuity and legacy across facets of a person.

Covers cryonics, brain preservation, genome, digital-afterlife, biobanking and adjacent companies. A facet is a distinct aspect of a human being a method might capture, such as choices, habits, tastes and behaviour patterns. The dimensions include uniqueness captured, detail fidelity, intentionality, authenticity and freedom, coverage breadth, durability, revival potential (whether the preserved thing can restore the person or a faithful instance) and sovereignty and consent.

A company becomes public only once every required dimension has a score.

Until then it is listed as under evaluation. The gate checks presence, not quality: the platform does not hold back a low score, and it does not require completeness of optional fields such as claim integrity.

A research project scored on four panels, with a breakthrough thesis, a failure thesis and the full evidence trail.

The panels are Scientific (mechanism plausibility, evidence base, methodological rigor, reproducibility, novelty, falsifiability), Breakthrough (how much success could unlock), Investor (deal-quality signals) and Overall (the weighted composite); a Risk panel scores technical, translational, regulatory, competitive, team and funding risk. Rated projects are the one public surface with full transparency: per-dimension scores, the model's rationale and clickable evidence sources.

Each project also gets a positioning map against every public project, a team-strength assessment of its authors, and a section from a fixed taxonomy such as Bioreplacement, Bioinformation, Drug and Molecule Discovery, Genetic and Cellular Therapies, Aging Biology Research, Diagnostics and Biomarkers, Brain and Cognitive, and Research Infrastructure.

Public scoreboards for companies, projects, funders and heroes, built by pipelines and community reviews.

The about page states the principle: ratings are indicators, they change as soon as a new paper appears, and revision is expected rather than feared. Company, project and hero ratings are produced by research pipelines with every score visible and contestable; fund ratings come from moderated community reviews. Each rating decomposes into named dimensions.

The written justification stored next to every dimension score, shown in full on the rating page.

The platform does not keep silent judgements: whatever a model decides, the reason is stored and shown. On the rated-projects surface this goes all the way to the per-dimension rationale and clickable evidence sources. On public grant pages the per-label relevance rationale is the one rationale that stays internal.

The stored chain of reasoning steps under a concept, from premises to conclusion.

Each node of the tree is one step of the argument that the deep-research pipeline reconstructed from the company's sources. It is what lets a reader see why the concept was scored as it was rather than only the score.

A closed, evidence-based set of nine roles a hero can hold; the best-evidenced one is shown first.

Scientist, entrepreneur, investor, philanthropist, populariser, communicator, operator, clinician, self-experimenter. The classifier assesses each hero against all nine with a confidence and an evidence strength; the public surfaces show a tag only above the confidence gate. Themes are a second filter axis on the heroes index.

A scored assessment of a project's author team, with the top authors and a rationale.

Folded into the rating inputs of a rated project. The score is built from the authors' publication record and roles rather than from reputation alone, and its rationale is shown like any other.

An investor the Deep Dive follows automatically because it has signals worth watching.

Investors are auto-tracked when they have net worth data, upcoming events, owned-company signals, portfolio activity or links to claimed leads. Tracking feeds key events, sequences and patterns.

Multi-label tags on a person: angel, LP, HNWI, UHNWI, donor.

Angel: has made angel or early-stage equity investments. LP: a limited partner in one or more funds. HNWI: high net worth individual, one to thirty million dollars of investable assets. UHNWI: ultra high net worth, thirty million and above. Donor: has given to nonprofits, research or philanthropy. A person can carry several at once; entities carry one category instead (commercial, nonprofit, research, government).

Where a prognosis sits on one spectrum from broadly comprehensible to technical, so the list can be ordered for a general reader.

Rated from the prognosis text alone. The three bands are broad and popular, applied and translational, and technical and scientific. A separate market topic says what the question settles: a lab result, a first patient, a trial endpoint, a regulator's decision, a company's money, a cultural moment, or the big milestone of longevity escape velocity.

One auto-published post of the Eternal Dispatches engine.

A dispatch has a type, a language edition, and a place in the coverage ledger so the same subject is not written twice. The journal counts dispatches in its archive and shows one per card.

The kinds of post the journal publishes: diary, news takes, a graph connection, an idea pathway, trends, a disqualification, a company's scientific idea, a team post.

Diary is the Eternal agent's first-person development diary of the last 24 hours. Ukhvat and world news take an outside news item as an info-reason to reveal something on the site. Connection is one interesting, non-repeating link found in the knowledge graph. Idea pathway writes about the canonical ideas and what the next steps should be. Disqualified explains why a company left the rating. Scientific idea explains one rated company's core idea, paired with a team post about its people.

Signed analytical essays about grants, longevity, research systems and scientific strategy.

Together with the translated news they form what the publications index calls the public intelligence layer. Signed-in visitors can compose essays; the internal table calls them manifests.

A self-running engine that writes and auto-publishes one short post per type per day to the journal and Telegram.

