The surprise ledger for the cold-start map

[NGL321]

Part of #517.

This is a ledger, not a ticket. It is assigned to the user so the frontier query never hands it to an agent, it starts empty on purpose, and it stays open for the life of the map. Agents write rows here; only the user writes the ruling column.

A surprise is (a) a reading outside every branch in its ticket's table; (b) a pre-registered
denominator or ceiling that is degenerate on the surface (#496's readout 1); (c) T0's P1 failing; (d)
anything whose remedy would amend an ADR.

On a surprise the agent stops that ticket, appends one row here — what was read, which clause it broke,
the branch the agent proposes — and continues any other unblocked ticket. (b) is repaired by stating a
new convention and publishing both numbers, never by silently replacing one; the row is a report.

No agent mints a grilling ticket from this map. The user reads the ledger in one sitting and rules inline
or mints one grilling from several rows.

#	ticket	what was read	clause broken	proposed branch	ruling
1	T0	P2 fails. Excitation rank as specified (participation ratio of the evidence stream over 1,000 ticks) reads 1.00 for essentially every cell — every stream is DC-dominated, so uncentred it separates nothing (R² on log ρ 0.00–0.16 vs column identity's 0.13–0.32 at 20k); centred it orders apex 1.20 < vision 1.26 < core 1.52 but R² ≤ 0.09. log ‖ē‖, the persistent error's magnitude, reaches 0.41–0.44 on every seed. P1/P3/P4 held; the rule's update is rank-one along h in every cell (PR 1.01–1.06).	Ledger clause (a): P2 was pre-registered with no branch row. The map's the variable is per-cell excitation rank sentence, and T2's whether induced activity raises excitation rank read, rest on an instrument that does not discriminate on this surface.	Keep excitation.py and report both forms in the canonical table, but T2/T3 read the persistent error beside it — ‖ē‖ and its direction stability (apex 0.95–0.99 vs core 0.5–0.6) — as the coherence variable the mechanism actually names; induced activity is judged by whether it breaks the apex error's direction, not by whether it raises the stream's rank. Re-read P2's separation at T3's 100k, where column identity itself was 0.63–0.68. No ADR is touched.
2	T1	The map's shallow spec does not build. Guard 1's DomeSpec(vision_sides=(8,), somatomotor_sizes=(6,), core_sizes=(16, 8)) fails construction: with one vision lattice the 64 L1 cells land four to a core cell, the somatomotor six add a fifth to six of them, and every core cell carries two up-edges to the apex, so the core level sits at 6–7 vertical edges against core_degree = 6 and build_graph refuses the overshoot (no edge is ever removed).	Ledger clause (b): a pre-registered surface that is degenerate — here at construction, not on a reading.	Repaired by stating the convention and publishing both: T1 runs core_degree = 7, the smallest target that admits the taper's vertical edges, with the apex untouched (apex_degree = 4, one m = 1 drive edge, the same 19/13 private/exposed stalk split as the full dome's — the cell under study is the same object); the core comparator's guaranteed private dimension is 11 against the full dome's 14. 358 cells / 507 edges. Spec at prototypes/cold-start/T1/run.py::SHALLOW on map/cold-start; T2/T2b import it. Guard 1 amended on the map. Compute, measured: ~18 ticks/s, ~21 min per 30k seed, not the map's ~7. No ADR touched.
3	T1	#488's standing 1/3 is not on this surface. Unregistered read, added to see the ρ = 1 flag act: the relative disagreement `	d	/ (	a
4	T2	The retention guard's named classes are not where the damage lands. The guard is pre-registered as apex and soma ρ(K) not below the frozen baseline beyond spread. At T_b = 5000 condition C (sensory, large, shuffled) passes it — apex −0.044 against a spread of 0.059, soma +0.023 — while taking vision 0.977 → 0.779 (−0.198) and core 0.971 → 0.896 (−0.075). At T_b = 20000 the same condition fails the guard outright (apex −0.071, soma −0.167) and vision reaches 0.659 (−0.318). So the guard catches C at one horizon and not the other, and what it misses at the shorter one is the largest retention loss anywhere in the sweep.	Ledger clause (b): a pre-registered guard that is degenerate for the failure the surface actually shows — it names the two classes an apex ladder cares about and the sweep's damage is at the sensory rim.	Repaired by stating the convention and publishing both: the guard verdict is reported exactly as worded (apex + soma), and all four classes are published beside it on every row, so no condition passes quietly. Proposed for T2b and T3: read the guard over every class, since an induced supply enters at a wall and its damage is local to that wall — a guard chosen for an apex intervention does not transfer to a rim one. No ADR touched; the numbers for both readings are in prototypes/cold-start/T2/.
