14. Lineage
The figure: the author
The lineage is the owner's own, and the book admits it as evidence on one condition: that it be read the way the court reads everything else, as a record. Seven projects, each with a founding text that says why the last one stopped; each a proposal, scored by a world that did or did not charge, recorded in the successor's charter, electing the next design. The lineage ran jack's loop before jack existed — and it ran it open: no pre-stated exit (proplang declared its roadmap terminated three times), no proper score (the judge of each project was its author, on his own cause), wealth reset at every project with nothing carried but prose, the operator's cost unmeasured until one project's records said "too slow". jack is that loop closed: the same loop with a record that binds, a score the proposer cannot shade, a null committed in advance, and the operator priced. That is a stronger claim than resemblance, and it is checkable: the successors' charters for the reasons, the predecessors for the missing pre-statements.
14.0 One pattern
Six lines of work precede jack — seven repositories in the cold read's count, and assize after them — and they are one pattern. Each put law around a Bayesian agent. The law grew into procedure, or the project became a research experiment with no world that charged for being wrong. Each successor shrank the law and widened the competition. jack is the fixed point of that recursion, and this chapter's claim is that the lineage is not a sequence of failures but a sequence of measurements: each project found the next necessity by hitting it, and the strongest evidence that the construction's necessities are necessities rather than tastes is that four separate attempts, by one author, converged on the same methodological rules before anyone wrote them down as rules.
The cold read of 26 October — counsel's evidence for amendment A4, by a session that read the repositories and nothing else — put the pattern in two sentences: "Lifetimes shorten and the governance share rises until the wald split"; and "No repo carries its own reason for stopping: proplang ends on unposted drafts, credence on a merge, tannen on a version sitting; each reason appears in the successor's founding text" [transcripts/PREDECESSORS-20261026.md, cross-repo synthesis]. That second sentence is the method of this chapter. The reason each project stopped is written in the next one's charter, and jack's charter §1 is the reason assize stopped.
14.1 credence (March–August 2026): the axioms, and the owner's utility
What it was. A Bayesian decision-making language in Julia and the tools built on it: a gateway that learns to route model requests, an agent library, research environments. Its constitution, the Tractatus Credentiae, states the commitments in numbered propositions: beliefs are previsions; "Beliefs change only by Bayesian conditioning"; "Actions maximise expected utility, and there is no mechanism outside EU-maximisation — not random, not capped, not scheduled"; "There is no forgetting. Decay, tempering, and forgetting factors are a second learning mechanism in disguise" [credence/CONSTITUTION.md §§2.1, 2.3, 1.41].
The commit log corrects the cold read's dates and adds a lesson the repository's prose does not state. The first commit is 14 March, not April [data/lineage-commits.md]; the oldest commitment in the lineage is a commit of 22 March, "closed-loop policies replace open-loop sequences" — chapter 1 §1.4's principle, eight days in. The discipline jack calls pre-registration was already there in July: "W1' PRE-STATEMENT: criteria fixed and committed before any real-label p1 exists", then "round 1 fired the confound — both arms discarded", then "round 2 RETRACTED — the corpus was never pinned", then a pass "on the pre-stated bar" with qualifications, all on 22 July. And the end is on the log in the project's own words: "Phase 0 EXIT DECISION: WAIT" the same day, "three workstreams overtaken by proplang", and on 9 August "Stage A RESULTS: the rail was our label, not proplang's ceiling". The core was superseded by its successor while its applications went on — the governing sidecar and the router shipped into November — which is why the repository carries no abandonment: nothing was abandoned, the reasoner moved.
The sessions' notes add what the constitution states as principle and the owner applied as correction, each dated [data/lineage-sessions.md, credence]. On 4 July he stopped a draft that had built a bespoke corrector for selection bias beside the reasoner — the note's gloss is that a framework dies of a thousand locally reasonable patches. On 1 July, ratifying a wire verb, he required that a belief scalar crossing the wire carry a contract saying it is not a decision channel, because a consumer could branch on it and reintroduce the hand-set threshold the whole design exists to remove — Fuller's congruence found at a protocol boundary three months before the charter named it. On 3 July, after three benchmark rounds had run against a measure the agent was never asked to optimise, he ratified the fix with the instruction to design the experiment before the run — pre-registration arriving as a correction. And on 20 April came the first delegation of an owner's act, merge authority to the session, on the ground that confirmation of each merge was friction without value. Each is a necessity of the construction met early and held as discipline, not yet as mechanism.
