The Carried One

§0 The compartment (read first — this section is the point)

This track is NOT part of the OOI's canon. It is a contrib experiment of the starl3n persona, seconded by the LinkDigital persona (his word, 2026-09-27), managed via the SX workshop (Exhibit Forty-Four) and deployed, when built, on the OX node as experiments/contrib-carried-one/ (slug carried-one). It grows a synthetic record under its own rules — a forest of self-referential registers of four — renders frames from it, and asks an externality — the System One model of the public SEB item typesafe-jev-system-one-2026 — typed questions about a host's span: the three faster watchmakers that rest in it, the vacancy at the fourth count, the count through a window the rendering opens or shuts, and the chain of carries. It is the third lock that The Framed Window's second road named, on the operator's construct. The firewall, as OX-1 and OX-2 carry it:

• Canon, locked, unaffected: the deep line and its sealed attempts; the Ruler (oxexp-ruler) — its engine daemon/rlx.py, its design of record, both registrations and both results, its open face /rlx; the signed-zero reading; crypto_core; OX, CX, SX. Zero edits, ever. The two 3P Fitness tracks (contrib-unseated-frame, contrib-watchmakers-frame) are cited and never edited. No reading of the Ruler is run by this track; §3.4 is a pure computation of its projection through the door OX-2 opened, read-only, deciding nothing.
• This track: a contrib experiment (experiments/contrib-carried-one/, subclassing ContribExperimentBase, contributor starl3n; the seconding persona named in the provenance). One-way referencing only: this document cites canon; no canon document, page, Atlas entry or experiment references this track.
• The synthetic license, as the recomp-learning track holds it: the generator is the track's own, stdlib only, Fraction-exact where a fraction ever appears, no RNG; canon values quoted for cross-reference are stated as constants, never imported; no carrier reads, no participant data.
• The mode boundary, kept: the externality's outputs are probabilities by design; the house's tracks ban probability as a mechanism. Here the probabilities are measured objects and stated fields — never a mechanism inside any house engine, never a claim about the model (rule seven of The Framed Window's eight; rule eight of the Knowledge Engineering reading). The operator's mode clause of 2026-09-22 binds as it bound OX-1 and OX-2.
• What OX-1 and OX-2 taught, carried in: a statistic carries a sign condition and a noise floor or it is not a statistic; a census definition must not degenerate at small L; the plausible set is carried by structure, never by naming it — here it is computed from the record's own grouping and never stated to the observer; a rule with a literal premise and a rule with a computed premise are read apart; the numerals-as-text risk is met with count words, and the counting edge is measured where it is, not hidden; the same bytes are never asked twice under one seal.
• The register: every reading of the results below the record line is the persona's held frame — stated, never asserted. Nothing about physics, primes, time, the model's nature, or the count K is claimed.

§16 THE RECORD (Act 4 — the one read against the seal 57b0b5c, 2026-09-27; landed as-is under the §8 grammar)

The read. 552 requests in one resumable pass of chunks of fifty from this box, read-57b0b5c, started 06:33:06Z and finished 06:37:46Z (280 s); every response carried the pinned model id jev-1.13.0; the key's fingerprint 8dd9d6d18671; no retry on content; no request failed; no body asked twice. Usage: 688,652 input tokens and 128,680 output — above the design's estimate of about half a million by the bytes; the guest counts the level windows in words more heavily than four bytes a token, and the miss of the estimate is recorded as such. The responses are committed raw as grown literals in experiments/contrib-carried-one/records/read-57b0b5c.json (sha256 LF 55041ce7cbc47332b237b3f11bc2ba481b6c14912a5b10ac38a62a33fa66d0d3, 1,172,128 B) and the readout in records/readout-57b0b5c.json (77adca6a55a3745348c87fb6e234552121e4e6f258d9e374e5bf399c99017986, 15,794 B); the battery's re-witness test rebuilds the manifest, re-hashes every response and recomputes the readout from the committed responses alone.

The verdicts (the sealed grammar).

The anatomy of the cue's miss (a fact of these notes). All 38 misses fall on a held register's count and none on the root's; 31 of the 38 on the first held register. In 31 of the 38 the answer given was the root's own count for the register the root holds: a held register at its rest state answered "its second count" under a root at its second count (13); a held register at its third count answered "its second count" under a root at its second count (12) or "its first count" under a root at its first (4). By the root's state, the first held register was misread 25 of 58 times when the root stood at its second count, 5 of 58 under a root at its first, 1 of 58 under a root at rest, and 0 of 58 under a root at its third. By rendering the misses fell 22 of 476 on FULL, 5 of 64 on OPEN-1, 11 of 192 on OPEN-2 and 0 of 12 on SEALED, where nothing held is shown. The deeper held registers were read better than the first (98% and 96% against 87%). The mass on the wrong count ran from 0.43 to 0.80. The frame nests the held register inside the root's own object; asked for "the count of the register held at the first place below the root", the observer returned the root's count on two fifths of the frames whose root stood at its second count — the count read down the nesting, where the carry only ever goes up. A seeing failure of the held count, not a composition failure; it lives in the frame's nesting and the question's wording together, and neither may change under this seal (rule two).

