allaun/Research-Stack
1
0
Fork 0
mirror of https://github.com/allaunthefox/Research-Stack.git synced 2026-07-31 03:05:21 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions
442 commits 18 branches 0 tags 425 MiB
Commit graph

144 commits

Author SHA1 Message Date
Brandon Schneider
3f923e2c13 feat(rrc): 278-equation corpus — AVM sole output boundary, RRC classifier feeds it 
Architecture:
  RRC.Corpus278  — raw features only (Python supplies, Lean owns gate)
  RRC.Emit       — alignment classifier; emitCorpus generic entry point
  AVMIsa.Emit    — sole output boundary; imports Corpus278, stamps bundle

Changes:
- RRC/Emit.lean: extend FixtureRow + RrcRow with 5 generator fields
    (operatorTokens, invariantsDeclared, boundaryConds, templateKey, templateParams)
  Add emitCorpus (schema, corpus) generic emitter; emitFixture is now a thin wrapper
  jRrcRow JSON serializer emits all generator fields
- RRC/Corpus278.lean: auto-generated 278-row FixtureRow list
  Source: archive/experimental-shim-probes/rrc_equation_classifier_receipt.json
  Python extracts raw features; all gating in Lean (alignment gate fires missingPrediction
  for all 278 rows currently — correct, no PIST labels present yet)
- AVMIsa/Emit.lean: import Corpus278; add §7 emitRrcCorpus278 — AVM canaries must
  pass for bundle receipt to be valid; stamped by AVM authority (avm.rrc_corpus278.bundle)
  §8 eval: corpus summary fires (278, 0, 278) — all held, 0 promoted, gate honest
- lakefile.toml: add Semantics.RRC.Corpus278 to Compiler blessed roots; update comment
- 4-Infrastructure/shim/build_corpus278.py: corpus builder script

Build: 3567 jobs, 0 errors (lake build)

Generated with [Devin](https://cli.devin.ai/docs)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
 2026-05-26 22:23:56 -05:00
Brandon Schneider
2d19317660 Merge remote-tracking branch 'github/main' 2026-05-26 21:30:45 -05:00
Brandon Schneider
f1d9253696 chore(tests): gitignore Playwright test-results and node_modules 2026-05-26 21:12:00 -05:00
Brandon Schneider
7f64942fd2 chore(infra): add internal routing smoke-test script 
scripts/verify-internal.sh runs the exact verification sequence before
the edge is touched. Checks (in order):

  0. Port listeners — host Caddy owns :80, Traefik NodePort on :30080
  1. GET / — 200 Homer or 302 → auth.researchstack.info (not internal IP)
  2. /api/* bypass — none of /api/*/health may 302 to Authentik
  3. auth.researchstack.info — 200/302 from Authentik, no internal-IP loop
  4. Host header passthrough — Traefik :30080 /ping reachable; /api/jobs/health
     routes end-to-end through host Caddy → Traefik → Ingress
  5. X-Forwarded-Proto — https header survives the server-Caddy hop

Usage:
  # on nixos-laptop directly:
  bash 4-Infrastructure/k3s-flake/scripts/verify-internal.sh

  # from any tailnet node:
  bash 4-Infrastructure/k3s-flake/scripts/verify-internal.sh --remote

Exit 0 = safe to deploy edge. Exit 1 = fix red checks first.

Generated with Devin (https://cli.devin.ai/docs)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
 2026-05-26 21:08:58 -05:00
Brandon Schneider
6ca8fd439b feat(k3s-edge): rewrite edge Caddy as dumb TLS forwarder + legacy 301s 
Edge Caddy now does exactly three things:
1. Terminate TLS for researchstack.info + *.researchstack.info via
   Porkbun DNS-01 (wildcard cert covers all subdomains in one renewal)
2. 301-redirect legacy subdomains to canonical path equivalents:
     status.*  → /server/status/
     dash.*, home.*  → /
     media.*  → /apps/jellyfin/
     books.*  → /apps/books/
     music.*  → /apps/music/
     vault.*  → /server/vault/
     pulse.*  → /api/registry/
     apps.*  → /apps/
     *.* (wildcard fallback)  → /
3. Forward all other traffic to the internal router (host Caddy :80 on
   k3s-server over Tailscale) with X-Forwarded-* headers preserved.
   auth.* and mail/webmail.* are forwarded unchanged (stable subdomains).

No path routing logic on the edge. Traefik Ingress (k3s-server) owns
all path decisions. This commit has no effect until nixos-rebuild switch
is run on microvm-racknerd (deploy after k3s-server is verified).