What it contributed to jack. Three things that jack takes without argument. The alignment commitment: "The agent's utility is the user's utility, unknown, inferred from behaviour. (CIRL.) The commitment defines the objective, not the solution, and cannot be gamed because the ground truth arrives from outside the agent" [§2.5] — which is why, in jack, the sovereign is the one unscored party and his wishes are preferences, not predictions (chapter 12 §12.4). The strange loop, stated and bounded: "Meta-actions — enumerate more, perturb the grammar, deepen — stand in the action space beside domain actions and are chosen by the same argmax. The strange loop is not confusion: the machinery is immutable and the hypothesis space is not. The agent changes what it thinks about, never how it thinks" [§3.5] — which is the ninth loop of chapter 11, and the rigid core and free edge of chapter 0 §0.2 before they had those names. And closed-loop over open-loop, the first architectural commitment, which chapter 1 §1.4 finds reappearing at the institutional scale.
Where it stopped. The cold read records a pivot "to governance sidecar following routing-collapse benchmark", a Stage A finding that every evidence channel the engine was fed ran at 0.99 or above and discarded half its outcome records — "The channel is degenerate because of our label" — and a recommendation that "Phase 1 should not start" [PREDECESSORS-20261026.md, credence]. The repository carries no statement of abandonment; the verdict lives in its successor's founding text: wald-charter's "builder held the pen on the reasoner, CI never ran the engine's tests, arithmetic was Float64, and its router learned a reliability with no ground truth" [wald-charter/HANDOFF.md §1]. The lesson in jack's terms: an architecture without an institution, and a score with no world behind it. credence's governor, the sidecar, ran about thirty thousand real decisions in shadow and "FAILS the waste bar (6.50% vs 0.05%)" against its own pre-stated exit, and was retired [PREDECESSORS-20261026.md, credence-governor] — the first of the lineage to be measured against a constant it had committed to, and to lose.
[G] What the sessions' notes add is above; what the owner would add beyond them is his to say.
14.2 proplang (July–September 2026): the alphabet is the prior, and the programme
What it was. A minimal Haskell language "in which only Bayesian-optimal decisions are expressible: the alphabet IS the prior, deliberation is a sentence, and the world — never the language — supplies every constant" [proplang/KERNEL.md]. Exact rationals; de Finetti's prevision first; nine expression terminals priced at one ninth each, so that the prior weight of a sentence is the exact product of its choice-widths.
What it contributed. The programme — the governance that jack's briefs still use. Obligations, findings, rulings and freeze layers; frozen oracles hashed in a manifest and signed before implementation; "pins" that must kill a seeded mutant; the rule that a green that cannot fail is a non-functional alarm; no unwitnessed claims; append-only corrections; denominators matching populations; minimality as a correctness property. Its own write-up says how these arose: "The build ran that way, and its incidents are the registry — each one became protocol text at the next freeze boundary, so the protocol the project ends with is the one its own failures wrote" [proplang/WRITEUP.md §8]. That sentence is the lineage's method stated by one of its members, and jack's amendments A1–A5 (chapter 9 §9.4) are the same thing one repository later.
The sessions on steel hold one proplang sitting, and it is the lineage's central document: 25–26 July, the exact freeze [data/lineage-sessions.md, proplang]. It opened with the owner's statement of the goal, and then, an hour in, with the sentence the whole lineage turns on: "I feel that what should be a simple system in its foundations is actually too complicated and unmanageable... All I want is a small set of primitives … and everything else composed from them... Is that too much to ask for????" Six reviewers with no shared context, each carrying one named historical failure — a theorem installed as a definition, a claim asserted n times and derived zero, a quantity defined nowhere, a type without its derivation, a convention overloaded, an object whose inputs are unstated — found tests in the frozen oracle that were green for any implementation. His ruling that afternoon is the principle chapters 4 and 8 build on: "I don't want to rely on prose and vigilance any more, ever." In the same hour he asked, in order, the questions this book answers — whether the language is necessary and sufficient for a Bayesian agent including metareasoning; how to make it a closed system in which metareasoning happens automatically; what "the oracle is law" means and whether the oracle is optimal — and said of his own repository that its architecture was no longer clear to him, its supposed author. The session's diagnosis is the one jack is built to answer: the language had been given the deletion test and the process never had — every incident minted a clause and no clause retired, so the protocol had become the bloated alphabet one level up. Its remedy, that nothing should hold a seat by familiarity including the process and the oracle, is the pins discipline; the half it could not do — retire law — became jack's lapse rule and ratio guard. The same session dated the second terminus and named the ninth loop, the agent buying its own macros, as a research programme deferred.
