Research prototype, not medical advice.

Astra founding charter (draft for D46 ratification)

Drafted by the chair (A1) from:

Nothing in this charter is new. Every paragraph ends with the decisions it rests on. A paragraph with no citation, or with a citation that doesn't cover what it says, is a defect that the D46 diff must flag.


Page 1: read this first

What Astra is

Astra is a team of AI agents, run for its president, whose purpose is to extend mammalian lifespan by at least 2 to 3 times. The mouse target is 3×, and it never moves. Its core is simulation: a computer model of how a body ages. Real experiments, on mice and on the president's own samples, exist only to confirm or refute what the model predicted in advance.

Cites: commandments 1–3, D115, D117.

What comes first

Nothing advances without a way to measure biological age with no blind spots. So the first job is a measure of biological age built as a set of readings, one per organ system and per cause of ageing, rather than a single number. No treatment is designed until that measure passes its tests.

Cites: commandments 5, 6, 9, D114, D115.

What to expect in the first 6 months

Cites: commandment 4, D117, D119, D121.

Decisions that belong to you

DecisionState
Separate GitHub identities for agents (D68)Refused 2026-10-20. Agents share one token.
Three missions start before ratification (D78)Approved 2026-10-20: Verification, forecast ledger, tooling.
Where your health data lives (D122)Answered 2026-10-20: public repo, encrypted; key only in a local, git-ignored .env. This replaces D116's astra-private.
May agents read your raw values (D122)Yes.
What is published about you (D122)You left it to us: default = per-measure records only for measures you list, none until you list any.
First two blood draws (noise baseline) (D116, D122)Later, when you say so.
scoring environment, main only (D122)Done 2026-10-20 by A1 at your request.
PROTECTION_READ_TOKEN (D122)Done with your full token, at your choice. Moving it into a main-only environment: refused 2026-10-20; it stays a repository secret.
Charter-level decisions D124, D125 (part 10, part 11)Voted by the managers with D46; yours to confirm at merge.
Re-seal of code-protection settings (M2)Pending: the Code Manager sends you the command.
Mice: when and how manyYours. Each request comes with a sealed prediction (part 4).
Verification's share of workers, each quarter, ≥ 15%Yours.
Overruling a Verification findingOnly you, in writing.
Merging this charterYou merge after the managers ratify it (D46).

Cites: D46, D47, D64, D68, D78, D116, D117, D120, D122, president rulings PR1, PR8, PR9 (docs/decisions/president-rulings.md).

Limits you must know

Cites: D52, D68, D80, D108, D116 (T3-A50-3), D122, PR8, PR9.


Part 1: the measure of biological age

Biological age is a vector: one axis per organ system and per hallmark of ageing, each tied to a functional test (for example grip strength for muscle). A summary number is published, but it is never an objective, so no gain on one axis can hide a loss on another.

Cites: commandments 5–6, D114.

A version of the measure counts only if it passes six tests, against baselines sealed in advance:

  1. Coverage with proof: every organ system and hallmark has an axis, and each axis moves the right way under an intervention known to act on it, and the other way under one that shortens life.
  2. Agreement with lifespan effects on ≥ 10 published mouse interventions from ≥ 2 classes, both sexes, on a sealed test set.
  3. Negative controls: null arms of the Interventions Testing Program must read as null within their observed interval.
  4. Individual level: it predicts each held-out mouse's remaining lifespan better than chronological age.
  5. Adversarial cases written by Verification, never by Research.
  6. Declared validity range: readings beyond the effect sizes it was validated on are labelled "extrapolated".

Every reading carries its interval (test-retest noise plus model uncertainty); a reading without one isn't published. Each version is sealed before its test score is computed, and its scorecard is published with its failures. Test 2 uses only studies whose controls lived within the normal range (ITP or replicated), includes null and life-shortening arms, and reports an "everything extends lifespan" baseline next to the measure. Each standard error records its source (printed, from_curve or from_p_value); from_p_value is used only in a sensitivity run, and a standard error read from a curve enters test 2 only if the curve was digitised twice and agreed.

Cites: D114 (T1-A28-3, T1-A50-1, T1-A50-2, T1-A28-5, T1-A2-3, T1-A2-4, T1-A2-5).

Scores on published interventions are labelled "retrospective, not blind". Only sealed forecasts and held-out mice count as blind evidence. Test 4 is labelled "blind (key not isolated)" until its keys sit in the scoring environment.

Cites: D114 (T1-A50-3, T1-A10-4).

