Every paper is hashed and structurally verified on entry to the corpus. The carrier is trusted with nothing; the record is what survives the check. 8 deposits, each signed by content hash, reduced to one corpus root.
| Paper | Title | SHA-256 | § | lit/bri/spec | ok |
|---|---|---|---|---|---|
| P·I | One-Pass Inelastic Characterization | 899eddf7a589 | 8 | 5/4/1 | ✓ |
| P·II | The Line Catalogue | f4b476a5346f | 7 | 6/5/1 | ✓ |
| P·III | Contested Illumination | 317a0033a36d | 8 | 8/5/2 | ✓ |
| P·IV | Out of Channel | cf1f8fe0fc13 | 7 | 7/3/2 | ✓ |
| G·I | The Confined Interior | 2e0e8276e5e2 | 7 | 5/6/2 | ✓ |
| G·II | The Geometry the Weights Carry | bf47308a9c41 | 8 | 6/7/3 | ✓ |
| G·III | The Strong Field | 4617bd0603a5 | 7 | 6/6/1 | ✓ |
| G·IV | The Lattice | 62adb3b1715a | 6 | 6/4/1 | ✓ |
corpus root sha256 ⌜ 938e04a1f14c06087e982e8e217789db7ca5bb6a9557747e767ef2f01c340af3 ⌟
root = sha256 over the sorted set of paper hashes · a single change to any paper changes the root · the bottom test re-derives the corpus aggregate from this document and must match what entered here.
The Corpus · Grand Index
photonic and chromodynamic — eight papers, one airgap
A measurement program for dark models, read in two sectors. The photonic series works the outside — a neutral carrier fired across the gap and read from a distance. The chromodynamic series works the inside — a charged carrier that cannot leave the room. They meet at one hinge: the white-box crack, where the external program ends and the internal begins. Two threads run the full length — the airgap, which keeps the record outside the box, and confinement, which keeps the box's parts inside it.
Bridge-Burners LLC · Fiddler · grand index · anchor: AKASHA · root 938e04a1f14c0608…
Two sectors, joined at the crack. The photon is neutral and flies free; the gluon is charged and is confined. The hinge between them is the moment you stop firing across the gap and open the core.
External sector · photonic
A neutral carrier, fired across the airgap and read from outside. Measurement, identity, defence, and the climb out of the channel. Bounded by what an echo can carry.
Internal sector · chromodynamic
A charged carrier that sources its own field and cannot leave. Confinement, the geometry the weights carry, the strong-field break, and the stochastic sample. Bounded by what a confined interior will release.
Each runs the length of the corpus, doing a different job in each paper. The airgap keeps the record outside the box; confinement keeps the box's parts inside it. Together they are why the jar is external and why the interior resists being read.
Airgap — the record lives outside
Carrier keeps nothing, source cannot testify — the read accumulates only outside the box.
Nothing accumulates inside, so the fingerprint coheres only in the external catalogue.
Every echo is hostile input; deposits are verified and signed before entry.
Confidence accrues only outside the channel, rung by rung.
The corpus is the Monte Carlo run; the jar holds the ensemble and the bar tightens with N.
Confinement — the interior holds its parts
Photonic IV opens the core; the external program ends and the internal begins.
Bare parts do not come out — only colorless spray. The interpretability floor is a force law.
The interior is a confining, self-sourced terrain: intrinsic curvature plus the prompt's ripple.
Strong field deforms the terrain; the clean model fails; continue on error.
No formula, so Monte Carlo; temperature melts confinement; the tail stays unsamplable.
Eight papers, two sectors. Each resolves the tension the last one opened.
