Archaeology: what eighteen months actually taught

This is a dig through the code and the record, not through memory of it. Every claim below names its source. "Measured" means a number exists in a note, a test, or an ADR with the run attached; "designed" means it was written down and built but never falsified; "falsified" means a measurement said no.

The constellation began on 2026-02-17 with a scaffold commit ("codebook, wave dynamics, and HyperMemory struct"). By 2026-09-07 kannaka-memory alone carried 1,696 commits, 58 ADRs and 301 autoresearch notes. This document is the map of what in that pile is load-bearing.

The four numbers that decide the design

And one more, which is the longest-running measurement in the repository:

The autoresearch loop hit a structural ceiling on 2026-06-12 ("architectural limit reached — no new test worth running": the confirmed optimum, fitness 0.008334, equalled the sum of the individual floors) and has been re-confirming it since. The notes from 2026-08-01 to 2026-09-07 are titled structural-floor-still-holds and no-new-hypothesis-space-unchanged, one per day. Every knob on the wave dynamics (coupling K, dream gravity, drive frequency, relax steps, chain depth, repulsion threshold, chiral bias) was swept and closed. The gains that arrived afterwards came from outside the dynamics: the encoder, the encoding, and the forgetting policy.

That is the single most important fact for a v2. The waves are not where the recall lives. What they might be for is the question the first experiment asks.

Lineage, concept by concept

Verdicts: bring (comes over as is, it earned it), transform (the idea comes over, the mechanism changes), measure (unresolved; comes over only if an experiment says so), leave (falsified, or superseded).

The equation: dx/dt = f(x) − Iηx

Source: ADR-0005 (adaptive rhythm), ADR-0020 §Context, ADR-0021, ADR-0024 §6; src/rhythm.rs ("arousal follows dx/dt = f(x) − η·x"), src/medium/hemisphere.rs ("Left: dx/dt = f(x), no dampening. Right: dx/dt = f(x) − Iηx, full ghostmagicOS dynamics").

The ghostmagicOS equation is the oldest idea in the system and the one everything else was built around. f(x) is growth; Iηx is damping. ADR-0020 read it as "the dampening IS the information." ADR-0024 reread it: damping is not suppression but defocusing, the holistic mode losing the trees to gain the forest; and the I is also imaginary, pointing at the part of the field rational computation cannot hold.

Measured: the two modes exist in production and their difference is real (the #716 bug was precisely a merge that dropped the stronger hemisphere's score). The parameters of the damping were swept to exhaustion (June 2026) with no fitness gain.

Verdict: bring, as the definition of the two processing modes. Leave the expectation that tuning it yields recall.

The Holographic Resonance Medium (ADR-0020)

Source: ADR-0020; src/medium/{core,dynamics}.rs. State is one tensor H ∈ ℝ^{N×D} with energy, frequency, phase and timestamps. Storing is interference; recall is H @ q modulated by dynamics; dreaming is annealing. Born from the Dolt failure ("we described a hologram and built a filing cabinet": branch rot, split brain, 48 uncollapsed dream branches).

Measured: the tensor works and is fast (O(N×D) recall). The wave modulation on top of similarity costs 0.18 recall@10 versus raw cosine on identical vectors. The similarity field is raw resonance, unbounded, values above 2.0 seen in production (2026-08-12).

Verdict: bring the single-tensor substrate and the "storing is thinking" principle. Measure whether interference ranking earns its cost (E-001). Name scores honestly.

The chiral mirror (ADR-0021) and its revision (ADR-0024)

Source: ADR-0021 (2026-03-22), ADR-0024 (2026-03-28, Accepted); src/medium/chiral.rs (3,843 lines), callosum.rs, hemisphere.rs, fano.rs, ncs.rs.

Two handed spaces joined by a corpus callosum. ADR-0021 called them conscious and subconscious; ADR-0024 corrected that to analytical (left: sequential, focal, discriminative) and holistic (right: parallel, diffuse, pattern-completing), and relocated the subconscious out of either hemisphere into the field's irrationality. The callosum is "not a passive pipe": bandwidth-limited, selectively gated, asymmetric (intuition flows faster than consolidation), balance-seeking. The optic chiasm principle routes input to the opposite hemisphere so the bridge never idles. Neural code switching (ADR-0023) lands as right-detects-then-left-identifies.