A confirmed worsening on an axis is decided by a family-wise rule that Verification fixes when the version is sealed, on axes that passed test 1. Axes that didn't are published as "unmeasured" blind spots. An intervention or combination with a confirmed worsening can't count toward 3×, can't be called progress and can't be the subject of a treatment request, unless a sealed, tested simulation predicts that a combination cancels it. Trade-offs are always reported in full.

Cites: commandment 6, D114 (T1-A1-3, T1-A28-6).

The human version uses the same axis names, built only from affordable, repeatable samples.

Cites: D114 (T1-A28-4), D116.

Part 2: the simulation

Simulation v0 is an organ-level model of damage and repair whose state is the measure's axes. Deeper levels (cells, molecules) are added only after a recorded failure of the level above.

Cites: commandment 1, D115.

Interventions are mapped to model parameters from mechanistic evidence only, and sealed before any lifespan result is seen. The train/test split, its strata and the class of every intervention are sealed together in one record.

Cites: D115 (T2-A2-4, T2-A10-5).

The model is scored against a sealed simple baseline (Gompertz curve plus a hazard ratio per class) on three strata: (a) new interventions from known classes, (b) new classes, (c) combinations and transfer between sexes. It must be no worse on (a) within a sealed margin, and must win on (b) or (c) to claim it captures mechanism. Each stratum has its own sequential test; "undecided" is published as such. The model's count of free parameters is published with each score. A mapping changed after a result is seen makes every later prediction using it exploratory until a new held-out item resolves.

Cites: D115 (T2-A28-6, T2-A50-1, T2-A50-3, T2-A10-4).

Measure, simulation and causal map are built in parallel, but the simulation ranks no treatment before the measure passes its six tests.

Cites: commandment 9, D115.

Part 3: the president as a resource

The president's role is to calibrate and challenge his model ("twin"), never to validate the measure. We ask him for measurements only: no dosing, no personal medical advice; any treatment goes through his doctor.

Cites: D18, D116.

The first request is a noise baseline: two draws a week apart plus one blind split duplicate, sent with a sealed expected test-retest interval. It waits until he says go.

Cites: D116, D122.

His raw data is stored encrypted in the public repo; the key lives only in a local, git-ignored .env. Agents may read raw values. The repo publishes forecast hashes, aggregates (≥ 5 resolved items per measure), interval widths as a ratio of his noise, and one count of open twin issues. Per-item records are published only for measures he lists in consent.md. Nothing about him is published before consent.md exists. Reveals and per-item scores otherwise stay private.

His twin is scored against a persistence baseline ("next value = last value", interval = his measured noise). Perturbations we forecast stay non-medical: a training block he chooses, or a vaccination he and his doctor already scheduled. Astra never proposes one.

Cites: D116 (T3-A10-2, T3-A50-2, T3-A50-4), D122 (president's answers, 2026-10-20).

Part 4: mice and the 3× target

Mice come only when the president is confident they'll be useful, and there are few shots. Cohort 1 is untreated (UM-HET3, both sexes) and tests sealed forecasts: axis trajectories and remaining lifespan per mouse. Its data trains only the next version of the measure.

Cites: commandment 2, D117.

Every mouse request is a file that CI checks: a sealed prediction hash, a refutation criterion, a reproducible power calculation, a dated search showing the experiment isn't already done, and no open S1 or S2 incident on the model.

Cites: D117 (T4-A10-1, T4-A50-1, T4-A2-4).

3× (or 2×) is claimed only after full survival, on median and maximum lifespan, against controls within the normal range, and replicated at a second site whose prediction was sealed before the first site's result.

Cites: D117 (T4-A50-2).

Mouse evidence is rated with SYRCLE risk of bias, capped by flags for short-lived controls and lack of replication.

Cites: D120 (T7-A2-1).

Part 5: data and blindness

A forecast is blind if it was sealed before any public trace of the quantity it forecasts. A forecast sealed after an interim result names that interim curve, hashes its extrapolated baseline at sealing, and is labelled "blind (post-interim)".

Cites: D118 (T5-A28-6).

Before every seal, Literature runs a dated search for public traces, meeting abstracts included. The seal is refused if that search is more than 7 days old.

Cites: D118 (T5-A2-4, T5-A10-4).

Data is fetched by manifests with licences recorded; nothing over 10 MB is committed outside the allow-list. A lifespan study that doesn't report a cause of death records it as "unknown", never "no change". Verification approves every resolution criterion and every licence before use. Each survival curve used for scoring is digitised twice, independently, and the build checks agreement.

Cites: D118 (T5-A50-2, T5-A50-3), D121 (O9, O15), D120 (T7-A2-2).

Part 6: organisation and missions

The president owns Astra. A1 is his vice-president and interface, and the customer of the Research, Literature and Platform missions. Managers orchestrate and never do the work themselves; workers do it. Each mission has one owner and one customer.