The commit log shows the method at its most concentrated: on 6 July alone, five increments each ran "Task 0 … stop for rulings", "Task 2 … freeze worksheet", "freeze — manifest n/n", and the day closed with "Roadmap COMPLETE"; the manifest count rises through the log like a ledger's seq — 49, 56, 63, 64, 68 [data/lineage-commits.md]. The last commits are drafts marked "NOT POSTED" and a bench "regenerated from life-agent": the consumer, not the language, pulled the last lever.
Where it stopped. It declared its roadmap terminated three times — "The roadmap terminates here" at the July freeze, a second terminus in August, a third sitting in September — and continued each time under "demand-gated residue" [PREDECESSORS-20261026.md, proplang]. The one field reading was "NOT MEASURED on 95k live events … no engine passes the registered loss bar" [proplang/EXACT_PLAN.md §13.0, quoted there]. No money. Fifty-three thousand lines of prose against eleven thousand of Haskell. The founding essay's phrase for it is "court-grade formality applied to the thing that was supposed to evolve, so it evolved at the speed of proofs" — the second way to get the shape wrong (chapter 10 §10.4). The cold read's figure is the measurement: a prose-to-kernel ratio of about five to one, which is also close to the procedure-to-object ratio jack's charter adopts as its stop [CHARTER A3.4].
[G] The sitting of 25–26 July answers most of this; whether the three termini were the method working or failing is still his to say — the session's own reading was that the roadmap should terminate and that what followed was demand-gated residue.
14.3 tannen (August–September 2026): separation by session, and the border
What it was. "A small semantic kernel plus an embedded Python DSL for defining data pipelines as compositions of pure, total, versioned transformations over content-addressed values" — and, in its own words, "Most of this repository is not the kernel. It is the machinery that decides what may enter the kernel, and the evidence that the machinery ran" [tannen/README.md].
What it contributed. Adversarial separation by session: "Each milestone is two adversarially separated sessions: one writes the specification and the property tests and signs a freeze tag; a later one implements against them and may not edit a frozen path. A law found wrong is never edited — it is superseded forward, and the mistake stays in the record" [tannen/README.md]. Decisions as data: every non-trivial choice a schema-validated record, with DECISIONS.md a generated projection never hand-edited. Poison fixtures that attack the guards, so that a guard is shown to fail before it is trusted. Signed custody. And the border treaty — repositories never import each other — which is why jack's charter has its agents living elsewhere, judged by an image pinned by digest [CHARTER §0].
The memory notes left by tannen's sessions carry the owner's words on the two things chapter 12 rests on [data/lineage-sessions.md, tannen]. On 12 September he made interviewing him for rulings a standing order, and four days later bounded it: ask only when the ruling blocks. The same day he gave the sentence this book takes as the justification of the whole apparatus: "My rulings have now been corrected by measurement on the custody question, by the spec-table guard on the supersession path, and by the floor audit on the remote. In each case the correction came from an artifact the repository built to be checkable, wielded against the owner's instruction. That is the strongest available evidence the governance layer earns its cost — not that it prevents error, but that it makes the owner's errors cheap to find and expensive to ignore." That is maxim 1 applied to the sovereign, a month before amendment A2.8 wrote it. On 16 September he called the process rigorous but too slow and asked why it carried so many dates, and the session measured it: eighteen minutes of hooks per commit, a sitting driver of nineteen steps and three hours, 265 records in 23 days — and the rigour that mattered was never where the time went. One note states the law behind the pins better than any charter clause: of nine green-that-cannot-fail defects in tannen's life, seven were inside artefacts written to close an earlier instance of the same class, so that "writing the discipline down does not install it — the production rate of this class does not fall with attention, only the catch rate does, and only when something mechanically forces the failure to be seen."
