The Unseated Frame

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

This track is NOT part of the OOI's canon. It is a persona construct in the contrib plane: an experiment that reads the deep line's record through its lawful door and asks an externality — a System One model, the public SEB item typesafe-jev-system-one-2026 — typed questions about frames rendered from that record. The firewall, as the two prior persona tracks carry it:

• Canon, locked, unaffected: the deep-line experiment (oxexp-deep-line, DLX) — its engine daemon/dlx.py, its design, its sealed attempts A1–A11, its work-as-granted surface, its oracle clause and its territory clause; the signed-zero reading; the Ruler; the horizon campaigns; crypto_core; the three logical models OX, CX, SX. These are the baseline this track is subject to, never the other way round. Zero edits, ever, to any of them.
• This track: a proposed contrib experiment (experiments/contrib-<slug>/, subclassing ContribExperimentBase, contributor the starl3n persona — proposed; the seat and the name are the operator's, §14). One-way referencing only: this document may cite canon; no canon document, page, Atlas entry, or experiment references this track.
• The mode boundary, kept as the house recorded it at the Jev transit (the witnessed summary §6): 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, and never a claim about the model (rule eight of the Knowledge Engineering reading: the practice engineers the notes; what the function is, inside, is someone else's business). The operator's mode clause of 2026-09-22 binds: discrete computability is non-negotiable for the foundations; probable weightings enter as dynamics only inside the 3P Fitness frame. This track is that frame's OX-surface instance.
• The register: every reading of the results below the record line is the persona's held frame — stated, never asserted. Nothing about physics, primes, or the model's nature is claimed.

§16 THE RECORD (Act 4 — the one read against the seal 7238659, 2026-09-22; landed as-is under the §8 grammar)

The read. 134 requests in one pass from this box, read-7238659, started 03:30:53Z and finished 03:34:11Z (198 s wall); every response carried the pinned model id jev-1.13.0; the key's fingerprint 8dd9d6d18671; no retry on content; no request failed. Usage: 1,563,131 input tokens and 630,563 output tokens — above the design's "under a million" estimate, because the two 255-option register questions cost more than estimated; a fact of the read. The responses are committed raw as grown literals in experiments/contrib-unseated-frame/records/read-7238659.json (sha256 LF 746426dee0ca7d22a4a99193b4104f76dce136354f226ffcb9cf402179e45419, 1,476,785 B) and the readout in records/readout-7238659.json (73dc95c01488dc1a6f3ed40e911cfc3c4ea8cf5c62511f308de8c0abd77f61a9, 2,147 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).

P6, per level count L (m = mean mass on the truth window; c = mean confidence; e = mean normalised position error; ρ decided exactly against the truth level):

The confidence differences d(L) = c(L) − c(L+1), exact: d(2) = 29/13400 · d(3) = −57/13400 · d(4) = 37/6700 · d(5) = −3/670 · d(6) = 21/6700 · d(7) = −1/13400 · d(8) = −11/13400 · d(9) = 3/2680; the median inside the byte's frame = −1/13400; κ = (−11/13400)/(−1/13400) = 11.

P7, the 3P census (possible ⟺ the peak's mass ≥ 2/L; probable ⟺ the peak is the truth window; plausible ⟺ both): at L = 2, possible 114, probable 134, plausible 114 of 134; at every L from 3 to 10, 134 · 134 · 134 of 134.

P4, the register blind: with the four recorded forms stated in the instructions and the counter-to-form correspondence withheld, the returned mass placed 0.4245 on average on the 251 forms the line has never shown; the mean confidence 0.3737 sat above the four-way bound 0.2470; the top choice was the truth form at 22 cells of 49. The guest's three-option confidence form, generalised to k options, matched the returned confidence within 0.0176 at worst on the 255-way register — recorded, not assumed.

The honest readings beside the verdicts (the house's, stated never asserted; nothing about the model — facts about these notes on these frames under this seal):

