94f05ba84f
`packs.compiler._blend_feature_versors` ignores its strength argument and
returns the target verbatim. All three "nudge" sites — cross-language alignment
(0.10), morphology clustering (0.40), mount-time domain resonance (0.40) — are
therefore total overwrites, and ADR-0015's two explicit requirements ("a
weighted graph, not a translation table"; "without flattening their
distinctions") are violated bit-exactly.
Measured: mounting the depth packs onto English REMOVES 37 distinct coordinates
(239 surfaces -> 202 points). English alone is collision-free (220/220), Greek
alone is collision-free (11/11), so the loss is created by the mount. Nine
English question words -- ask, question, what, who, where, how, why, when,
which -- share one coordinate, dragged there by a single Greek verb in their
primary domain. `word`, `דבר`, `דברים` and `λόγος` are one point. English is the
mount-time prototype, so the design's depth-anchor roles run backwards: the
depth languages are deleted into the articulation surface.
Four more findings in the same layer:
* 63 of 83 authored alignment edges resolve to nothing -- including every edge
of the two packs serving actually grounds HE/GRC against -- because
`_infer_foreign_pack_ids` maps id prefixes through a table hardcoded to the
original three packs, and a miss is a silent `continue`. The corpus was
grown 7.5x; the growth was never connected.
* `algebra.holonomy.holonomy_encode` never closes. The reverse walk was
deleted at fca6216e and the docstring describing it left in place; `alpha`
is validated and never read (alpha=0.0 and alpha=1.0 return bit-identical
arrays) while core/physics/biography.py:94 passes it.
* Four token-level "resonance proofs" are green because the compared arrays
are the same array -- the open follow-up the 2026-06-14 clause-level finding
explicitly deferred, now closed as decoration.
* The cure was already one layer down and never called: algebra/rotor.py has
shipped `rotor_power` + `word_transition_rotor` (the group geodesic) since
the algebra layer's first commit. A naive lerp leaves the versor group and
VocabManifold.update() correctly refuses it -- the overwrite is what a lerp
degrades into once the guard bites. The guard was right; the escape was not.
This is the missing half of every "geometry adds nothing" verdict: no operator
can distinguish what its ground does not distinguish.
Lands as instrument, not repair -- ../core is the quarry, and L2 is rebuilt in
the keel with these measurements as its provenance. The four decoration tests
are deliberately left standing and named in the analysis doc: green-for-the-
wrong-reason and red are different signals, and replacing them belongs to the
rebuild.
* evals/logos/manifold_collapse.py -- deterministic, bit-exact census
* tests/test_manifold_collapse_floor.py -- both directions plus the
English-alone control that attributes the loss to the mount. Two sabotages
observed red: a real interpolation (ValueError from the versor guard, which
is how the cure was found) and disabling the mount site (37 -> 3, which is
how the 34/2/1 site attribution was measured).
Also lands the FA-1 gate PRE-REGISTRATION -- criterion, corpus, negative
classes and anti-gaming rules committed here so they demonstrably predate any
discrimination number, per the ADR-0252 §5 precedent. The repairs it tests are
mechanism (the geodesic, the closed loop, fail-closed edge resolution), not
tuning; the declared strengths are used as written and are not free parameters.
Charter and gap register corrected where this contradicts them: the corpus is
83 edges not 11, the holonomy machinery is not core/physics/digest.py, and
FA-1's "the gate has never been measured" was false -- the 2026-06-14 negative
exists and is pinned. That error, made in the document that records the
three-negatives correction, is the fourth instance of the pattern this arc
keeps finding, and is recorded as such rather than quietly fixed.