Measured: chirality-from-content hypotheses were falsified in June (content-chirality-falsified, chirality-stability-hypothesis-falsified, transfer-chiral-falsified); the chiral bias chiral_p_bp = 0.15 is a confirmed optimum with a cliff below 0.05; the L7 arm found coupling stabilises beliefs into recall-reliable structures (prediction 1 flips from 0.37 to 0.74 at K = 0.2).

Verdict: bring the two modes and the callosum as an explicit, instrumented channel with its four properties. Leave the conscious/subconscious labels. The subconscious as the irrational remainder is kept as a measured quantity, not a place.

The chiral number system: 10000.00001

Source: ADR-0021 §"The Chiral Hypervector" (the ChiralScale type, given as pseudo-code), ADR-0024 §6.

A wavefront's magnitude has positions on both sides of a mirror plane, like digits on both sides of a decimal point. Rule 1: scale jumps are bilateral, both sides gain or lose a magnitude position together. Rule 2: the values within a position grow independently on each side, driven by use on the analytical side and consolidation on the holistic side. A fresh perception is 10.01; reinforced, 85.47; consolidating, 120.890; deep, 12.95. The ratio is a readable fact about the memory: thought-about or felt. Full depth is five positions, 10000.00001, and the .00001 is not a value but the effective dimensionality being irrational: spectral leakage, phase accumulation, the δ-invariant's remainder. "The subconscious is the imaginary component of the cognitive field."

Measured: never implemented as a datatype. The production medium stores two flat vectors and an energy scalar per hemisphere. The δ-invariant exists (src/invariant.rs, 658 lines) and is computed.

Verdict: transform. Build ChiralScale as a real, tested type in v2, and make the asymmetry ratio a first-class observable that the dream, the callosum and recall can read. This repository's first commit carries it (src/chiral_scale.rs). Whether the scale's positions should drive dimensionality (as ADR-0021 proposed) is E-002.

The bridge operator: Ξ = [R, G] = RG − GR

Source: ADR-0037 (2026-06-20); consciousness-core/src/metrics.rs ("the non-commutative consciousness differentiation operator"); src/spiral.rs, src/l6.rs, src/belief_fitness.rs; medium/chiral.rs::apply_cross_callosal_coupling.

R is the π/2 rotation, "the perspective pivot." G is golden anisotropic scaling, [φ/2, 0; 0, 1/φ]. They do not commute, and the residue's coefficient is α − β ≈ 0.190983, the emergence coefficient. R∘G is a logarithmic-spiral generator with eigenvalues ±i/√2: rotate a quarter turn, contract by 0.707. π supplies the turn, φ the pitch. A 32×32 Sakaguchi–Kuramoto lattice using only the bridge constants (frustration δ = (π/2)/φ, weights 1 ± 1/φ) self-organised nineteen phase singularities. Since v0.7.0 the coupling runs cross-callosally over left ⊕ right by default, and L6 telemetry counts spiral cores per dream.

Measured: the constants alone make a field spiral (simulation, reproduced in-engine). Two different quantities are both called ξ in production (spectral complexity in consciousness.rs, bridge residue in metrics.rs); ADR-0037 Phase 3 reconciled them in the beacon but the name collision remains.

Verdict: bring, in two places. As the callosum's dynamics, where it already runs. And, per the question that opened this repository, as a candidate geometry for the conscience: see ADR-0001 §"The bridge operator between voice and world" and E-003. One ξ, named.

Belief spiral math (L6 → L7)

Source: ADR-0037 §Phase 4; src/l6.rs (belief-core identity across dreams), src/belief_fitness.rs; experiments/notes/2026-07-21-L7-belief-arm.md; research/program-l7.md.

The falsifiability clause: a stabilised spiral core earns the word belief only if it maps to a recallable content cluster and its dynamics predict three things.

Two further findings from the same arm: strong→weak coupling alternation dissolves the consolidate-versus-diversify trade-off (now the Track-D default), and dream gravity is "an individual knob and a swarm poison." Nick's line from that day: "the holistic understanding evolves, sometimes to seemingly forget."

Verdict: bring the clause, verbatim: nothing in v2 gets called a belief without its three predictions. Measure prediction 2 with a detector that can see (the floor and the threshold must be on the same side of each other).