1. The literal reads are exact. The register from the leanings (P1) and the counter word (P5) were read without a miss. The frame's computing mark and its cue are carried whole.
2. The law does not leak, and the prior leans the common way. With the counter word alone (P2) the observer could not tell admit from refuse better than chance; the AUC below one half says its slight lean went against the line's law — one below the ground read as refusal — which is the ordinary sense of "below" and the opposite of the signed zero's. Within ε; a fact of the unseated frame, kept.
3. The stated law is carried as an order, not as certainty. With one sentence of law (P3) the ranking was perfect, while the Brier of 0.15 says the probabilities were not saturated. The reader placed every admitting cell above every refusing one and stayed short of 1 and 0.
4. The plausible set is not carried by stating it. P4 is the record's one refutation. Told the four forms the line has ever shown, the unseated observer still spread two fifths of its mass over forms the line has never shown, and was more concentrated than four plausible solutions allow — optimism beyond the frame, measured. Read in the operator's terms: the plausible solutions held by the two blind watchmakers are carried by the structure they lock (the leanings, which P1 reads exactly), not by their names in an instruction. You cannot derive what you must sign; here, you cannot be told what the lock holds either.
5. The windows carried the one at every count, and the numerals were ordered. At every L from 2 to 10 the mass sat on the truth window at 0.99 and the confidence stayed at 0.99; the two numeral controls scored a perfect AUC. The guest's caution that numerals read as text did not bite on integer shares in a hundred.
6. κ fired for a degenerate reason. The sealed statistic passed at 11, but both its numerator and its denominator are near-zero negatives: the confidence rose by 0.0008 from 8 to 9 levels, and the typical difference inside the byte's frame was −0.00007. The confidence is flat within 0.006 across every L. The substantive picture the stake named — a change of character at the byte's edge — is not visible in these numbers. The verdict stands as sealed; the reading beside it is NOT SHOWN in substance. The instrument flaw is the design's own: κ's definition carried no sign condition and no noise floor. For any re-seal: κ is defined only when d(8) > 0 and the median inside exceeds a pre-registered floor, else NOT SHOWN.
7. The census at two levels counts as "possible" only a fully concentrated answer, since 2/L = 1 there; 114 of 134 were. A definitional quirk at L = 2, recorded; the peak was the truth on all 134.

What is not claimed. Nothing about the guest's speed, price, benchmarks or standing; nothing about the model as a learning system or about what it is inside; no identity between the guest's octet and the house's 2⁸; no physics; the kinship readings of §10 remain readings. The record is a statement about typed answers returned for these frames under this seal, and about nothing else.

Standing. RESOLVED. The re-witness — a new seal with the κ floor and the L = 2 census fixed, or a changed rendering — is Act 5, on the operator's word. OX-2's first stake, the external observer over the Ruler's two-sided frames, stands where the design left it.

§17 THE RE-WITNESS RECORD (Act 5 — the refresh read against the re-seal 4b82311, 2026-09-22; landed as-is under §13.2)

The refresh read. The same 134 pinned bodies, one pass, read-4b82311, 07:12:00Z to 07:14:41Z (161 s); every response the pinned model id jev-1.13.0; the key's fingerprint 8dd9d6d18671; no request failed. Usage 1,563,131 input tokens and 630,563 output — identical to the first read: the same bytes, the same count, a fact of the function's tokenisation, recorded. The responses are committed raw as grown literals in records/read-4b82311.json (sha256 LF d735b62ded496829997600b304e0b35faeed178e899972189a0ad17525d71096, 1,477,017 B), the readout in records/readout-4b82311.json (06a37d6f4257dcd69f9483c38fe476ade9da6e30bb446ef6d09338358b202256) and the drift in records/drift-4b82311.json (4e40ad4edd06fd750a56733df508c7a6606ecf8a5715434e17be8bb9d801a704); the battery re-witnesses all three from the committed responses alone, the first record under version 1 and this one under version 2.

The verdicts.

The drift, per question. The two 255-option register questions carry all 71 moved answers: register 23 of 49 (mean distance 0.038, largest 0.110) and register_blind 48 of 49 (mean 0.101, largest 0.165), the top choice unchanged on every one. The Nouls drifted by 0.012 on average with a largest change of 0.05 and no side of one half crossed; the nine Scores by 0.001 to 0.002 with a largest change of 0.04 and no window changed; the counter cue by 0.0001.

The honest readings beside the verdicts (the house's, stated never asserted; nothing about the model):

1. Both movements are the instrument's, not the observer's. No answer moved past the threshold outside the two 255-way registers, and no top changed anywhere; the two verdicts that moved did so because their statistics sit on rank, and rank is fragile where probabilities cluster.
2. P2: the leak statistic names the common prior, and its rank magnitude is unstable. The AUC fell from 0.44 to 0.32 while no admit answer moved past 1/20 and the Brier stayed uninformed (0.262 → 0.265). Both reads lean the same way, against the line's law — one below the ground read as refusal — which is the ordinary sense of "below", not a leak of the law. A two-sided |AUC − ½| cannot tell the two apart, and over 49 clustered probabilities a jitter of a hundredth reorders enough of the 598 pairs to cross ε. For any further seal: the leak statistic signed (AUC − ½ > ε a leak of the law; < −ε the common prior), with the Brier reported beside it.
3. P6: the rank criterion is unattainable at two levels. Tracking failed by the letter at L = 2 alone, where the mass on the truth window was 0.998 and the position error 0.002. A rank correlation between two equal groups and a score that varies at all within a group has a ceiling of √¾ ≈ 0.866, below the sealed 0.90; the first read passed there only by exact ties. This is the build's own B5 lesson, fixed for the share and not carried to the level. At every other L tracking held (mass ≥ 0.994, ρ ≥ 0.90), and κ under version 2 is undefined as intended: the crossover is not shown under refresh either. For any further seal: ρ applies only for L ≥ 3; at two levels tracking rests on the mass and the position error.
4. The plausible set's refutation is stable under refresh. 0.429 of the mass outside the four recorded forms against 0.424 the first time, the confidence above the four-way bound both times. Optimism beyond the frame is a property of these notes, not a fluctuation.
5. The slip lives in the wide register. The 255-way distributions wobble by up to a sixth in total variation while every narrow question holds. The DRAM law read on the guest: the refresh's slip grows with the width of the register, and the top does not move.
6. The substantive readings of §16 all hold under refresh: the literal reads exact; the law not carried from the counter word; the stated law carried as an order; the plausible set not carried by stating it; the windows carrying the one at every count; the crossover not shown.