[Verification]: uv run core test --suite smoke -q -> 782 passed in 224.37s, EXIT=0
259 lines
15 KiB
Markdown
259 lines
15 KiB
Markdown
# Finding: the semantic ground collapses — "alignment" is implemented as overwrite
|
||
|
||
**Date:** 2026-07-28 · measured at `472fc0a8` · **FA-1** of the Foundations Audit · **G-25**
|
||
**Status:** substrate finding, L2. Load-bearing for every layer above it, for the keel's
|
||
`logos/` and `algebra/` contracts (K1/K3), and for the standing question of why capability
|
||
has not moved.
|
||
|
||
---
|
||
|
||
## 0. The one-sentence finding
|
||
|
||
`packs.compiler._blend_feature_versors` — the function every cross-language and
|
||
morphological "nudge" in the system calls, at three sites, with three different declared
|
||
strengths — **ignores its strength argument and returns the target verbatim**. Nothing in
|
||
CORE is blended toward anything. Aligned words are *overwritten* by their partners, so the
|
||
three-language manifold ADR-0005/0015 designs as three charts on one semantic space is, in
|
||
the mounted configuration the runtime actually uses, **one chart with aliases**.
|
||
|
||
```python
|
||
def _blend_feature_versors(source, target, strength):
|
||
strength = max(0.0, min(1.0, float(strength)))
|
||
if strength <= 0.0:
|
||
return np.asarray(source, dtype=np.float32).copy()
|
||
return np.asarray(target, dtype=np.float32).copy() # ← any strength > 0
|
||
```
|
||
|
||
ADR-0015 states the requirement this violates twice over: aligned clauses must resonate
|
||
*"without flattening their distinctions"*, and *"cross-language alignment is a weighted
|
||
graph, **not** a translation table."* It is a translation table, applied destructively.
|
||
|
||
## 1. What it costs, measured
|
||
|
||
Instrument: `evals/logos/manifold_collapse.py`. Coordinates are compared **bit-exactly**
|
||
(`ndarray.tobytes()`) — there is no tolerance to tune, and no metric to argue about. Pinned
|
||
by `tests/test_manifold_collapse_floor.py` (registered in `smoke`).
|
||
|
||
| Configuration | surfaces | distinct coordinates | **lost** | groups |
|
||
|---|---|---|---|---|
|
||
| `en_minimal_v1` alone | 220 | 220 | **0** | 0 |
|
||
| `grc_logos_micro_v1` alone | 11 | 11 | **0** | 0 |
|
||
| `he_logos_micro_v1` alone | 9 | 8 | 1 | `אור`/`אוֹר` |
|
||
| **trilingual mounted** | **239** | **202** | **37** | **11** |
|
||
|
||
**Mounting the depth packs — the operation designed to *add* depth — removes 37 distinct
|
||
coordinates.** English alone is collision-free; English mounted with Hebrew and Greek is not.
|
||
|
||
The eleven collapsed groups, verbatim:
|
||
|
||
```
|
||
ask · how · question · what · when · where · which · who · why · ἐρωτάω
|
||
begin · beginning · origin · source · ראשית · ἀρχή
|
||
breath · spirit · πνεῦμα · רוח
|
||
create · image · κτίζω · ברא
|
||
light · φῶς · אוֹר · אור
|
||
read · receive · silence · λαμβάνω
|
||
say · speak · voice · λέγω
|
||
word · λόγος · דבר · דברים
|
||
answer · respond · ἀποκρίνομαι
|
||
truth · אמת · ἀλήθεια
|
||
life · חיים
|
||
```
|
||
|
||
Read the first line again. **Nine English question words occupy one coordinate.** In the
|
||
semantic ground the reader stands on, *who*, *why*, *when*, *where*, *how* and *which* are
|
||
the same point — and they are collapsed not by any English-side decision but because one
|
||
Greek verb, `ἐρωτάω`, shares their primary semantic domain and drags the whole group into
|
||
the multi-language branch. `read`, `receive` and `silence` are likewise one point.
|
||
|
||
**Attribution** (measured by disabling each site in turn):
|
||
|
||
| Site | lost coordinates | what it does |
|
||
|---|---|---|
|
||
| `_apply_mounted_primary_domain_resonance` (mount time, strength `0.40`) | **34** | overwrites *every member* of a multi-language primary-domain group with the English prototype |
|
||
| `_apply_alignment_corrections` (pack load, strength `weight × 0.10`) | 2 | overwrites the home token with the foreign token, last foreign pack wins |
|
||
| `_apply_morphology_cluster_corrections` (compile, strength `0.40`) | 1 | overwrites every same-root form with the cluster prototype |
|
||
|
||
The mount-time site is the dominant cause and carries a comment that named the open question
|
||
precisely — *"it does not yet isolate structural derivation (Hebrew/Greek morphology
|
||
operators) from this function's nudge"* — and pointed at `docs/handoff/ADR-0167-FOLLOWUPS.md`
|
||
§6, **a file that does not exist**. The isolation question is now answered: there is no
|
||
structural derivation left to isolate, because the 0.40 "blend factor" is a 1.00 replacement.
