Verifier layer: grounding + consistency over the §5 geometry / reasoning ops

The disposes half of the propose->verify loop. Catches the plausible lie a
grammar check never sees: fluent, confident, wrong.

GROUNDING (anti-hallucination): fit a claim point against every real evidence
neighborhood via engram_reason_point_fit; grounded iff best fit clears an
absolute threshold. Off-model orthogonal residual is the hallucination signal.
Distinct from abduction: asks 'is there any real support at all' and may say no.

CONSISTENCY (contradiction): (a) polarity/negation inversion -- claim lands on
the opposite side of a real polarity axis from the grounded truth (the
reassurance->accusation catch: 'never fought'->'argued'); (b) geometric --
claim inside a forbidden region or beyond a max-distance constraint.

Pure C11, read-only, composes existing primitives only. 29 constructed-case
checks, 0 failures across PERF and ASan/UBSan; macOS leaks 0. el-exposure
deferred (point/variadic-set inputs -- matches reasoning-agent precedent).
Reach checks (formal/causal/predictive) not started; documented.
This commit is contained in:
2026-08-13 01:43:36 -05:00
parent a3358dfc95
commit ca13471745
5 changed files with 671 additions and 0 deletions
+118
View File
@@ -0,0 +1,118 @@
/* engram_verify.h — the VERIFIER layer: GROUNDING + CONSISTENCY over the live
* geometry (engram_geometry.h) and reasoning (engram_reason.h) operators.
*
* The geometry PROPOSES (cheap, creative, sometimes wrong); the verifier DISPOSES.
* This layer catches the class of failure a grammar check never sees: a fluent,
* confident, WRONG output — the "plausible lie". The motivating case: a translation
* that DELETED a negation so "you never fought" became "you argued" — reassurance
* inverted into accusation, grammatical and invisible, catchable ONLY by the geometry.
*
* GROUNDING claim → is there ANY real structure that supports it, or is it
* floating free of the manifold? (anti-hallucination gate)
* CONSISTENCY claim → does it CONTRADICT the established structure? Two catches:
* (a) POLARITY: the claim lands on the OPPOSITE side of a negation
* axis from the grounded truth (the reassurance→accusation catch),
* (b) GEOMETRIC: the claim sits inside a region it must be far from,
* or violates a max-distance constraint to its context.
*
* PURE + READ-ONLY (stdlib + libm only): every function consumes a claim POINT
* (float* in R^dim) plus GeoDescriptor(s), and NEVER touches the store, index, or
* activation. All geometry is delegated to engram_reason_point_fit / engram_geo_*;
* this file only composes and applies thresholds.
*
* FRAME CONTRACT (inherited): the claim point and every descriptor passed together
* MUST share emb `dim` and the same `global_mean` frame — exactly the §5 operator
* contract. A function returns <0 on a dim/frame mismatch or bad argument.
*/
#ifndef ENGRAM_VERIFY_H
#define ENGRAM_VERIFY_H
#include "engram_geometry.h"
#include "engram_reason.h"
/* ═══════════════════════════════════════════════════════════════════════════
* GROUNDING — anti-hallucination. Score how well a claimed POINT is supported by
* the ACTUAL structure: fit the claim against every real evidence neighborhood
* (engram_reason_point_fit → in-distribution Mahalanobis + off-model orthogonal
* residual) and take the BEST supporter. A claim that sits inside real structure
* scores high (grounded); a claim floating far from every neighborhood scores low
* on all of them → flagged UNGROUNDED (a hallucination).
*
* This is an ABSOLUTE-THRESHOLD gate, deliberately distinct from ABDUCTION (which
* always RANKS and picks a winner among competing hypotheses): grounding asks the
* prior question — "is there any real support at all?" — and is allowed to answer no.
* The off-model `ortho_residual` is the sharpest hallucination signal: energy in a
* direction the manifold does not even span.
* ═══════════════════════════════════════════════════════════════════════════ */
typedef struct {