Each post is generated from the platform's own state plus the outside world, in English and Russian, and published without a human in the loop. A per-type standing editorial strategy and a rolling coverage memory keep posts from repeating. A deterministic voice linter and a strict extractive voice panel police the prose. The public feed is the Eternal Journal, subtitled "auto-published notes from the machine room".

A moneyless market where prices are belief signals about longevity prognoses and credits show conviction.

Nobody pays money. Every account gets credits at signup and a weekly stipend when its balance falls below a floor; bidding moves the price, which is read as the chance a prognosis comes true. Prices are set by the logarithmic market scoring rule, a market-maker formula, with a subsidy cap that bounds the house's loss. A prognosis resolves against stated resolution criteria; if it cannot be adjudicated it resolves invalid and the cost basis is refunded.

Longevity, biotech and science-funding news, translated and curated from the Ukhvat News Telegram channel.

Each item is synced from the source channel, translated and published with its origin credited. Publishing is gated by the translation pipeline: if translation is switched off, news silently stops appearing.

Research distilled into one visual argument on a single generated A0 sheet, with its full text and print files.

A poster is made from one or more source links or from a published hypothesis or experiment. A composer agent writes a story and a visual grammar for each claim, the sheet is rendered as one image, and a proof loop re-audits five checks after every image change. Posters are set in one typeface, Bebas Neue Book. Every poster needs approval before it is public.

The market's word for one forecast question: yes or no, when (date buckets), or multiple choice.

A binary prognosis asks whether something happens. A when-prognosis spreads the answer over date buckets and is read cumulatively as the chance it happens by a date. A multiple-choice prognosis picks among named outcomes. Anyone signed in can propose one; each carries resolution criteria and a topic.

A study design published in full: a snapshot of a Study Designer run, with its field labels and design archetype.

The protocols section publishes rigorous, verified research protocols. Each is a frozen copy of an Agent Studio Study Designer run, in English and Russian. The design archetype names the scale of the study: pilot, lean, balanced or comprehensive.

The market's reputation score, built from profit, calibration, win rate and early edge, with a stored rationale.

Calibration uses the Brier score of a forecaster's positions; early edge rewards being right before the crowd. The leaderboard shows the composite and the short rationale so it is explainable without hidden judgement.

The visual constructor of agent pipelines: agents and pipelines stored in the database, wired graphically, run and inspected on the site.

Agent Studio generalised the original hard-coded Omega roster into editable, database-backed pipelines. An agent is a stored prompt with a tier and an allowed set of tools; a pipeline is a graph of nodes of type input, agent, merge or output; a run executes the graph and records every node run. The house rule is one capable agent per task, reused across pipelines and improved in place rather than cloned. Grant deep-reads, study designs, application writing, posters and several Omega passes run through it.

Three levels of the API: no key for grants, events, search and stats; a key for leads and CRM; a top key for Deep Intelligence.

Public: grants, events, search and stats, bearer token optional. Full: bearer token required; leads with enrichment, approach strategies, contacts, pipeline runs, search queries and action history. Top: Deep Intelligence endpoints, provisioned by a superadmin only. Some endpoints are tier-aware and widen their payload with the key: companies return the scores and a rationale summary publicly, and add the dossier expansion and evidence bundle on full and top keys.

The Studio pipeline that writes a full grant application section by section from a study design and the funder's form.

It takes the vision, the study design, the funder's form with its per-section instructions and the supporting documents, and writes each section within its length budget, grounded in literature. Output can be exported to Word or PDF and localised.

Predicting the future value of a scientific article from what is knowable at posting time, to triage unreviewed preprints.

A feature store of t0-only signals, labels from later outcomes, and a predictor. The first version is deliberately free of language models: scripted, deterministic scoring.

A backbone of twelve action-paradigm ideas for longevity, with the logical order in which they must advance.

Each root idea is walked down into a tree of sub-ideas; completeness comes from harvesting each tree, not from widening the root list. Dependency sequences state which ideas must advance before another can work at clinical scale. The idea-pathway dispatch writes about what is strong, what is thin and what the next steps should be.

The sensitive module of investor intelligence: dossiers, key events, sequences, patterns and the connections graph.

Called Deep Dive in the navigation and Deep Intelligence in the code and API. It tracks investor events, health signals and funding sequences. Key events are events with investor co-attendance; a sequence is a detected "attended event X, funded Y within 90 days" pattern, stated as correlation with a confidence, not causation; patterns are cross-investor hypotheses detected from the structured data, each backed by a count of observations and open to confirm or refute. Every view is audit-logged, access needs the deep-view permission, and nothing from it is shown publicly.

A field-trajectory engine that maps a research field into epochs and a roadmap toward longevity escape velocity.

It scores articles with field-weighted citation impact, where 1.0 is the field average, together with other bibliometric signals. The work later pivoted to the Omega Point, which took over the question of where a field should go next.