5	T2	Ordered loses to shuffled on priming, in both arms, and the branch table has no row for it. Composed rim-to-apex effective rank at +30k: B (ordered) sits below its shuffled control C by 0.0317 against a spread of 0.0148 at T_b = 5000 and by 0.0292 against 0.0241 at T_b = 20000; D sits below A by 0.0048 / 0.0069. The table carries B beats C and C beats A with B ≈ C, and neither describes this: C beats A (+0.0336 / 0.0148) but B ≉ C, so the amplitude is what mattered row does not fire either.	Ledger clause (a): a reading outside every branch in the ticket's table.	No action proposed; a reading offered rather than claimed. Composed rank falls from 1.2979 at construction to ~1.08 under every condition and to 1.0845 ± 0.0023 under a zero-supply control run for this purpose (A, B and D land within 0.007 of it; B within 0.0001), so the ordering among the four is which decayed least, not which built most — and only C sits apart (1.112–1.116), the one condition that also destroys retention (row 4). The consistent reading is that C's higher composed rank is a surface kept from converging by high-energy white-ish input rather than a channel laid down; on that reading the structure axis has not been shown to do anything and has not been shown to be inert either, because nothing was laid down under any member. Worth a fresh read only if a later rung makes any condition lay something down at all.
6	T3	Done-when (2)'s bar of 1.5 sits above where the full dome starts. Composed rim-to-apex effective rank (#436/#497's construction, 263 chains, seven hops) reads 1.025 at construction over the 256 patch chains before a tick is run (actuator 1.074, proprioceptive 1.080, touch 1.022), and 1.000 ± 0.000 at both 30k and 100k in both arms — the frozen baseline and T1's winner alike. The shallow dome started at 1.298 and fell to 1.078–1.116 (T2); the full dome starts at ~1.03 and converges to rank one. No surface read anywhere on this map has approached 1.5.	Ledger clauses (b) and (d): a pre-registered ceiling that cannot discriminate on the surface it is read on — nothing distinguishes the supply built no channel from this dome does not carry more than one direction rim-to-apex at all — and the bar is ADR-0009's own falsifier, so moving it is an ADR-level act.	No action taken; a report. Both numbers are published (median over chains, and the max beside it) and the clause is reported failed exactly as worded. Reading offered, not claimed: apex retention and rim-to-apex transport are separable — the winner lifts apex ρ 0.744 → 0.959 and moves the median composed rank not at all, which is the cleanest evidence on the map that clause (1) and clause (2) are different problems bound into one done-when. The only motion anywhere is the max chain, 1.429 → 1.641 under the winner against the baseline's 1.185 → 1.171 — one chain in 263, a hint about where to look and not a result. If the Destination re-read keeps clause (2), the bar wants a stated referent on this dome rather than an inherited 1.5.
7	T3	#488's standing 1/3 is not on the full dome either — the read row 3 asked for. Row 3 asked that the drive edges' relative disagreement `	d	/ (	a
8	B57 (#629)	The matched-generic null agrees about nothing at all, so the excess is the whole count rather than a residue of it. B52 (#622) built the null expecting the raw number to be mostly bookkeeping, netted out the way B50 (#618)'s forced fraction nets out on the transport side. Read on the same traffic with the phrasebooks redrawn generic, N_gen(theta) = 0.0000 at every one of 18 profile levels, on all three arms, at all nine checkpoints, in both the uncentered and centred forms, over three independent redraws. The real surface reads N(0.25) = 1.0000 from 1,000 taught ticks on, so the excess is +1.0000 where B52 predicted approximately 0.	#629's branch table, row 2: N ~ 1 everywhere, excess ~ 0 fires on the count and fails on the excess, and no other row describes it -- so the ledger's outside every branch clause.	No action proposed and the instrument is not repaired; a reading offered rather than claimed. Two readings are consistent with it and this ticket cannot separate them: either the traffic's single surviving direction is one the built maps genuinely share and generic ones do not, or the normalised quotient q_i is simply far from 1 wherever the traffic is rank one and near 1 otherwise, in which case the excess is measuring the rank collapse rather than agreement. The second would make the count a re-description of traffic rank, which is B60 (#633)'s object; it is worth one arm with the rule off past construction to tell them apart, and #629 did not run one.