|
||
|
||
## 2. The inversion this produces
|
||
|
||
ADR-0005 assigns the roles: Hebrew as **depth anchor**, Greek as **relational depth**,
|
||
English as **articulation surface**. The mount-time prototype rule is
|
||
`next((surface for language, surface in surfaces if language == "en"), surfaces[0][1])` —
|
||
**English is preferred as the prototype**, and every other language is replaced by it.
|
||
|
||
The design's dependency runs exactly backwards. The depth languages do not anchor English;
|
||
they are deleted into it. `דבר` is not *near* `word` in the mounted manifold — it **is**
|
||
`word`, bit for bit, and so is `λόγος`, and so is `דברים`.
|
||
|
||
At pack-load time the same function runs with the foreign packs in `sorted()` order, so
|
||
`he[דבר]` is first overwritten by `en[word]` and then by `grc[λόγος]`: **which language a
|
||
Hebrew token becomes is decided by the alphabetical order of pack ids.**
|
||
|
||
## 3. Four "resonance proofs" that pass because the compared quantities are identical
|
||
|
||
`docs/analysis/holonomy-resonance-proof-not-robust-2026-06-14.md` audited the *clause*-level
|
||
resonance and found it decoration. It explicitly deferred the token-level tests: *"whether
|
||
each is robust or another cherry-picked configuration is an open follow-up, not covered by
|
||
this finding."* That follow-up is now closed, and the answer is worse than cherry-picking:
|
||
|
||
| Test | Why it passes |
|
||
|---|---|
|
||
| `test_light_alignment_clusters_across_mounted_trilingual_field` | `cga_inner(light, אוֹר)` = `3.0220894813537598` = `cga_inner(light, light)`. The versors are the same array. |
|
||
| `test_triple_alignment_closer_than_other_triples` | same — aligned pairs are self-comparisons |
|
||
| `test_same_root_hebrew_forms_land_closer_than_unrelated_noun` | `דבר` and `דברים` are bit-identical after clustering |
|
||
| `test_structured_morphology_improves_same_root_hebrew_resonance` | asserts structured > unstructured for the same-root pair; structured makes them **identical**, so the "improvement" is total collapse |
|
||
|
||
The last one is the sharpest: a test named for an *improvement in resonance* is green
|
||
precisely because the distinction it should preserve was destroyed.
|
||
|
||
This is the repo's own disease definition met exactly — *"a test that passes under conditions
|
||
that bypass the obligation it nominally proves is decoration, not proof"* — and it is the
|
||
fourth independent instance this arc (CLAIMS.md staleness G-22, `DOMAIN_PACKS` G-23, the
|
||
hollow distinct-evidence guard in PR-12, and now the semantic ground itself).
|
||
|
||
## 4. The second defect: 63 of the 83 alignment edges are silently discarded
|
||
|
||
The audit charter recorded the alignment corpus as *"seed-scale: 11 edges… the ADR's
|
||
prescribed first cases, never grown beyond them."* **That is wrong, and the way it is wrong
|
||
is the finding.** Four packs carry `alignment.jsonl`:
|
||
|
||
| Pack | edges authored | edges that resolve | in serving? |
|
||
|---|---|---|---|
|
||
| `he_logos_micro_v1` | 11 | 11 (as overwrites) | secondary |
|
||
| `grc_logos_micro_v1` | 9 | 9 (as overwrites) | secondary |
|
||
| `he_core_cognition_v1` | **22** | **0** | **primary** (`chat/pack_grounding.py:57`) |
|
||
| `grc_logos_cognition_v1` | **41** | **0** | **primary** (`chat/pack_grounding.py:56`) |
|
||
| **total** | **83** | **20** | |
|
||
|
||
The corpus *was* grown — 7.5× beyond the seed, to 83 edges — and **76% of it is inert**.