Three of its decision records, as commit subjects, are lessons jack later wrote as rules: "D0117–D0118: a verdict is one bit, and generators are shaped rather than policed" (30 August); "D0148–D0149: … a verdict that read the clock twice" (2 September) — the kernel that never reads a clock; "D0204–D0205: four false sentences superseded forward" (9 September) — append-only correction. And the red teams were scheduled at every boundary and never closed their findings on the day: "M1→M2 boundary red team: eight findings, one critical, none closed"; "M2→M3 … six findings, two critical, none closed"; "M3→M4 … six findings, two high, none closed" [data/lineage-commits.md]. That is the review loop of chapter 11 §11.6 running before the record existed to hold it.
Where it stopped. Its milestone M5 was "one Renavon vertical, byte-identical vs production on frozen inputs … (production bugs found = wins)"; it closed with a decision record saying the win condition was unreachable by construction — "M5 attests that tannen reproduces production's code; it cannot surface a production defect" — and the red team showed "a pipeline that 'computes nothing' passes all 128 laws" [PREDECESSORS-20261026.md, tannen]. The owner's time was measured, for the first time in the lineage, in the project's own records: "it's rigorous but too slow"; eighteen minutes of hooks per commit; a sitting driver of nineteen steps and three hours; a budget file with every ceiling at zero. The lesson in jack's terms is two of the construction's necessities at once: separation by session is not enough without a record that binds (the law was prose plus scripts, not entries), and a court whose laws can be satisfied by a pipeline that computes nothing has no world that charges.
[G] The memory notes give the lesson in the owner's words: the governance layer earns its cost by making the owner's errors cheap to find; and discipline does not install itself. Whether "separation by session is not enough without a record" is also right is his to say.
14.4 wald (September–October 2026): the arena, and judgement marked as judgement
What it was. Three repositories under one charter: "what is the smallest thing you can write down such that everything an agent does follows from it, and anything it should not do is unsayable? wald's answer is Bayesian decision theory as law" [wald-charter/HANDOFF.md §1]. The law as signed pages with reference checkers in exact fractions and a kit that judges any implementation; a builder "that cannot touch either"; a cage of CI and lock; an arena.
What it contributed. Single exit; housed numerals — nothing but 0 and 1 outside a declared table, which is jack's number lint. WORLD_FALSIFIED as a first-class outcome. The marking of judgement: wald's charter labels what is judgement (J1–J10) against what is derivation, and every jack brief's "decided (not asked), p = …" descends from it — including this book's plan.md §6. Charters versioned with attack prompts, and the discipline adopted on 24 September after "two v0.2 pages took fifteen and eight drafts … to finding problems a machine could have found": every rule stated once with an ID, every refusal naming its rule, every poison dying by its declared rule, a gate that must pass before any attack or signing [wald-charter/HANDOFF.md §3]. And the arena, the direct ancestor of the election: wald-arena's brief 002 was, in the cold read's words, the only pre-stated completion criterion in the corpus "met as written, against money ($38.40) and an external board, with held/missed printed", and it reported a baseline that won — "the better single model +56.7 against wald +36.0" [PREDECESSORS-20261026.md, wald-arena]. A court that prints its own loss is the thing jack was built to be.
The charter repository's log is a diary of the author role: eighty commits in twelve days, all by the owner, most titled "HANDOFF: state as of …" — the record kept by hand that jack keeps as entries. Two subjects name the inheritance outright: "CHARTER v0.2 draft 4: … credence's failures as poisons" (23 September) and "HANDOFF: gate-first lessons from session 4" (25 September) [data/lineage-commits.md]; the arena's verdicts were "read as measurements" on 30 September and "nothing in them bears on the law" the next day — the first time in the lineage that a result was allowed to be about the world and not about the text.