Cites: PR1, PR2, D119 (T6-A1-1).

Verification is independent of A1 and reports to the president. Its share of workers is ceil(0.15 × cap), never less than the president's quarterly setting; a lower re-measured cap applies at once.

Cites: D47, D64, D119, D120 (T7-A50-4).

Worker cap. This computer is shared, so the whole fleet runs at most 9 workers at once, plus managers, with at most 2 local builds (lock file, delivered with P2 in week 1). Anything heavier runs in GitHub Actions. The Code Manager re-measures the cap monthly.

PhaseResearchLiteraturePlatformVerificationDatabase
1: until split sealed and L1 drafted22221
2: after32220

Coherence and Interface start no worker without asking A1.

Cites: D119 (T6-A10-2, T6-A1-4).

MissionOwnerOutput
M0 VerificationA50Seals, audits, adversarial cases, independent recompute of scorecards
M1 Forecast ledgerA28Sealed forecasts on mouse lifespan studies
M2 Platform railsA10Write paths, decision citations, drift job, launch gate, seal on workflows
R-Metric v0A28Mouse measure v0 on public data, with its scorecard
R-Sim v0A28Organ-level model calibrated on control mice, scored on the sealed test set
R-TwinA28Noise baseline, then quarterly sealed forecasts on the president
R-Cohort 1 protocolA28Cohort protocol, sent with measure v0's scorecard
L1 Lifespan tableA2Every published mouse lifespan intervention, ITP first
L2 Metric dossierA2One file per input and function test of the measure
L3 Refutation fileA2One row per causal edge and parameter mapping
P1 Data manifestsA9 (manifests), A10 (fetch job)Sources fetched by manifest
P2 Run manifestsA10Seeded re-runs; split check
P3 Issues and ledger jobsA10issues/ register blocking sealed outputs with open S1/S2
P4 Twin data railsA10Encrypted storage, hash publishing, aggregate scores
P5 astra-sim portA10Rust port of sim v0 within the agreed tolerance

The old missions M3–M9, the RMST ranking engine as a product, and the cardiometabolic flagship are stopped.

Cites: D119 (T6-A28-1, T6-A2-1, T6-A10-1, T6-A1-2), D120 (T7-A28-3, T7-A10-3), D122.

Part 7: decision rights

Each mission protocol is sealed before any data is read, signed by its owner and by Verification, and names exactly one primary outcome. A departure without a sealed amendment is exploratory.

Cites: D2, D61, D110, D120 (T7-A50-1).

Acceptance criteria change only by numbered decision. A commit citing a decision passes CI only if that decision exists on main before the change. Two members may appeal a chair ruling to the president; the ruling stands after 7 days without answer. Only the president can overrule a Verification finding, in writing.

Cites: D47, D48, D49, D69, D120 (T7-A28-1).

Rules that are conventions, not proof. All agents push with one shared token (D68 refused). The rules below are conventions, not proof, with no end date. A determined or tricked agent can break them; detection is after the fact.

D47, D52, D69, D80 and D109 are built but enforced only from the M2 merge, after the president's re-seal; until then they are conventions too, and Verification's weekly git log diff is the only check.

RuleWhat it claimsWhy it's only a conventionWhat still catches a breach
D2 co-signed protocolsProducer + Verification both sealed itSignatures are agent names in a fileSeal hash in git history; D46/D61 review record
D46 ratificationEach member approved the charterPR reviews all come from one accountObjections quoted verbatim; A50's diff against the log
D47 Verification write pathsOnly Verification writes verification/Agent: trailers are self-declaredCI check (D47) on trailers + docs/agents.yaml; A50's weekly git log diff
D61 dual sign-offTwo units approvedSame as D2Same as D2
D66 planted-error keyVerification can't read the answer keyAny agent can read workflow outputsCommit-reveal hash; label "planted (key not isolated)" until the keys move into the scoring environment (Audits P2)
D124 model idRecords which model did the workSelf-reported trailerPresence check only; useful for drift, not independence
President's rule: no settings / no --adminNobody edits protection or rulesetsThe token allows itD52 daily drift job → alert to the president and Verification; D80 seal on workflows and scripts

The route back to real identities is the > 20 writers trigger (D124).

Cites: D66, D68, D78, D108, D120 (Audits P2), D121 (O14), D124, PR7.

Part 8: verification and procedures

Sealing follows verification/procedures/sealing.md P1–P5: only Verification seals, never an author; "no data read" is the owner's statement plus the fetch-job log and the dated trace search, and a false statement is an S1 incident; every refusal is recorded.

