Contents
Where this sits
The spikes and the golden corpus feed the preregistration, and the preregistration feeds the Python reference that will decide every claim.
What the spikes can decide
Nothing in the ledger. The spikes are feasibility checks: they ran on the existing code to find out whether the design could work and what would break it, and several of the design’s choices come straight from what they found.
- They use the repositories’ code unchanged: Owl Academy’s engram core copied verbatim, PolyStep’s browser cores and the fence loaded as they are, and the orbital pages’ code read from their build templates (since Phase 1, required from the shared module
orbital_core.js). - Their outputs are committed beside them, and every run is deterministic, so anyone gets the same lines.
- They ran before the preregistration was written. A claim tested on the same cases that inspired it would be circular, so each claim is decided later, by the Python reference, on the instances the preregistration fixes.
Spike A showed what the key really is, which is why the Scope section and claim E9 exist. Spike B’s depth limit is why position and frame are kept apart. Spike C’s same-bit case is why the Merkle chain is required, and its clutch measurement chose the free-run setting. Spike D’s straight leaves are why the fence only screens and an exact check decides.
Spike A · The key
spikes/engram_core_audit.cjs: what engram/1 is, measured on the verbatim core. 15 of 15 checks pass.
- The copy reproduces Owl Academy’s vectors:
GEODESIC → 5c896a4801c9c24c,chiralKey(STRIX) = O1::S::48d835dfdfc5484e. - The 26 letter grids have rank 26 over GF(2): at most 2²⁶ keys occur, and two words share a key exactly when every letter count has the same parity.
- Distances between distinct words look random: mean 32.4 of 64 over 190 pairs, from 20 to 44.
- The window works as a triple quad over GF(2): any one of α, β, γ is rebuilt from the other two and Y.
Loading the captured output…
Spike B · Quaternion addresses
spikes/quaternion_address.cjs: engrams as exact addresses in PolyStep’s norm-5 tree, on its browser cores. 8 of 8 checks pass.
- 2000 of 2000 random 64-bit words round-trip through one integer quaternion of norm 5²⁸, components of at most 33 bits.
- A chain of 40 packets is one quaternion that unfolds in order, 40 of 40. It is exact, not compressed: 5,204 bits for 2,560.
- Float placement in H³ decodes to depth 12 (50/50) and fails from depth 14 (36/50 at 14, 9/50 at 16, 0/50 at 18).
- The packet quaternion is an exact rotation, and the coupler’s mirror law keeps B = q₂q₁ on it.
Loading the captured output…
Spike C · The orbital memory code
spikes/orbital_memory.cjs: the window on the tiles, Merkle chain, gauge and pump of the orbital pages. In Phase 0 it read them by name from the build templates; since Phase 1 it requires scripts/geometric-build/orbital_core.js, with the same results. 9 of 9 checks pass.
- All 60 words of a sample dialogue travel as two Form 5 tiles and decode with their header fields intact.
- On 58 windows a flipped bit is detected by the parity, located by the Part 1 Merkle chain and rebuilt by XOR, with the root matching again: 58 of 58 for each step.
- Two altered cells are caught by the chain; the same bit flipped in two cells passes the parity and is caught only by the chain.
- The convergence gauge recovers the three footer distances exactly, 57 of 57; one packet in 57 locked.
- The pump identity holds on every channel. In the hold setting a register leaves P = 360/720; in free-run it closes.
Loading the captured output…
Spike D · The fence on a memory tree
spikes/fence_memory.cjs: the 0.7 Claude fence and its ln 2 guard on the norm-5 tree in H³, using fence.js as it is. 4 of 4 checks pass.
- Memories placed two levels too deep are always flagged, 5 of 5 in every run; one level too deep, 4 of 5 and 5 of 5.
- At depth 3 the fence also flags the six straightest leaves, 6 of 187 memories in their proper place, because it measures distance from the centre and straight paths reach farther.
- Comparing each flagged memory’s exact point with the exact point of its recorded address clears those six and confirms every misplaced memory.
- The guard passes clean distances (worst defect 0.6923, below ln 2 = 0.6931) and blames a corrupted distance on its own pair.