double grounding; /* ∈[0,1]: overall support = best fit score */
int grounded; /* 1 iff grounding >= ground_threshold */
int best; /* index of best-supporting evidence structure, or 1 */
double best_distance; /* full point-to-manifold distance to the best */
double best_ortho; /* off-model orthogonal residual of the best fit */
double nearest_centroid_l2;/* raw L2 to the nearest evidence centroid (coarse) */
int n_evidence;
double* scores; /* per-evidence fit score, higher = better (owned)*/
} GeoGrounding;
/* ext_floor>0 guards zero-extent axes (default 1.0). ground_threshold∈(0,1): the
* minimum best-fit score to call the claim grounded (default 0.5). */
int engram_verify_grounding(const float* claim, int dim,
const GeoDescriptor* const* evidence, int n_evidence,
double ext_floor, double ground_threshold,
GeoGrounding* out);
void engram_verify_grounding_free(GeoGrounding* out);
/* ═══════════════════════════════════════════════════════════════════════════
* CONSISTENCY — contradiction detection. Does the claim contradict the established
* structure? Two independent sub-checks (either can fire; both flags are reported):
*
* (a) POLARITY / negation inversion. A polarity axis p is defined by two REAL
* poles — pole_pos (asserts X) and pole_neg (asserts ¬X):
* p = (c_pos c_neg)/‖·‖ , midpoint o = ½(c_pos + c_neg).
* The claim's side = p·(claim o); the reference's side = p·(c_context o).
* If the two sides have OPPOSITE sign AND both clear the neutral deadzone, the
* claim asserts the polarity opposite to the grounded truth → INVERSION flagged.
* This is the "you never fought"→"you argued" catch: the truth ("never fought")
* sits on the negate pole, the claim ("argued") on the affirm pole → opposite
* sides → flagged, though every word is grammatical.
*
* (b) GEOMETRIC contradiction. The claim sits INSIDE a `forbidden` region it must
* be far from (point_fit score to forbidden ≥ forbidden_thresh), OR it violates
* a max-distance constraint to its context centroid (L2 > max_distance).
*
* pole_pos/pole_neg may both be NULL to skip the polarity check; forbidden may be
* NULL and max_distance≤0 to skip the geometric check. `context` (the grounded truth
* region) is required whenever polarity or the distance constraint is used.
* ═══════════════════════════════════════════════════════════════════════════ */
typedef enum {
GEO_CONSIST_OK = 0, /* consistent with context */
GEO_CONSIST_POLARITY = 1, /* polarity/negation inversion (asserts ¬X where X) */
GEO_CONSIST_GEOMETRIC = 2 /* geometric contradiction (in forbidden / too far) */
} GeoConsistencyVerdict;
typedef struct {
GeoConsistencyVerdict verdict; /* headline (polarity takes precedence) */
double consistency; /* ∈[0,1]: min over the checks (1 = fully consistent)*/
/* polarity sub-check */
int inverted; /* 1 iff a polarity inversion was detected */
double polarity_claim; /* p·(claim o) (signed position on the axis)*/
double polarity_reference; /* p·(c_context o) (the grounded truth's side) */
double polarity_separation; /* ½‖c_pos c_neg‖ (the axis half-length / scale)*/
/* geometric sub-check */
int geo_violation; /* 1 iff a geometric contradiction was detected */
double forbidden_fit; /* claim's point_fit score to the forbidden region*/
double context_distance; /* L2(claim, c_context) */
} GeoConsistency;
/* ext_floor>0 (default 1.0). deadzone_frac∈[0,1): a polarity side within
* deadzone_frac·separation of the midpoint is "neutral" and never triggers inversion
* (default 0.10). forbidden_thresh∈(0,1): fit-to-forbidden at/above which the claim
* counts as inside the forbidden region (default 0.5). max_distance>0 enables the
* distance constraint; ≤0 disables it. */
int engram_verify_consistency(const float* claim, int dim,
const GeoDescriptor* context,
const GeoDescriptor* pole_pos, const GeoDescriptor* pole_neg,
const GeoDescriptor* forbidden,
double ext_floor, double deadzone_frac,
double forbidden_thresh, double max_distance,
GeoConsistency* out);
#endif /* ENGRAM_VERIFY_H */