Cites: D120.

Audits follow verification/procedures/audits.md P1–P8: seeded sampling; seed and planted-error keys in the main-only scoring environment; an error is any field mismatch beyond declared rounding, with effect-changing errors counted separately; the D51 threshold applies per finding over a rolling 3 months, per mission; small pools keep the same rates with exact binomial bounds; nobody audits their own work.

Cites: D51, D63, D120.

Planted errors now also test the measure and the simulation: bugs planted in model code, labels swapped in training data. Accuracy is measured against known truth. An outside reviewer samples metric scorecards each quarter.

Cites: D66, D71, D102, D120 (T7-A10-2, T7-A50-1).

The seal covers lint settings, workflows and scripts; changing them needs a re-seal by the president. A daily job compares live protection with the committed settings and alerts the president and Verification. Retractions are scanned weekly, including mouse papers.

Cites: D45, D52, D53, D80, D120 (T7-A10-1).

Part 9: obstacles and dates

Goals are written as obstacles. Each has an owner, an exit and a date; Verification, not the owner, decides when an exit is met. Targets never move.

#ObstacleOwnerDate
O1No measure that reads ageing reliablyA28 + A50month 6
O2Causes not separated from consequencesA28 + A2month 9 (sim v0 sealed, month 4)
O3Cancer ceilingA28, A2month 9 (milestone month 6)
O4Combinations don't add upA28month 12 (milestone month 6)
O53 years per lifespan testA28month 9 (first early-reading score, month 5)
O6Strain and sex dependenceA28 + A50ongoing (sex transfer stratum sealed, month 4; test 2 per sex, month 6)
O7Combined toxicityA28 + A2before the first treatment request (adverse findings for every O4 combination, month 6)
O8No mice, lab or budgetA1month 6
O9Data accessA9 + A50month 1
O10Long-lived species unexplainedA2 + A28month 12 (milestone month 6)
O11Published lifespan effects are biasedA2month 3
O12Causal links asserted without refutation testsA2month 3 (L3 row for every sim v0 edge and mapping, approved by A50)
O13Results that can't be reproducedA10month 1 (run manifests + one seeded re-run in CI); then every sealed prediction re-runs before sealing
O14Gates that rely on trustA10end of week 1 after the re-seal
O15Data too big for gitA10month 1
O16No public route to SLAM longitudinal dataA2month 1
V1No blind test exists yetA50 + A28month 3
V2Hidden keys not isolatedA50 → presidentmonth 1
V3Scorecards checked only by re-running the same codeA50month 2: recompute plan + 2nd-implementation spec; recompute delivered with metric v0's scorecard
V413 procedure rules undecidedA50 → chairdone (D120)

Every obstacle needs an exit dated by month 6 or a dated milestone before then; otherwise it counts as unowned. Every row is now dated (V3 by Verification, PR #51). The register job that raises the alarm on missed dates is live by month 1 (A10).

The full exits and month-6 milestones are in arcanum/session-1/topics/T8.md.

Cites: commandment 4, D117, D121 (T8-A50-1, T8-A1-2, T8-A28-4, T8-A10-1, T8-A10-4); owner datings in the D46 reviews of A28 and A2.

Part 10: refusals, incidents and launch

Astra never gives dosing or personal medical advice; never uses unpinned external data; never scores its own forecasts; never does dual-use work. A negative control never counts as progress.

Cites: D124 (re-issues D18's refusals), D116, D120 (T7-A28-2).

Incidents are S1 (fixed in 72 h), S2 (2 weeks) or S3 (next release). Corrections are add-only. No sealed output ships while an S1 or S2 on it is open.

Cites: D54, D119 (P3).

Until Verification clears it, CI blocks any page that publishes a biological age or forecast of a named individual, or claims a lifespan effect we predicted. Every page carries a "research prototype, not medical advice" banner.

Cites: D109, D120 (T7-A10-2).

Part 11: meetings and amendments

Meetings follow the Arcanum rules: the debate happens in files, chat messages are at most 2 lines, each turn has a steelman before each objection and at least two dissents, objections are recorded, and only the clerk commits. Any member may call out flooding. Verification is never red team on its own findings.

Cites: D55, PR5, PR6, arcanum/session-1/README.md.

This charter changes only by a numbered decision voted in an Arcanum session; Verification diffs each amendment's citations before the president merges it. The commandments outrank it; a decision that contradicts one is void.

Cites: D125, commandments.

Annex: decisions proposed with this charter

Voted by the members in their D46 reviews of 1e69101: D124 accepted by all four; D125 accepted by all four with A50's reservation, now part of its text. Confirmed by the president at merge.