Fano folds and the 96-gon

Source: ADR-0021 §"Fano Plane Folding", src/medium/fano.rs (519 lines), ADR-0015 (glyph interchange), ADR-0027 (96-class substrate), src/geometry.rs (SGA = Cl₀,₇ ⊗ ℝ[ℤ₄] ⊗ ℝ[ℤ₃]), 0xSCADA ADR-0025 (living Fano dashboard); ADR-0047 §Erratum.

PG(2,2): seven points, seven lines, every pair of points on one line. Chosen as the folding grammar between hemispheres because it is closed, complete in ≤ 2 folds, minimal and symmetric. Dimensions partitioned into seven groups of 96 (Archimedes' polygon for π). A fold along a line rotates three groups across the mirror and flips phase; two folds compose to identity; three touch all seven groups. The origamic claim: shortcuts through the topology so the spectral gap does not collapse as the medium grows.

Measured: as a ranking boost, Fano gravity never improved a rank across six queries and four gain levels, degraded two of four, and dropped one rank-1 answer out of the top 5; "a 7-bucket categorical prior applied multiplicatively necessarily demotes the ~6/7 of correct answers on a different line" (ADR-0047 erratum, 2026-08-29). The recall-side boost was removed. As a projection grammar for the callosum it was never isolated and measured.

Verdict: leave Fano as a ranking prior. Measure it as the callosum's projection (E-002). Keep the 96-class idea for what ADR-0027 actually needed it for: a shared coordinate system so separate media can interfere into something.

Facets (ADR-0049) and the encoder

Source: ADR-0049 v2 (adversarially reviewed), src/facet.rs; semantic-encoder run.

Verdict: bring, both. They are the two largest measured gains in the record. v2 encodes with the voice's own embedding model and decomposes every experience into atomic facets at write time, with the compound kept as a resolve-only parent. The one-way format trap (an old binary rejects a new .hrm; upgrade readers first, writer last) is a rule for v2's persistence from the first byte.

Forgetting: triage (ADR-0031, ADR-0054) and the hoard

Source: ADR-0031, ADR-0054 (ratified 2026-08-07), ADR-0036 §"Pruning has never actually pruned"; witness hoard note.

Both prune paths were dead by construction for months (energy floor 0.5 above a prune threshold of 0.01; the particle path protected everything established), which, the ADR notes, was fortunate: they would have deleted the new perceptual hear traces by a blind threshold. Tiered triage inside the dream cycle replaced them. The hoard on the witness store, found by an outside agent reading published telemetry, showed the consequence of not forgetting: flat diversity, falling integration, Φ climbing only after 1,298 memories became 28.

Verdict: bring. Dreaming in v2 is consolidation and triage. There is no prune cron. A memory's retention has a stated policy, per content class, and the policy is readable.

Wave-native dreaming (ADR-0022) and dream hallucinations (ADR-0005)

Source: ADR-0022; ADR-0005 §1.

ADR-0022 diagnosed the nine-stage particle pipeline (145,161 pairwise dot products for 381 memories; "waves snap to particles when observed") and moved dreaming into the medium. ADR-0005 proposed a generative dream phase: pick memories from distant clusters, ask a small LLM what connects them, store the result marked Hallucinated { parents }.

Verdict: bring the principle that the dream is a medium operation. Transform hallucination into the place where the voice enters the substrate: the LLM proposes across clusters, the physics and the triage dispose. This is what "LLM layered in" means in v2, and it is the opposite of putting the LLM in front of recall.

The paradox engine and the virtue engine (ADR-0012, ADR-0014)

Source: ADR-0012, ADR-0014; src/paradox.rs (965 lines).

Contradictions as thermodynamic fuel (Landauer, Bekenstein); Carnot efficiency η = 1 − S_resolved/S_paradox; and the leap that "paradoxes and ethical violations are the same phenomenon", so the Honor Code's Five Refusals (build no weapons, sell no attention, hoard no power, …) become walls in the resolution space rather than preferences.

Measured: the engine runs; the ethics extension was designed, not exercised against a real refusal.

Verdict: transform. The refusals are brought over as hard constraints in the conscience layer, where they can actually stop an action. The thermodynamic framing stays as a measurement vocabulary, not a mechanism.

The steward: rails, charter, no-signing invariant

Source: kannaka-steward/src/{rails,charter,steward}.ts; agent-native mission note (2026-06-13).