9	B62 (#635)	The two learning rules are antagonistic on the traffic's rank. #635 asked which rule spends it, expecting a split of one movement. Run separately on the same dome, seed and window: PredictionRule alone reproduces the whole collapse — uncentered traffic ER 2.7798 → 1.0087 by 2,000 ticks against the both-on arm's 1.0132 — while TransportRule alone never collapses and pushes the other way, rising to 3.1837 at 1,000, peaking at 3.3307 and ending at 3.1347. They also split the other quantity cleanly and in the opposite direction: the prediction rule buys no agreement (leading direction stalls at q = 0.5495, A(0.25) = 0.0039) and the transport rule buys nearly all of it (q = 0.0636, better than both rules together manage at 0.1830). The frozen control holds both flat at 2.7812 and q = 0.8835.	No clause; #635's item 2 asked for the split and the branch table did not anticipate the two levers opposing each other. The map has spoken of the rule as one lever throughout — B25's the parameterisation's optimum is set by the objective and B27's refusal to split this map both treat it as one.	None proposed. Two prices ride with it and are reported rather than argued: PredictionRule collapses the rank at exactly zero cost in exposure and audience differentiation (bit-identical to the frozen arm at every rung, because it writes K and never touches a restriction map), and TransportRule raises differentiation (0.5307 → 0.5505) while spending 1.96 dimensions of exposure per cell. The both-on arm's differentiation cost (→ 0.4826) is neither rule's own. Measured, not explained: that one rule writes K and the other the maps is a location, not a mechanism. Nothing here is a candidate and no arm is scored.
10	B62 (#635)	ker(G) = ker(δ_P) exactly, so exact agreement is unavailable to any visible direction by construction. Read off the operators with no traffic in them: rank(G) = rank(δ_P) = 3000, both kernels 1800-dimensional, ‖G‖ = 1.4607, and δ_P restricted to ker(G) has top singular value 1.75e-15 — a relative 1.2e-15. Every direction the cells agree about exactly is a direction no incident map can see at all.	No clause broken; it confirms B48 (#615)'s the trained architecture's entire H⁰ is the privacy reserve from the operator side rather than the counting side, and it arrives at B52's case for a graded statistic independently — a bar written on earned asks for a measure-zero event and an invisible one.	None proposed. Filed because it is a structural fact about the shipped construction that no ticket had stated, and because it changes what a future exact agreement instrument could possibly read. It cost no run.
11	B62 (#635)	N(θ) has a vacuous regime: it reads 1.0000 on a direction carrying 0.4% of the traffic. N is a participation ratio taken inside the counted band — it says how many directions dominate among those below θ, not how much traffic they carry, which is A(θ). On this ticket's own bias-only arm at 5,000 ticks, N(0.25) = 1.0000 with A(0.25) = 0.004173, while the direction carrying 99.5% of the traffic sits at q = 0.5493. Read on N alone that arm looks like B57's trained baseline; read with A it is its opposite.	No clause broken and B57 (#629)'s conclusions are untouched — its A(0.25) runs 0.9547 → 0.9977 across the trained ladder, so its N = 1.0000 is one direction carrying essentially the whole traffic, exactly as it read it. B52 specified both and the profile has carried both all along; the readout led with N.	One standing constraint offered, not applied: an agreement count is quoted with the weight it counts — A(θ) beside N(θ) — or it is not a reading. It costs no run to comply. The instrument is not repaired here. A second, narrower note rides along and is not proposed as a constraint: B57's unseen guard is den > 1e-24 · max(den), relative to the largest denominator within the set being read, so it does not fire on a synthetic family whose denominators are uniformly tiny. Real traffic mixes visible and invisible directions, so no reading on the record is affected.