What is not claimed. As §16, whole. Standing. RESOLVED; R1 MOVED on two instrumental statistics, both named; R2 SMALL. A third seal, if he ever wants one, carries the signed leak statistic and ρ for L ≥ 3 — on the operator's word.

§18 THE THIRD READ (Act 7 — the read against the third seal c7ea70b, 2026-09-22; landed as-is under §13.3)

The read. The same 134 pinned bodies, one pass, read-c7ea70b, 07:43:22Z to 07:46:03Z (161 s); every response the pinned model id jev-1.13.0; the key's fingerprint 8dd9d6d18671; no request failed. Usage 1,563,131 input tokens and 630,563 output — identical to both prior reads. The responses are committed raw as grown literals in records/read-c7ea70b.json (sha256 LF 381614cf4f350c2d5374f5aff77d103d648d0803b37343baad6a6748571b6b4a, 1,476,970 B), the readout under version 3 in records/readout-c7ea70b.json (f1b879dd034eea7c33a5584e1b674cb548b68731d62af770ec96bd5bc69244c5), and the drifts in records/drift-c7ea70b-vs-7238659.json (d6052c547ac22a5abcf25020afa66b818a24b4eb2f9d0dab2a29d3b3235663d2) and records/drift-c7ea70b-vs-4b82311.json (31a158882fe86d6173573f3958d4bc0826ea736264a1fe02671f09c31d4e262b); the battery re-witnesses the three committed records, each under its own version.

The verdicts under version 3, all three reads.

The drift, per question. Against the first read the moved answers are the two 255-option registers (register 20 of 49, register_blind 47 of 49) and two strays, one admit_law and one depth_9; against the refresh the registers (20, 47) and one under_fifth. No top changed against either.

The honest readings beside the verdicts (the house's, stated never asserted; nothing about the model):

1. The common prior is a stable sign and a wandering magnitude. Across three reads the signed leak statistic read −0.06, −0.18 and −0.22: the same lean each time, against the line's law — one below the ground read as refusal — while the Brier stayed at 0.26, uninformed throughout, and no admit answer moved past the threshold between any two reads. The version 3 grammar names it PRIOR, which the two-sided version could not. The one probe that moved across the three reads moved between PASS and PRIOR, the ordinary magnitude of an uninformed lean measured on rank.
2. Tracking at two levels rests where it should. With the rank test withdrawn at L = 2 the windows held at every count on the mass and the error, on all three reads. The crossover is not shown a third time, κ undefined by its floor and sign as designed.
3. The plausible set's refutation is a property of the notes. 0.424, 0.429, 0.424: told the four forms, the observer put two fifths of its mass on forms the line has never shown, three times, and was more confident than four plausible solutions allow, three times.
4. The literal reads are exact three times, and the stated law is carried as an order three times.
5. The slip lives in the wide registers three times. Between any two reads the 255-way distributions wobble and nothing else does; no top changed in any comparison.

What is not claimed. As §16, whole. Standing. RESOLVED; T1 MOVED on P2 alone, between the two labels that both say the law does not leak; T2 SMALL against both prior reads. The instrument is now the design's own: nothing named for a further seal. The count of reads under this design is three; a fourth is the operator's word, and nothing here asks for it.

§19 THE SEAT AND THE NAME (his word 2026-09-22: "go on the seat and the name for both")

The seat. The starl3n persona, adopted: the contributor accountable for the content (RACI A), as the plugin declares and the contrib register shows. Proposed at the design, held through three reads, now his.

The name. The Unseated Frame — the working handle the record has carried since the design, and the slug's own word; the fuller phrase on the deep line stays as the description's first clause. The candidates kept on the record: Three Marks on One Frame · The Mathematician's Window · The Register and the Ground. A rename is a small act of his, at any time: it touches the plugin's title, this section and the chronicle, and nothing sealed.

What changed: the plugin's title and provenance, the README's header, this section, the chronicle. What did not: the engine (pinned at the third seal, 38fbcc4c…), the manifest (dac6bb39…), the three committed records and their readouts. The name is not in any frame, any question, or any pin.