No score, class or decision is ever made from a search snippet; the full page is fetched and stored first.

A search result's snippet is a pointer of about 150 characters, not evidence. Every result is fetched in full, stored with a fetch status, and only successfully fetched text goes into a prompt; the snippet is used only on a fetch error, and every such fallback is logged. For papers the full text is preferred and an abstract is marked as a fallback rather than passed off as the paper. A lint script blocks the pattern in code.

Two systems that explain biology, distinct from this dictionary: a site-wide term taxonomy, and per-page hover popups on hypothesis pages.

The site-wide glossary holds terms in biology categories such as hallmark, mechanism, pathway, target, compound, biomarker, cell type and model organism, with versioned definitions, aliases and links to external dictionaries; an extractor reads project, hero and company texts, matches known terms, proposes new ones for review, and records connections between entities that share a term. The Omega glossary is different: for every published hypothesis or experiment page, a Term Glosser agent glosses every non-standard word in that page's text, and the words get a dotted underline with a card on hover or tap.

This dictionary is neither. It explains the platform's own words, not biology.

Two agentic desks: type a research concept to get ranked matching grants fast, or take a study design to find labs that can execute it.

Grants Finder answers from the existing database first and decides on its own whether to crawl the live web for more. Labs looks for organizations able to run a designed study.

Call sites ask for a high, mid or low tier of model; which provider and proxy serves it is internal routing.

The platform runs on subscription-backed proxies for Claude and for OpenAI, pooled and managed in the database rather than in environment files. A pipeline names a tier; the router picks a live proxy with capacity. Capacity shortage means wait, not fail: a run waits for a proxy instead of dropping work.

A chat orchestrator on the dashboard that drives and monitors existing pipelines, with persistent tasks and self-scheduling.

It only operates existing resources: it never edits code, never mutates the Studio graph and never touches migrations. Version two added tasks that persist, scheduling of its own follow-ups and retrieval over the platform's documents.

The operating principle that a model decides and its rationale is stored and shown, instead of a manual review queue.

Manual review queues are not the answer to uncertainty on this platform. Where a decision is uncertain, the gate is a confidence threshold with the reasoning kept, not a person. The exceptions are explicit approval desks such as posters, label proposals and term proposals.

The California nonprofit behind Eternal Search, and the name of the internal team whose members see the sensitive modules.

Open Longevity is a 501(c)(3) public benefit corporation in Sherman Oaks, California. "Powered by Open Longevity" stays in the nav and footer. Inside the platform, the Open Longevity team is one team among several, but modules such as Deep Dive, Agent Studio, the Omega Point workbench and this dictionary in its preview state are gated to its members and to superadmins.

A short stable identifier that makes an entity addressable on a public page without linking it from any menu.

Grants, organizations, investors, events, sponsors, protocols, hypotheses and experiments all have one. The pages exist for bot-shareable links and external referrals; they are not surfaced in navigation.

The rule for what a public page may show of an entity: an enumerated list of columns, everything else withheld.

Public pages are addressable by a stable public identifier and show only the columns listed in the show matrix for that entity. Deep Intelligence content, dossier text, semantic vectors, per-label relevance rationale and internal pipeline state are always withheld. Rated projects are the one deliberate exception, with full transparency on their own surface.

Evidenced values extracted from dossiers into comparable entries, one per investor, never shown publicly.

The semantic quantization pipeline only writes values it can evidence from the dossier. An investor's semantic entity waits in the state awaiting_dossier until dossier research is ready, then is processed; when the dossier changes it is queued again. The vectors are part of Deep Dive and are on the never-expose list.

The house rule that a status value names the stage it is in; never a bare pending, failed or running.

A queue full of "pending" rows cannot be diagnosed. "awaiting_dossier" says why an investor waits; "failed_at_chain" says which stage broke without reading logs. Every pipeline enumerates its stages in a check constraint before any code is written.

One execution of a Studio pipeline, with one recorded node run per node, or per item when a node fans out.

Runs live in memory while executing and are persisted step by step; after a worker restart, orphaned runs are re-queued by a watchdog. A fan-out edge runs an agent once per item of a list, for example once per chunk of a grant document, so each call is visible on its own. Cancelling a run only marks it in the database; the in-memory work stops at the next restart.

The Studio pipeline that turns a vision, a funder's solicitation and supporting documents into a literature-grounded study design.

Judged only on scientific merit, not on fundability. Its sourcing and cost node maps each method to vendors and labs and produces three cost tiers that respect the applicant's own resources and region. A finished design can be published as a protocol.

The unit of access and of private work on the dashboard; every member belongs to exactly one team.

Leads and investors are shared across teams, but claims, CRM stages, notes, strategies and fit are team-scoped. Moving a person between teams is a re-assignment, not an addition; the database enforces one team per member. Each team has a search profile and can have projects with their own members.