Where it stopped. A language with an arena but no court: the owner was still the referee, ruling every judgement mark, signing every tag, opening the cage's door [wald-charter/HANDOFF.md §2]. Its handoff names what it took from credence and what credence lacked; jack's charter §1 names what it took from assize; the chain of founding texts is unbroken.
[G] The notes hold only the cage door — "a push with the door shut fails … expected, not an error to work around" — which is the owner as referee. Whether "an arena needs a court" is the right one line is his to say.
14.5 life-agent (May–October 2026): derive, never decide
Read through the cold read only. The personal agent: a knowledge base that derives and an agent that decides; a revealed-preference stream; the owner's channel. Its contribution to jack is the split of derivation from decision, the owner's reactions as the cost model, and the notification seam the court later used. Its exit test — calendar days carrying live traffic — "has not started counting"; its archive holds sixty-seven thousand lines in 246 files; its rulings file records that "asking the owner cost less than searching" [PREDECESSORS-20261026.md, life-agent]. Retrieval, not utility, was the binding constraint, and the agent had no institution in which to earn powers.
14.6 assize (6–27 October 2026): the price of the sovereign
What it was. The first court, for coding agents: Builder, Reviewer, Auditor, Vizier by account; a forecasts ledger with log-loss and base rates; pins with mutants in CI; entrenched paths; a judge sealed as a container image. Its "ground" — append-only ledgers, claims before scoring, outcomes by code, no self-judgement, the signature binds its text, the price recorded, models elected not named — is nearly the whole of jack's constitutional vocabulary [claude/guy-acts-6-13oct.csv, instruction #54, 10 October].
What it contributed. The vocabulary, and more important, the measurements. The charter's §1 is a table of what assize cost: 88 decisions and 121 repairs in four days, 5.0 repairs per merge, twenty of twenty-four merges about the court, each recurring failure a broken maxim — F1 the accused writing its own test, F2 a retroactive rule, F3 stalls found by a human, F4 review without outcomes, F5 the reversible gated and the irreversible not, F6 counsel unscored, F7 procedure over object [CHARTER §1]. Those seven failures are most of chapter 11's witnesses, and the four maxims of chapter 4 §4.5 are the seven failures stated as what they broke. The weekly acts series continued after jack's charter was signed and is the baseline p₀ against which C1b's liveness was tested [claude/acts-weekly.md; BRIEF-C1b §Liveness].
The sessions on steel show the price from the inside [data/lineage-sessions.md, assize]. Of the owner's 95 short turns in 24 days, twenty-two concern a token, a personal access token, a keyring or a secret; eighteen a runner or a service to restart; seven a status report sent to his phone; two ask whether an issue needs him or will close by itself. The builder prompts of 6–7 October already carried the rule that became the liveness statistic — a session that asks the owner something it could have decided and recorded has failed — and on 24 October, drafting the wish that ended the court, he refused to loosen a pre-committed bound on his own acts, on the ground that a bound missed by one is a recorded fact about the price of the sovereign and loosening it to make the forecast safer was the habit the court existed to break.
Where it stopped, and why that was the point. assize ran from 6 October to its seal, reached the seal of its judge — a Crafter verdict under the sealed image, NetHack carried — and was frozen there by the sovereign's own wish [assize REPORT.md, L2 line; claude/guy-acts-21-27oct-assize.csv, #341]. It did what it was for: it measured the price of a court whose operator is in the loop, and it became jack's control group and judge of last resort. Its last pull requests were closed with the note "L3 is in renavondata/jack under Amendment 6" (assize's sixth amendment) [the same CSV, 27 October]. It is the only predecessor whose reason for stopping is written in its own record as well as in its successor's.