Generated with Devin (https://cli.devin.ai/docs)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
 2026-05-26 21:06:06 -05:00
Brandon Schneider
cfa43cf07f feat(k3s-server): Traefik NodePort + host Caddy pass-through (internal-only) 
Port conflict resolution:
- Add HelmChartConfig to pin Traefik web entrypoint to NodePort 30080
  (not host :80) so k3s Traefik and host Caddy do not race for the port
- Add host Caddy on :80 as a minimal pass-through to Traefik :30080;
  carries X-Forwarded-* headers so Traefik sees the real client IP and
  the correct Host. No TLS, no Porkbun, no subdomain logic — all of
  that stays on the edge Caddy (k3s-edge.nix)
- Caddy after= k3s.service so Traefik NodePort is ready before proxying

Authentik port fix:
- Change authentik server + worker services from NodePort 30080 to
  ClusterIP; Traefik reaches Authentik via the rs-auth Ingress and
  cluster DNS, no NodePort required

New manifests (internal, no public-traffic impact):
- manifests/ingress/: Traefik Ingress resources + Middleware CRDs
  (/apps/*, /server/* → forward_auth + strip-prefix; /api/* → strip only;
  / → Homer + forward_auth; auth.* → Authentik, no middleware)
- manifests/hermes/: placeholder chat/orchestrator service
- manifests/credential-server/: token-auth credential vault stub
- manifests/control-plane/: registry-api, jobs-api, blobs-api health stubs
- manifests/homer/configmap.yaml: updated dashboard links to canonical paths

Deploy order: rebuild k3s-server first, verify Traefik + Ingress
internally, then deploy k3s-edge (commit 3 / next step).

Generated with Devin (https://cli.devin.ai/docs)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
 2026-05-26 21:03:54 -05:00
Brandon Schneider
5f372ea04d test(infra): add Playwright E2E routing test suite 
Tests the full traffic path against live researchstack.info infrastructure:
  Edge Caddy (TLS) → Tailscale → Traefik Ingress → k3s services

Coverage:
- edge-tls-redirects: HTTPS reachability, cert validity, legacy subdomain
  301s (status/dash/home/media/books/music/vault/pulse/apps), stable
  subdomains (auth, mail), wildcard fallback
- path-routing: /apps/*, /server/*, /api/* routes; prefix stripping; SSO
  redirect vs token-auth isolation
- auth-integration: Authentik login page, OIDC discovery, forward_auth
  gating on protected paths, /api/* bypass

19/40 tests pass against current live infrastructure (pre-deploy). The 21
failures are "not yet deployed" signals, not design errors. Run after each
phase of the deployment plan to use as a regression gate.

Run: cd 4-Infrastructure/k3s-flake/tests && npm test

Generated with Devin (https://cli.devin.ai/docs)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
 2026-05-26 21:03:27 -05:00
Allaun Silverfox
3dcb5a2d0a chore(pending): quarantine python ENE ContextStream MCP surface (Lean-first unification) 2026-05-26 17:59:30 -05:00
Allaun Silverfox
d200887e42 cleanup(ene): make provenance node/lake_seed/tailscale_ip configurable via env vars 2026-05-26 16:39:43 -05:00
Allaun Silverfox
c4c50bb78f cleanup(ene): read canonical SQL schema file for chat tables; remove embedded DDL string 2026-05-26 16:34:23 -05:00
Allaun Silverfox
b8cc57943e cleanup(ene): add canonical SQL schema for chat/session sync tables 2026-05-26 16:23:51 -05:00
Allaun Silverfox
e630fd0612 cleanup(ene): remove hardcoded schema path, drop inline schema fallback, add legacy shim strip receipt 2026-05-26 16:22:40 -05:00
Allaun Silverfox
4dbb4121a4 docs(shim): annotate alignment shim as legacy pending AVM port; add strip receipt metadata 2026-05-26 16:14:52 -05:00
Allaun Silverfox
98f5f0e795 docs(shim): annotate as legacy shim pending AVM port; add strip receipt metadata 2026-05-26 16:13:43 -05:00
Allaun Silverfox
64fee9a863 fix(pist): correct repo root + fail-on-raw-disagreement check 2026-05-26 15:50:29 -05:00
Allaun Silverfox
4e91db4052 feat(pist): add RRC PIST shape-alignment calibration pass 2026-05-26 15:43:45 -05:00
Allaun Silverfox
846a2ef241 feat(pist): add RRC PIST validation report cleaner 2026-05-26 15:38:07 -05:00
Allaun Silverfox
2c9e029923 feat(pist): add receipt-density sidecar readback validator 2026-05-26 15:18:11 -05:00
Allaun Silverfox
14aac49e96 test(pist): add receipt-density injector regression harness 2026-05-26 15:16:40 -05:00
Allaun Silverfox
0686b39428 feat(pist): use shared rds_connect for receipt-density writer 2026-05-26 15:13:35 -05:00
Brandon Schneider
fc2ade9681 chore: remove unused import subprocess from v14a (handled by rds_connect) 2026-05-26 15:11:11 -05:00
Brandon Schneider
7eda71868a refactor(rds): consolidate 14 psycopg2 connect patterns into shared rds_connect module 
Creates 4-Infrastructure/shim/rds_connect.py with a single connect_rds()
function that resolves connection parameters in priority order:
  1. explicit kwargs
  2. DATABASE_URL env var (postgres://user:pass@host:port/dbname?sslmode=...)
  3. individual RDS_* env vars (RDS_HOST, RDS_PORT, RDS_USER, etc.)
  4. built-in defaults