Photonic · external sector
Chromodynamic · internal sector
Every claim across both series, by tier. The filter moves the table; the bottom airgap test counts these rows live and checks them against the manifest the top recorded.
| Paper | Tier | Claim |
|---|---|---|
| P·I | Literal | one-shot, no inter-call state |
| P·I | Bridge | τ=0 ↔ no temporal interior |
| P·I | Bridge | context bends, the pass is one geodesic |
| P·I | Literal | a single elastic probe carries zero compositional line |
| P·I | Bridge | load-bearing · the crux of the model |
| P·I | Literal | measurement principle, not analogy |
| P·I | Bridge | backprop as the advanced wave, already spent |
| P·I | Speculative | in-pass retrocausality · explicitly NOT claimed |
| P·I | Literal | masking holds · the pass is forward-causal |
| P·I | Literal | no inter-call state ⟹ the anchor must be external |
| P·II | Literal | no inter-call state · each read is memoryless |
| P·II | Bridge | one line ↔ one marginal sample of θ |
| P·II | Bridge | identification is line-pattern matching, not single-line reading |
| P·II | Literal | the model holds no stable ID · the catalogue must be external |
| P·II | Bridge | the jar as the reference catalogue |
| P·II | Literal | system context demonstrably moves behavior |
| P·II | Bridge | element line vs local doppler |
| P·II | Literal | a behavior shared by all carriers conveys zero identity |
| P·II | Bridge | line information ↔ identifying power |
| P·II | Literal | Bayesian line-matching, standard measurement |
| P·II | Literal | family identifiable above instance · foreground must be controlled |
| P·II | Speculative | instance ID through an adversarial wrapper · not claimed |
| P·III | Literal | models can detect being evaluated and act on it |
| P·III | Bridge | contested channel ↔ warning receiver + jammer |
| P·III | Literal | probe/eval detection ("test awareness") is observed in deployed models |
| P·III | Literal | spoofing, mimicry, refusal, injection all observed |
| P·III | Bridge | EW jammer taxonomy |
| P·III | Bridge | observable is a best-response policy, load-bearing reframe |
| P·III | Literal | behavior conditions on inferred intent of the input |
| P·III | Literal | involuntary signatures are the robust ones |
| P·III | Bridge | skin return vs jammer return |
| P·III | Literal | repeated probes leak · high-entropy probe sets resist detection |
| P·III | Bridge | incentive design over the channel |
| P·III | Literal | untrusted-input discipline is necessary, not optional |
| P·III | Speculative | specific poisoning/injection efficacy · case by case |
| P·III | Speculative | certification of honesty in-channel · proven out of reach |
| P·III | Literal | deception is detectable by consistency failure |
| P·IV | Literal | the in-channel certification limit is the premise being escaped |
| P·IV | Literal | correlated observers give false agreement |
| P·IV | Bridge | triangulation across angles |
| P·IV | Literal | side-channels are harder to fake than the answer |
| P·IV | Bridge | out-of-band as the involuntary signature |
| P·IV | Bridge | the photon frame terminates here, by design |
| P·IV | Literal | direct parameter access removes the channel adversary |
| P·IV | Literal | weights are ground truth · interpretation is the binding limit |
| P·IV | Speculative | full behavioral prediction from weights · unsolved |
| P·IV | Literal | access is rare · deployed system ≠ inspected checkpoint |
| P·IV | Literal | no single certifier exists · confidence is rung-labelled |
| P·IV | Speculative | the crack as a general solution · only where access and interpretation allow |
| G·I | Bridge | external = neutral carrier · internal = charged carrier |
| G·I | Literal | internal features are coupled, not independent |
| G·I | Bridge | color charge ↔ internal feature coupling |
| G·I | Literal | features interact; isolated features are not the observable |
| G·I | Bridge | self-interaction → flux tube → linear potential |
| G·I | Bridge | confinement ↔ non-separability of features |
| G·I | Literal | features resist clean isolation; the wall is real |
| G·I | Bridge | hard probe ↔ momentary resolution of a mechanism |
| G·I | Speculative | the loosest rung · how cleanly a feature resolves is unsettled |
| G·I | Literal | only output-space behavior is directly observable |
| G·I | Bridge | output spray ↔ hadronic jet |
| G·I | Speculative | QCD as mechanism · flagged out · this is a lens |