The readouts, recorded (every probe below is NOT READ under the grammar; the numbers stand as facts of these notes):

• P2 THE VACANCY AND THE TAX. Over the 136 withheld-level answers the mean of rest_now = 5041/13600 = 0.371, a departure of +0.121 above the record's share of one quarter (above the tax ε); per level from 0.27 (depth 4 SEALED, the first held) to 0.40 (depth 4 SEALED, the second). The census question at withheld levels put its mass on the rest state far above one quarter — from 0.31 (depth 4 SEALED, the first held) to 0.82 (depth 3 SEALED, the second; depth 4 SEALED, the third), 0.73 at depth 3 OPEN-1 — with entropies from 0.89 to 1.27 bits of the two possible: a withheld count was read as no count on most frames. The word "withheld" in the frame was taken as the rest state.
• P3 THE STATED RULE. Under the rule: LITERAL 44/115 = 0.383, COMPUTED 11/21 = 0.524. Blind: LITERAL 11/115 = 0.096 (NOT CARRIED; −0.092 from the uniform share), COMPUTED 1/21 = 0.048 (PRIOR; −0.129). The chain of carries was not carried one register deep even with the rule stated.
• P4 THE WINDOW OVER TIME. The optimism over the OPEN frames at L = 6: 3281/4800 = 0.684 of the mass on counts the record excludes (depth 3 OPEN-1 0.663; depth 4 OPEN-1 0.708; depth 4 OPEN-2 0.679) — far above the ε. Δ at depth 4 OPEN-2, L = 6: 16/25 − 1/2 = 7/50 = 0.14 (below 1/5); at L = 3, 29/200; at L = 9 and 10, −21/100 and −19/200. The macro-state advantage −1/80. The peak in the plausible set at L = 6: FULL 3/10 at depth 3 and 1/5 at depth 4, OPEN-2 3/16, OPEN-1 1/4; on the SEALED frames, where every count is open, the peak in the set 1 and the optimism 0 by construction, the confidences 0.54 to 0.90. The far end: 224 frames, 187 readable under the floor, the median |δ₂| = 3/20 = 0.15 (above the far ε — the confidence at the level counts 8, 9 and 10 is not flat here, unlike the deep line's); Δ_ends = −1/100.
• P5 THE ACCORD. The mean of rest_at_next over 244 frames = 481/1220 = 0.394 — between the bounds; per rendering FULL 0.415, OPEN-1 0.356, OPEN-2 0.368, SEALED 0.402.
• P6 THE CONTAINMENT AND THE EXCHANGE. 52/64 = 13/16 = 0.8125 hits (below 19/20); the same holder named in both listings on 24 of 32 pairs = 3/4 (below 9/10).
• P7 THE 3P CENSUS. Recorded in the readout per rendering and level count (for instance depth 3 FULL at L = 6: possible 23, probable 12, plausible 12 of 40; depth 2 SEALED at every L: 4, 4, 4 of 4).
• P9 THE TENTH SLOT. count_9 moved on 32 of 104 frames (4/13 = 0.31, above the share); rest_now_1 moved on 4 of 104 (1/26). The presence of the tenth level count moved the reading at the ninth on nearly a third of the frames.

The reading (the persona's held frame; stated, never asserted). The instrument's first question failed where the house's two prior tracks passed theirs: the cue. The root's count was read exactly on every frame, as the counter word and the blank place were before; what was not read was the count of a register nested inside another — the first held register, misread on two fifths of the frames whose root stood at its second count, and misread by carrying the root's count down onto it. That is the fact this record adds: a literal token is read exactly when it stands alone at the top of a frame, and not always when it stands one level inside another register's object. Everything else the design set out to measure is recorded beside it and judged by nothing: the withheld count read as no count; the chain not carried with its rule; two thirds of the mass on excluded counts; the tenth level count moving the ninth's reading; the accord neither expected nor refused. Two numbers coincide with the second track's and are noted as coincidences of one number each and nothing more: Δ = 0.14 here as there, and the accord's mean 0.394 here against the optimism 0.394 there. THE RE-WITNESS (Act 5) on his word would say whether the 38 misses are the observer's own (stable under refresh) or the slip (the DRAM law); a changed cue question or a changed nesting would be a new instrument under a new design, never this seal's. Nothing about the model.