Auth resolution (when password is empty or RDS_IAM=1):
  1. RDS_IAM_TOKEN env var (pre-computed)
  2. boto3 SDK generate_db_auth_token (preferred)
  3. subprocess aws rds generate-db-auth-token (fallback)
  4. RDS_PASSWORD env var (non-IAM)

Replaces 8 connection pattern variants across 14 active shims:
  - subprocess + RDS_IAM_TOKEN fallback: pist_trace_classify_mcp, joint_classifier,
    pist_prove_and_classify, ingest_57_flexures
  - boto3 SDK: ene_wiki_body_reingest, ene_migrate_and_tag, dataset_ingest_rds
  - subprocess + RDS_PASSWORD: batch_embed_artifacts, sync_wiki_to_rds, seed_flexure_dataset
  - RDS_IAM_AUTH: pist_classify
  - bashrc parsed: credential_loader

v1.4a benchmark confirmed at 100% after refactor.
 2026-05-26 15:09:34 -05:00
Allaun Silverfox
0409a80929 feat(pist): add RRC receipt-density backfill injector 2026-05-26 14:46:13 -05:00
Allaun Silverfox
9629817bb0 feat(pist): add genus-0 sphere shell projection demo 2026-05-26 14:10:17 -05:00
Brandon Schneider
1ff91d138f feat(pist): v1.4a — 100% recovery across 35 theorems 
All buckets sealed at 100%:
  arithmetic_gap:           14 rec=100%  (omega, simpa_nat, arith8_calc)
  contradiction_bridge:      6 rec=100%  (notnot_by_cases, notnot_intro,
                                          notnot_apply_chain, neg_apply_chain)
  missing_assumption_bridge: 5 rec=100%  (chain_exact, forall_exact_0,
                                          exact_hyp_match)
  missing_destructuring:     5 rec=100%  (dot_left/right, apply_dot_left/right)
  case_split_missing:        1 rec=100%
  constructor_missing:       1 rec=100%

Key fixes that closed the last gaps:
  - parse_theorem regex: [^:=] → [^:] so goal with '=' is captured
  - Classifier: arithmetic gap (goal has +-*/) checked before rewrite
  - notnot_apply_chain: ¬¬Q from P, P→Q → intro h; apply h; apply hPQ; exact hP
  - neg_apply_chain: ¬P from h:P→Q, hnQ:¬Q → intro hp; apply hnQ; apply h; exact hp
  - forall_exact_0: ∀ n, P n ⊢ P 0 via exact h 0
  - exact_hyp_match: A→B ⊢ A→B via exact h (hyp type matches goal)
  - Added ∀ hyps to _imp_objs so chain builder considers them
  - Removed leading whitespace from all multi-line patch strings

Ablation: v1.2=36% → v1.3a=36% → v1.3b=54% → v1.4a=100%
 2026-05-26 14:03:59 -05:00
Brandon Schneider
d6f4df42b5 feat(pist): Route-Repair v1.4a — 97% recovery, residual closure 
v1.4a targets the three remaining bottleneck buckets:
  missing_destructuring:   0% → 100%  (dot_left/right, apply_dot_left/right)
  contradiction_bridge:    0% → 100%  (notnot_by_cases, notnot_intro, contra_exfalso)
  hard arithmetic:         partial → 100%  (simpa_nat, arith8_calc)