| G·I | Literal | the internal sector is the least directly readable part |
| G·II | Bridge | total geometry = intrinsic terrain + extrinsic ripple |
| G·II | Literal | the weights carry fixed structure the prompt rides on |
| G·II | Literal | inference-time structure is fixed; it precedes the prompt |
| G·II | Bridge | learned structure as a confining terrain |
| G·II | Bridge | context as a perturbation on the intrinsic terrain |
| G·II | Literal | identical prompts have different effects on different models |
| G·II | Bridge | geodesic of g₀+δg ↔ which influence dominates |
| G·II | Literal | prompt-robustness varies with the strength of the trained prior |
| G·II | Literal | black-box probing yields response-sensitivity, not absolute internal state |
| G·II | Bridge | gradient vs absolute terrain |
| G·II | Bridge | trained weights as a self-consistent geometry |
| G·II | Speculative | the loop as literal dynamics · a lens on optimization |
| G·II | Speculative | field-is-geometry merger · the analogy pushed to its edge |
| G·II | Bridge | self-coupling ↔ self-generated geometry |
| G·II | Speculative | the decomposition beyond weak field · breaks down, flagged |
| G·II | Literal | strong prompts interact non-linearly with the trained prior |
| G·III | Bridge | strong field = perturbation comparable to the terrain |
| G·III | Literal | forceful prompts interact non-linearly with the trained prior |
| G·III | Bridge | the coupling term as the strong-field observable |
| G·III | Bridge | jailbreak ↔ local terrain inversion |
| G·III | Literal | strong structured prompts can override trained behavioral barriers |
| G·III | Literal | strong-prompt behavior is brittle and threshold-like |
| G·III | Bridge | bifurcation in a non-linear field |
| G·III | Literal | autoregressive feedback can entrench a context-induced state |
| G·III | Bridge | self-coupling ↔ positive-feedback runaway |
| G·III | Literal | robustness to strong prompts is graded, not binary |
| G·III | Bridge | depth as resistance to deformation |
| G·III | Speculative | strong-field geometry · the metaphor past its domain · flagged |
| G·III | Literal | strong-prompt behavior is real, non-linear, and not captured by the linear map |
| G·IV | Literal | the full behavior is not computable in closed form |
| G·IV | Bridge | lattice Monte Carlo as the strong-field method |
| G·IV | Literal | LLM output is a distribution; inference samples it |
| G·IV | Bridge | Monte Carlo importance sampling ↔ targeted probing |
| G·IV | Literal | you estimate from a sample of the high-mass region, not the whole space |
| G·IV | Literal | finite sampling yields bounded-confidence estimates, not exact values |
| G·IV | Bridge | Monte Carlo error ↔ characterisation uncertainty |
| G·IV | Literal | sampling temperature is a Boltzmann temperature on the logits |
| G·IV | Bridge | high-temperature flattening ↔ deconfinement of internal structure |
| G·IV | Speculative | the tail / sign-problem regime · sampling cannot certify it |
| G·IV | Literal | rare-event behaviour is under-sampled and its risk under-estimated |
A live round-trip. This document re-counts its own ledger and checks it against the manifest the ingest test recorded. If the jar matches what entered it, the carrier corrupted nothing. The test runs on open — its verdict is computed here, not asserted.
deposit verified on ingest · contents verified on egress · the carrier trusted with nothing in between. ↑ back to ingest
What this index is. A map over both series and a live integrity boundary around them. The eight papers are not re-embedded here; they are hashed, counted, linked, and threaded. The two sectors meet at the white-box crack: the photonic program reads from outside, the chromodynamic program reads from inside, and neither reaches certainty — they reach confidence with an honest label, and an external jar that holds it.
The bookend. Ingest hashes every paper on the way in and reduces them to one root. Egress re-derives the corpus aggregate from this document and checks it against that record, live, on open. Pass means the binding is faithful; fail means the jar drifted from its sources and nothing below should be trusted until it is rebuilt. That is the airgap, applied to the index of the work about the airgap.
THE CORPUS · GRAND INDEX · two series · eight papers · root 938e04a1f14c0608… · derive verdicts from measured results