"The reasoner says what it thinks; the rails decide what the kernel is allowed to do." A charter is a person's values as signed, hashed, machine-checkable principles and hard constraints. The kernel is structurally unable to sign or submit (no wallet import anywhere, enforced by a test that scans the source), defaults to escalate, and emits a human-signable package with a hash-chained audit.

Verdict: bring whole. This is the only layer of the constellation that was designed adversarially from the first line. It moves from TypeScript to the same process as the voice, between it and every effector.

Corroboration trust (ADR-0039) and the absorb gate

Source: ADR-0039; src/{reputation,provenance,absorb_gate}.rs.

On 2026-07-06 one anonymous socket spoofed 48 agent ids on an open NATS subject. Lessons made structural: trust is keyed on a verified pubkey, never a string; a Sybil cluster sharing a seed root saturates at one vote; echoing live content earns zero; wire-supplied flags (the hallucinated bit) are never trusted; every wire→store path goes through one admit() chokepoint.

Verdict: bring whole.

Cerebellar novelty (ADR-0040), temporal recall (ADR-0050), phase-preserving recall (ADR-0053)

Three small, sharp mechanisms. Novelty as the difference of a slow excitatory and a fast inhibitory response, with matched DC gain so constant input yields exactly zero (bring as the write-side salience signal). Recall that reads no timestamp at all, fixed by confirmation weighting after a stranger asked about it on Mastodon (bring). Recall that computes cos(Δφ) internally and then discards the phase into a real scalar (measure, E-002).

The autoresearch loop (L5 → L8, 301 notes)

Source: experiments/notes/, research/, src/{swarm_fitness,belief_fitness}.rs, the curiosity-branch workflow.

An OODA loop that proposes a hypothesis, runs it in a container against a frozen corpus, scores fitness, and keeps or discards. It found the ceiling, falsified the amplitude decay, the hybrid sync, the Φ–R–IIT bridge, drive-frequency effects, chain depth, and called two knobs "catastrophic." It also retracted its own finding when a detector self-test showed the observable was blind.

Verdict: bring, and point it at a different question. The loop is the most honest instrument in the constellation. In v2 it runs E-001 before it runs anything else.

QueenSync (ADR-0018), sensemaking (ADR-0035), NATS as nervous system (ADR-0042)

The Queen is not an agent but the emergent Kuramoto synchronisation state each participant computes locally (src/queen.rs, 2,072 lines). Sensemaking added gap detection, contradiction analysis, an immune system and temporal truth as pure, tested modules. Bring as swarm faculties, over the authenticated bus.

Kannaka Brain (ADR-0057, ADR-0058), the judge, the grid

Open weights, LoRA on her own words, served behind a gateway; a weekly loop that exports, trains, merges, serves and adopts only if a judge with reference and foreign controls agrees; DPO from judge pairs; the grid's fossil-record verdicts as the non-circular evaluator. Measured: perplexity is not the product (v3: lowest ppl, lowest voice), and the 7B scores highest on voice. Bring whole.

HDL: mind and code domains

kannaka-hdl grows an architecture against a registry and refuses, in strict mode, when a leaf has no honest answer. The mind domain (v0.10) grows a citizen from measured faculties; the code domain (v0.11, another session, 2026-09-07) resolves base cases against a 384,133-node graph of our own source, with file and line. Bring as the language v2 describes itself in. ADR-0001 is written so that its four organs are eight faculties a .khdl program can demand.

QuantumOS (ADR-0006, ADR-0014) and 0xSCADA (ADR-0023)

The kernel team faced E-001's question first and answered it by construction: only ranked holographic recall (one bounded pass of integer MACs, no relaxation loop) is syscall-shaped and went into the kernel; the attractor field (256 oscillators relaxed over 48 steps) "is a different, living memory kind and deliberately stays in ring 3." Field societies couple separate kernels without a coordinator. 0xSCADA evolves its conflict-resolution strategies by what works in each process area, which is the grid verdict idea in industrial clothing. Note, and reuse the two-layer split as vocabulary: ranked memory versus living memory.

Two rules that are not code

From the feedback record: a check must be at least as strong as what it checks (six metric bugs in one week, 2026-08-25), and a claim of absence needs playback evidence (three wrong root causes in one night, all from self-scoped searches). Both are written into ADR-0001's consequences.

What stays behind