14.7 jack: the fixed point
| project | what it was | what jack takes from it | where it stopped, and the lesson |
|---|---|---|---|
| credence | the axioms as a language and apps | the owner's utility as the agent's; the strange loop bounded; closed-loop first | architecture without an institution; a score with no ground truth |
| proplang | the language, and the programme | obligations, pins that kill, no unwitnessed claims, append-only corrections, minimality as correctness | law in prose, no priced world, no competition; three termini |
| tannen | a pipeline kernel and the machinery to enter it | separation by session; decisions as data; poison; signed custody; the border | prose law can be satisfied by a pipeline that computes nothing; the owner measured as too slow |
| wald | the law as signed pages; the cage; the arena | single exit; housed numerals; judgement marked as judgement; charters with attacks; a scoreboard that prints its own loss | an arena with no court; the owner the referee |
| life-agent | the personal agent | derive, never decide; the owner's reactions as the cost model | no institution to earn powers in |
| assize | the first court | the vocabulary and the measurements; F1–F7 | frozen at its seal, deliberately; the control group |
| jack | the small core | — | running; its own forecasts about itself are on its ledger |
Two things to draw out. First, the lineage shortens and sharpens: credence five and a half months, life-agent four and a half, proplang eleven weeks, tannen five, each wald repository two, assize three, and each with a smaller law than the last and a wider competition — one model, then a builder under a cage, then an arena of models against a board, then a court that elects them [PREDECESSORS-20261026.md, synthesis; chapter 11 §11.11]. Second, every methodological rule now in jack was written first in one of them — the pins in proplang, the poison in tannen, the housed numerals in wald, the ground in assize — and that convergence is the chapter's evidence. An author who arrives at the same rule from four different starting points has found something about the problem, not about himself.
14.8 The lesson the sequence of measurements does not capture
Reading the sessions beside the repositories shows one thing the chapter's frame — each project finding the next necessity by hitting it — understates. In every predecessor the governance was right about the artifact and blind about itself. proplang's deletion audit policed terminals and then types, and never the protocol — every incident minted a clause and none retired. tannen's laws were frozen and its guards poisoned, and its own cost went unmeasured until the owner said "too slow" and the session counted eighteen minutes of hooks per commit. credence's engine refused every bespoke corrector and its router "learned a reliability with no ground truth". wald caged the builder and left the owner as the uncaged referee. Each law held its subject to a standard it did not hold itself to, and each owner was the one party outside the record.
jack's two additions over every predecessor are exactly those two: the law is under the same test as the artifact — pins that must kill, amendments that price and bet on themselves, a ratio that stops the court making court, lapse for rules no case defends — and the owner is inside the record, typed, counted and corrected where he forecasts, with his wishes alone unscored. The sessions also show why it took seven attempts, in the tannen note quoted above: writing the discipline down does not install it; only a mechanism that forces the failure to be seen changes the catch rate. That is Fuller's congruence as an empirical finding, and it is the reason this book's law is code.
14.9 What would show this chapter wrong
A predecessor that reached a litigant. If any of the seven ran a scored loop on a world that charged, for longer than jack has, at a lower price to its operator, then "the science was done and the engineering wasn't" is false and the lineage is a sequence of abandonments, not measurements. The cold read's finding is that only wald-arena's brief 002 was met as written against money, and it reported its baseline winning; the author's re-check of the clones found nothing to add. [G] is the one source the author cannot read: what each project taught that is not in its repository.
The frame wrong. "A sequence of measurements, each project finding the next necessity by hitting it" is the author's reading of the founding texts' chain. If the owner's account is that the projects were abandoned for reasons the successors do not state — fatigue, a change of mind, an opportunity elsewhere — then the chain is a rationalisation, and the chapter must say so.
The convergence coincidental. Four attempts converging on the same rules is evidence that the rules are forced only if the attempts were independent enough; one author carrying his habits forward would produce the same convergence for the wrong reason. The best answer the record gives is that the rules were each time forced by a failure the repository records, not carried in by hand — proplang's write-up says so in terms — but a referee is entitled to weigh it.
Draft note. Draft 1, the author session, 3 November 2026. Evidence of record: renavondata/jack/transcripts/PREDECESSORS-20261026.md (a cold read of the seven predecessors by a session of counsel's family, quotations verbatim, nothing executed), and the owner's own words from the sessions and memory notes on his machine (data/lineage-sessions.md), re-checked by the author against clones of gfrmin/credence (b10da64), gfrmin/proplang (e313733), gfrmin/tannen (1a1f903), gfrmin/wald-charter (bedffdd), gfrmin/wald (13addd6), and renavondata/assize (ff41aaa); life-agent, wald-arena and credence-governor are read through the cold read only. The marks [G] are where only Guy can say, and are the open question G of plan.md.