|
||
|
||
The mechanism is four lines. `_infer_foreign_pack_ids` decides which foreign pack an edge
|
||
points into by splitting `target_id` on `"-"` and looking the first field up in a hardcoded
|
||
three-entry table:
|
||
|
||
```python
|
||
_PREFIX_TO_PACK = {"he": "he_logos_micro_v1", "grc": "grc_logos_micro_v1", "en": "en_minimal_v1"}
|
||
prefix = edge.target_id.split("-")[0]
|
||
```
|
||
|
||
`en-core-cog-001` has prefix `"en"`, so it resolves to `en_minimal_v1` — a pack with no such
|
||
entry id. The lookup misses, `_apply_alignment_corrections` hits `continue`, and the edge
|
||
vanishes without a warning, a counter, or a test. Entry ids are pack-local; the resolver is
|
||
global and frozen at three pack names, so **any pack beyond the original three is unreachable
|
||
by construction** — including the two that serving actually grounds Hebrew and Greek against.
|
||
|
||
Both branches of the pillar therefore fail, in opposite directions: the packs whose alignment
|
||
*works* destroy distinctions, and the packs serving actually uses have alignment that does
|
||
nothing at all. FA-1's work-order item 4 — *"grow-or-declare the alignment corpus"* — is
|
||
answered before it begins: it was grown, and the growth was never connected.
|
||
|
||
## 5. The third defect: `holonomy` does not close
|
||
|
||
Independent of the collapse, the object named `holonomy` is not one.
|
||
|
||
`algebra/holonomy.py` documents, in the module docstring and in the function docstring:
|
||
|
||
> *A prompt is encoded as the geometric holonomy of its forward+reverse versor walk.
|
||
> **The walk closes**… Forward walk: F = w1·…·wn · Reverse walk: R = (1−alpha)·reverse(wn)·…·reverse(w1) · Holonomy: H = F · R*
|
||
|
||
The implementation computes `F` and returns it. There is no `R`. Commit `fca6216e`
|
||
("Stabilize holonomy accumulation", 2026-05-13) **deleted the reverse accumulation and kept
|
||
the docstring that describes it** — the original at `b80dd57a` did compute
|
||
`H = geometric_product(F, R)`. Consequences, measured:
|
||
|
||
* `alpha` is validated (`0.0 <= alpha <= 1.0`) and then **never read**. `holonomy_encode(v, alpha=0.0)`
|
||
and `holonomy_encode(v, alpha=1.0)` return bit-identical arrays. `core/physics/biography.py:94`
|
||
passes `alpha=alpha` into it — a caller steering a parameter that does nothing.
|
||
* A walk followed by its own reverse does not return to identity (scalar `164.6`, non-scalar
|
||
norm `4044.2`). A closed-loop transport must; an open path product need not, and this is one.
|
||
|
||
ADR-0015's validation gate is **closure** — `holonomy(he) ≈ holonomy(grc) ≈ holonomy(en)`.
|
||
A quantity that never closes has no reason to be invariant across representations, so the
|
||
2026-06-14 negative result was not a surprising failure of the geometry: it is what this
|
||
implementation predicts. That doc's open question RQ2 — *"is the per-token `_position_rotor`
|
||
injection drowning the semantic signal in path-order geometry?"* — was aimed correctly and
|
||
could not see the deletion, because the docstring said the closure was there.
|
||
|
||
## 6. The cure was already in the algebra layer, one import away
|
||
|
||
A linear blend of two versors is not a versor — the Cl(4,1) versor group is not closed under
|
||
addition — and `VocabManifold.update()` correctly refuses non-versors (`residual > 1e-5`).
|
||
Substituting a naive lerp for the overwrite raises `ValueError: Word 'דברים': replacement
|
||
versor residual 3.68e-02 exceeds 1.0e-05` immediately. **The overwrite is what a lerp
|
||
degrades into once the guard rejects it.** The guard was right; the escape was wrong.
|
||
|
||
The correct operation is interpolation *on the group* — the geodesic from source to target in
|
||
Spin(4,1) — and `algebra/rotor.py` has implemented it since the algebra layer's first commit:
|
||
|
||
```python
|
||
blended = geometric_product(rotor_power(word_transition_rotor(source, target), strength), source)
|
||
```
|
||
|
||
Measured on `en[word] → en[light]`: unit residual stays at `~1e-7` for every `t` (three
|
||
orders of magnitude inside the manifold's tolerance), `t=0` returns the source exactly,
|
||
`t=1` the target exactly, and intermediate `t` genuinely interpolates.