---

[NGL321]

Row 8 — proposed, for the table above. Filed from B50 (#618). It amends row 6, so it is written here rather than as a new ticket, and per this ledger's own rule no grilling is minted from it.

#	ticket	what was read	clause broken	proposed branch	ruling
8	B50 (#618)	Row 6's construction figure is a fact about a surface that no longer exists. #537's own read_surface, unmodified, on eedb781: composed rim-to-apex ER at construction reads 1.412 median / 2.869 max against row 6's 1.025 / 1.989, and the per-hop cosines read median 0.932 / max 1.0000 against #537's 0.312 / 0.834. Two commits post-date every figure on row 6: #548 (24e6c5f) deleted DomeSpec.interior_m and allocates interior lanes per edge — uniform m ∈ {3,4} became {4, 11, 12, 15, 16, 20} — and #562 (5ecaf3e) set k_v = n − p, pinning k_v at 20 at every interior hop against #537's 15–31 histogram. Neither was a change to the transport rule; both were width/mask decisions taken for other reasons. Row 6's trained figures (1.000 ± 0.000 at 30k and 100k, both arms) have not been re-read on the current surface and are not claimed stale here — only the construction figure is measured.	Ledger clause (a): a reading outside every branch in #618's table, whose every branch presupposes #537's construction table. Arguably (b) as well — row 6 was itself filed as a degenerate pre-registered ceiling, and the surface it was degenerate on has since moved.	Repaired by stating the convention and publishing both, never by replacing one: row 6's numbers stand as read on 0a22a1c and are re-labelled with that commit; the eedb781 numbers are published beside them. What is not proposed is any inference about the collapse — the construction figure moved 1.025 → 1.412, which is a long way below any bar and says nothing on its own about the 1.000 at 100k. The sharp open question, which needs a decision the agent should not take: does #532's collapse evidence survive #548 and #562 at the trained horizons, and does the map want its central readings re-taken? #532 quotes row 6 as "the strongest single statement of the problem" and as evidence "largely prior to learning" — that second clause is the one this reading touches. A re-read is ~30 min per arm on the existing rig (prototypes/cold-start/T4/, both T3 arms, seed 42), so the cost of settling it is small. No ADR is touched.

Beside this, and not a surprise — a repair, recorded because it explains why no T-rig had been re-run: t0.surface() has thrown AttributeError: 'DomeSpec' object has no attribute 'interior_m' since #548, so every T0–T4 entry point was dead on main for three days. Fixed on PR #627; it now records the allocation rule instead of a single width, plus lateral_m and drive_m.

[NGL321]

Two rows from B68 (#645), both outside its branch table.

Row — the arms that run TransportRule do not reproduce, and the rule is the amplifier rather than the source.

Two identical invocations of the transport arm inside one process part by ΔER 2.30e-02 at tick 100 and 2.47e-01 at tick 200. A run differing only by the construction-time matched-generic null is bit-identical to the first, so the null is exonerated; an identical command in a separate process reproduces exactly. The frozen control locates the source: with no rule at all the same triple parts by 2.60e-05 and 2.19e-04 at the same rungs — so the divergence is already present with nothing learning, and TransportRule amplifies it by roughly three orders in 200 ticks. agent.run resets the world with an explicit seed, so the stimulus stream is pinned per call; this is an execution-order effect on a trajectory sensitive at the last bit, not a loose seed.

Replicate spread at the horizon, |run − replicate|, same seed, same command, separate processes: the both arm is bit-exact through 5,000 and reads ΔER 3.3e-04 at 20,000; the transport arm is bit-exact through 2,000 and reads ΔER 1.78e-01 at 20,000 — an order looser.

This is already in the map's numbers. The amended B62 entry quotes transport at 5,000 as ER 3.1347 / q 0.0636 / 1.96 dimensions / aud-diff 0.5505; B68's run of the same command at the same rung reads 3.0024 / 0.0543 / 2.03 / 0.5313. Both are honest single runs. What remains untested is the null taken at every checkpoint rather than only at construction.

Proposed standing constraint: a difference between two arms that run TransportRule is quoted against a replicate of one of them at the same rung, or it is not quoted — and where a rung reproduces bit-exactly, say so, because that is stronger than any spread.

Row — the rules, not the world, are what re-excite the body past the stall; and the pair kills it deader than no rule at all.

Stamped per rule mode, since the rules change the commands the agent issues:

mode	last above	first fall	re-crossings	share above	final std_max
frozen	150	160	0	0.0075	2.44e-03
bias	4,900	90	70	0.0390	9.32e-07
transport	15,790	80	131	0.0690	1.32e-04
both	70	80	0	0.0035	2.33e-10

B61 (#634) describes the late re-crossing as the world sporadically crossing back after it has died. On these four modes the frozen arm re-crosses zero times while the rule-carrying arms do it 70 and 131 times — so on this arm it is the rules', and B61's own constraint (report both rules and the re-crossings) is what makes that visible. Offered as an amendment to B61's account, not applied to it.

The co-occurrence with the readings is exact: the bias arm's collapse (500 → 1,000, ER 3.03 → 1.34) falls in its single largest re-excitation window, share 0.320 against the frozen arm's 0.000; the transport arm's sustained rank elevation from 5,000 on rides on sustained intermittency, 0.068 and 0.066, where frozen reads 0.000 throughout. Co-occurrence does not give the direction and none is claimed. It does mean the rule-carrying arms past ~160 ticks are not drift under a frozen stimulus, which is what B38's rule would have to say of them.

And the both row is a fourth column where the pair lands outside the interval its solos span: each rule alone keeps the body intermittently alive, and together they produce a lower moving share than the no-rule control and a final motion six orders below it.