Loading the captured output…
The golden corpus
reference/golden/engram_v1.json is the target claim E1 must hit. It was written by make_golden_v1.cjs from two files of the Owl-Academy repository, read unchanged: the reference page Engram_Mind_Eye.html and the second implementation, scripts/biomesh.js.
| Part | Contents |
|---|---|
| Letters | All 26 |
| Concepts | The 15 names in STRIX’s concept library |
| Window | The 3 seeds the page starts with: MIND, EYE, STRIX |
| Dialogue | 477 distinct words of the STRIX introduction and the concept library’s responses |
| Edge cases | 15, among them the empty string, numbers, mixed case, an apostrophe, a hyphen, an anagram pair and a 28-letter word |
| Tokenizers | 3 sentences split by both of Owl Academy’s tokenizers |
| Packet chain | The 69 packets of a fresh session: the introduction logged as STRIX, then one human question, with every header, body, footer and the final parity |
Every entry is computed by both Owl Academy implementations, and the script refuses to write if they disagree. They agree on all 536. The file records the Owl-Academy commit (9638040) and the SHA-256 of both source files, and the preregistration records the SHA-256 of the corpus itself.
Two tokenizers. The Mind Eye page splits on anything that is not a letter or digit, so don’t becomes DON and T; biomesh.js keeps it as one token. A sentence gives different packet counts in the two, so the corpus keeps both.
Two hex spellings. The chiral key prints the grid byte by byte, least significant bit first (5c896a48… for GEODESIC). The packet body prints the same 64 bits as one number, most significant first (4CC2C901…). Both are pinned, so neither can be confused for the other later.
The nine claims
Registered on 2026-11-04 in PREREGISTRATION.md. An exact claim is supported only if every case passes, and refuted by any single failure. Every exact test has a control, a broken variant that must fail, or the test is void.
| # | Test | Control that must fail | Decision |
|---|---|---|---|
| E1 | Every field of the 536 entries, both tokenizations, the engram stream and the 69-packet chain | Registered: seed multiplier 37 in place of 31, found void (it can never fail). Amendment 1: Mulberry32’s increment 0x6D2B79F7 | Exact |
| E2 | Chains of 1, 2, 10 and 72 packets (20 each) and 720 packets (5): norm 5^(28K), every packet recovered in order | A backtracking packet; two packets swapped | Exact |
| E3 | Every single-bit flip of every cell of every window: detected, located, rebuilt, root restored. 200 double alterations: detected, repair refused | Parity alone misses a same-position double alteration | Exact |
| E4 | Every complete register of 20 synthetic dialogues: packet, block and group products equal; free-run + membrane closes every register | Two blocks swapped; one packet replaced | Exact |
| E5 | The golden session and 20 dialogues: B unchanged at every zero point; carry bit-identical to no coupler | The same-chirality coupler | Exact |
| E6 | All 274,625 distance triples through the gauge; the lock on the dialogues, 65 equal triples and 5,000 random ones | A gauge with Y = b + c | Exact |
| E7 | Memory trees of depth 2 and 3 with 5 or 10 memories placed 1 or 2 levels too deep, 50 instances per cell | — | Settlement exact; detection at depth +2 ≥ 0.99 with the interval’s low end ≥ 0.95 |
| E8 | 50 matrices of 24 memories, one distance raised by 0.5 to 3.0 | — | Clean bound exact; blame at +1.5 and +3.0 in ≥ 99% with the low end ≥ 0.95 |
| E9 | a: five parity-keeping variants of each stored word. b: five one-letter substitutions, against a shuffled-key store | — | a exact; b supported if the difference’s interval lies within ±0.05 |
E9b predicts that recall from a one-letter misspelling is no better than chance. Substituting one letter XORs two letter grids into the key and moves it about 32 bits, as far as an unrelated word. If the runs show otherwise, the claim is refuted and the Scope section is rewritten.
The reference ran every claim on 2026-11-04, its runner committed beforehand. E1–E7 and E9 are supported. E8 is inconclusive: the ln 2 bound held exactly, but the guard named the corrupted pair 96 times in 100 at +1.5, short of the registered 99. The prediction above held: misspelling recall was at chance. Every number is in Explainer 4.
Running it yourself
Everything here runs with Node 18 or later and nothing else, from the repository root.
node "Pages/Engram Minds Eye/spikes/quaternion_address.cjs"
node "Pages/Engram Minds Eye/spikes/orbital_memory.cjs"
node "Pages/Engram Minds Eye/spikes/fence_memory.cjs"
node "Pages/Engram Minds Eye/reference/golden/make_golden_v1.cjs" ../Owl-Academy --check
The last line needs an Owl-Academy checkout beside this one. It recomputes the corpus from Owl Academy’s files and says whether the committed file still matches, so a change on that side is seen.