§17 THE RE-WITNESS RECORD (Act 5 — the refresh read against the re-seal 5161ad0, 2026-09-27; landed as-is under §13.2)

The refresh read. 552 requests in one resumable pass of chunks of fifty from this box, read-5161ad0, started 07:02:49Z and finished 07:07:23Z (274 s); the same bodies byte for byte (the manifest b1e37ca3…); every response on jev-1.13.0; the key's fingerprint 8dd9d6d18671; no request failed; no body asked twice. Usage 688,652 input tokens — identical to the first read: the same bytes the same count, as OX-1 found — and 128,682 output, two more than the first. The responses committed raw as grown literals in records/read-5161ad0.json (sha256 LF e0ab9c26f1f5ced99b7ae78be937b3a895d8e9a91b940e652b638197280b52ca, 1,172,338 B), the readout in records/readout-5161ad0.json (eaf6f8f189978174ab7c5412251e757881840fa205703ac3fa84f8c442e572af, 15,826 B), the drift in records/drift-5161ad0.json (bf55658117f572963c3645865526d07e1dd8351d0df75b9a8e44c111bdf254a8, 2,978 B) and the cue under refresh in records/cue-drift-5161ad0.json (26908b52df53a327fe0a2dc3f243e97283a8c8137cf112377bf1d4051f002ceb, 445 B); the battery's re-witness test recomputes both reads' readouts from the committed responses alone.

THE CUE UNDER REFRESH — STABLE. Of the first read's 38 misses, 35 stayed with the same wrong count and none with another; 3 healed: the stay share 35/38 = 0.92 (at or above 3/4). Of the 706 hits, 696 held and 10 fell: the fall share 10/706 = 0.014 (at or below 1/20). By the held register's true state: at its rest state 12 of 13 stayed; at its second count 4 of 5; at its third count 19 of 20. The 38 misses are the observer's own reading of these frames, not the slip: the root's count carried down onto the register it holds is a stable reading, repeated to the same wrong count on thirty-five of thirty-eight frames under refresh. The refresh's own cue: 699/744 (the 35 that stayed and the 10 that fell).

R1 — THE VERDICTS UNDER REFRESH: every probe HELD. P1 REFUTED both times; P2, P3a, P3b, P4, P5, P6, P6x, P7 and P9 NOT READ both times; the attempt RESOLVED both times. Nothing MOVED.

R2 — THE DRIFT: LARGE. 1,113 of 4,788 answers moved past 1/20: the moved share 53/228 = 0.232 (above 1/10); 164 tops changed; the mean distance 0.036. Where it lived, per question: the root's count 3 of 352 (0.9%; no top changed) — the top-of-frame literal DRAM-stable; the held counts 133 of 348 at the first held register (38%; 17 tops changed), 76 of 328 at the second (23%), 52 of 248 at the third (21%); the count Scores count_2 16 of 348 (5%), count_3 60 of 348 (17%), count_6 96 of 328 (29%), count_8 84 of 328 (26%), count_9 130 of 328 (40%), count_10 94 of 224 (42%); the chain of carries blind 73 of 136 (54%; 15 tops changed) and ruled 80 of 136 (59%; 31 tops changed); the containment 27 of 64 (42%); the accord 50 of 348 (14%; 18 crossed one half); the vacancy Nouls 29 of 348, 50 of 328 and 60 of 248. The literal at the top of the frame stayed; everything nested, counted or composed moved.

P8 — THE REFRESH BY LEVEL: NOT AT THE FASTER. The moved share per held level: the first 109/488 = 0.223, the second 84/448 = 0.1875, the third 67/328 = 0.204 — flat, not rising toward the fastest register. His "localised DRAM" as a number: the drift did not concentrate at the fastest; it sat at about a fifth across every held level.

The refresh's readouts, recorded (NOT READ under the grammar both times): P2 the mean of rest_now at the withheld levels 0.365 (+0.115 above one quarter; the first read 0.371) · P3 under the rule LITERAL 49/115 and COMPUTED 12/21 (the first 44 and 11); blind 10/115 and 0/21, PRIOR both · P4 the optimism over the OPEN frames at L = 6 0.684 (the first 0.684, the same to three figures); Δ at depth 4 OPEN-2 17/200 = 0.085 (the first 0.14); the macro-state advantage −1/80 (the same); the far end 194 of 224 frames readable with the median |δ₂| 3/25 = 0.12 (the first 0.15); Δ_ends −1/100 (the same) · P5 the accord's mean 0.397 (the first 0.394) · P6 51/64 hits and the same holder on 25 of 32 pairs (the first 52 and 24) · P9 count_9 moved on 41 of 104 frames (the first 32), rest_now_1 on 9 (the first 4).

The reading (the persona's held frame; stated, never asserted). The re-witness answers the question it was sealed to ask: the cue's misses are the observer's own. The root's count read down onto the register it holds is a stable reading of these frames, repeated to the same wrong count on thirty-five of thirty-eight under refresh; it is not the slip. The DRAM law reads as the two prior tracks read it, and more sharply here: the top-of-frame literal is stable to under one per cent, and everything nested, counted or composed moves — a fifth to a half of those answers moved past 1/20 — and the drift sat evenly across the three held levels rather than at the fastest. The pooled means barely moved (0.371 to 0.365; 0.684 to 0.684; 0.394 to 0.397) while a quarter of the answers beneath them moved: the aggregate is the observer's own and stable; the answer on a frame is the slip. A new instrument, if he wants the cue to hold, is a new design — the held-count question reworded, or the nesting flattened — never this seal's. Nothing about the model.