|
||
|
||
So the deepest layer built the exact tool the layer above it needed, and the layer above
|
||
never called it. That is the Foundations Audit's thesis — *layers assuming connections that
|
||
were never made* — at the foundation, with both halves of the connection present and unjoined.
|
||
|
||
## 7. Why this is the answer to "why hasn't capability moved"
|
||
|
||
Every geometric experiment this project has run was run on this ground:
|
||
|
||
* the field-reasoner wedge (2026-06-04) — **decoration**,
|
||
* the relational-operator ablation (2026-07-19) — **identical to the symbolic baseline**,
|
||
* ADR-0252 §5 structure-mapping (2026-07-28) — **NO-GO at every attribute weight**,
|
||
* cross-language holonomy resonance (2026-06-14) — **anti-correlated**.
|
||
|
||
The convergent explanation recorded in the Perception Arc was *"wherever the geometric
|
||
operation is isomorphic to the arithmetic it replaces, it reproduces it exactly."* That is
|
||
true and insufficient. The measurement here adds the missing half: **the geometry had been
|
||
erased before the experiments started.** A fifth of the mounted vocabulary shares coordinates
|
||
with something else; the tokens that carry question type are one point; the depth languages
|
||
are aliases of English. No operator defined over that ground can distinguish what the ground
|
||
does not distinguish, so "geometry adds nothing" was the only available result.
|
||
|
||
This does **not** retroactively make those experiments positive — they were well-controlled
|
||
and their verdicts stand for the encodings they tested. It changes what they are evidence
|
||
*of*: they measured operators over a collapsed ground, not the value of geometric operators.
|
||
The reopening criterion in the foundations audit §4 — *"a domain where the geometric encoding
|
||
is not isomorphic to a trivial symbolic computation"* — is now joined by a second, stronger
|
||
one: **re-run on a ground that has not been flattened.**
|
||
|
||
## 8. Disposition
|
||
|
||
Not repaired here. `../core` is the quarry and the evidence source; L2 is rebuilt in the keel
|
||
(`coreai`, K3), where the corrected contract is admitted with these measurements as its
|
||
provenance. Repairing the compiler in place would move every versor in the system, and the
|
||
blast radius includes four tests whose current green depends on the collapse — those are
|
||
**decoration to be replaced with real proofs**, not regressions to be preserved, and doing
|
||
that properly is the keel's K3 work, not a patch here.
|
||
|
||
What lands here instead:
|
||
|
||
1. **The instrument** — `evals/logos/manifold_collapse.py`, deterministic, bit-exact.
|
||
2. **The pin** — `tests/test_manifold_collapse_floor.py` in `smoke`, both directions: the
|
||
collapse may not widen silently, and it may not shrink silently either. Two sabotages
|
||
observed red (a real interpolation → `ValueError` from the versor guard; disabling the
|
||
mount-time site → `37 → 3`, which is also how the 34/2/1 attribution above was measured).
|
||
3. **The four decoration tests are left standing and named here.** They are not weakened in
|
||
this commit: a test that is green for the wrong reason and a test that is red are different
|
||
signals, and the honest move is the one the 2026-06-14 doc made — record exactly why each
|
||
is green, so replacing it is a decision someone makes with the evidence in hand.
|
||
|
||
## 9. What FA-1 asks next — pre-registered separately
|
||
|
||
Whether a *repaired* ground discriminates meaning is a question with a criterion, and the
|
||
criterion is committed before the run (`docs/analysis/fa1-holonomy-gate-preregistration.md`),
|
||
the way ADR-0252 §5 was decided. The mechanism is proven; the capability is not claimed.
|
||
|
||
**Related:** `docs/analysis/holonomy-resonance-proof-not-robust-2026-06-14.md` (the clause-level
|
||
negative this extends) · `docs/plans/2026-07-28-foundations-audit.md` (FA-1) ·
|
||
`docs/assessment/30-gap-register.md` (G-25) · `coreai/docs/atlas/00-atlas-overview.md` (K1/K3).
|