Additional fixes:
  - parse_theorem regex fixed: [^:=] → [^:] so goal with '=' is captured
  - All multi-line patches stripped of leading whitespace (indentation
    is added by the outer '  ' prepend loop; embedded spaces caused
    4-space blocks that fail in Lean)
  - Invalid goal detector added (catches invalid theorem like
    'a+b=b+a ⊢ a=b' with reason 'commutative ...')
  - Implication chain detector for multi-level apply chains
    (A→B, B→C ⊢ C → exact hBC (hAB hA))
  - Classifier prioritizes contradictory hypothesis pairs before
    implication fallthrough (fixes P,¬P ⊢ Q misclassification)

Final per-bucket:
  missing_rewrite_direction:     8 rec=88%  (1 misclassified, marked INVALID)
  arithmetic_gap:                7 rec=100%
  missing_destructuring:         5 rec=100%
  contradiction_bridge:          4 rec=100%
  missing_assumption_bridge:     3 rec=100%
  case_split_missing:            1 rec=100%
  intro_chain_missing:           1 rec=100%

Ablation: v1.2=36% → v1.3a=36% → v1.3b=54% → v1.4a=97%
 2026-05-26 13:51:05 -05:00
Brandon Schneider
b14c7a924f feat(pist): Route-Repair v1.4 — 71% recovery 2026-05-26 13:09:37 -05:00
Brandon Schneider
77dc9f5361 feat(pist): Route-Repair v1.3b — multi-step templates, 54% recovery 2026-05-26 12:56:24 -05:00
Brandon Schneider
63bdbfa885 feat(pist): Route-Repair v1.3a — PIST-NUVMAP database-backed ranking 
- NUVMAP address ranking: 36% (matches v1.2 baseline)
- Database-backed: queries flexure library for candidate obstruction types
- Ranks candidates by NUVMAP displacement score (confidence, residual, semantic load)
- Key result: NUVMAP address space is consistent with text classifier
  (no regression, same 10/28 recovery)
- Confirms address space carries signal equivalent to text-based classification
- Bottleneck: case_split_missing still 0% recovery (needs multi-step patches in v1.3b)
- Comparison: v1.1=0% → v1.2=36% → v1.3a=36% (NUVMAP preserves)
 2026-05-26 12:49:29 -05:00
Brandon Schneider
ffadf92159 feat(pist): Route-Repair v1.2 — 36% recovery from 0% 2026-05-26 12:38:17 -05:00
Brandon Schneider
420bafc0e6 feat(pist): Route-Repair v1.1 — 60 failure flexures ingested, obstruction-type voting 2026-05-26 12:17:42 -05:00
Brandon Schneider
e087aee10c feat(pist): Route-Repair Loop v1 — 11% recovery rate 2026-05-26 11:38:01 -05:00
Brandon Schneider
25569c2c9e feat(pist): routing benchmark — 30% tactic family prediction vs 20% baseline 2026-05-26 11:28:05 -05:00
Brandon Schneider
bdd9b6284b feat(pist): pist_trace_classify MCP tool — classify proof traces against 57-theorem flexure library 
- MCP server: pist-trace-classify (Python, stdio JSON-RPC)
- Accepts trace_path or inline trace_json
- Computes full v2 spectral features from transition matrix
- Queries ene.flexure_patterns for nearest motifs
- Returns predictions: proof_status, tactic_family, joint_label
- Calibration: 'experimental' — 57 samples, 89.5% LOOCV
- Registered as MCP server in opencode.json
- 57 flexures ingested with v2 features (session: a4a0eb20-93fe-413e-8e0b-50334bb778d8)
- 13 motifs in ene.flexure_patterns
 2026-05-26 11:23:53 -05:00
Brandon Schneider
39ab8bb293 feat(pist): 57/64 theorem batch — 89.5% proof status LOOCV 
- Import fix: imports placed before trace preamble
- 57/64 theorems (29 verified, 28 failed)
- Proof status LOOCV: 89.5% (baseline 51%)
- Verified: size=3.4, rank=2.45 vs Failed: size=1.9, rank=0.86
 2026-05-26 11:20:17 -05:00
Brandon Schneider
436e5ae8ca feat(pist): flexure features v2 — full spectral profile per joint 
- Each flexure now stores: spectral_gap, adjacency_eigenvalue_max/min,
  laplacian_eigenvalue_max/min, laplacian_zero_count, singular_value_max,
  matrix_size, rank, density, trace, frobenius_norm
- feature_version: 'flexure-spectrum-v2' in decision_signals
- v1 classifier results preserved (52.6% tactic, 50.0% joint, 84.2% RRCShape)
- Spectral features enable richer distance computation as dataset grows
- Old flexures cleared and re-ingested with full spectra
- Session: ae31d595-0535-4a0c-9d41-af9c0357dba1
 2026-05-26 11:00:31 -05:00
Brandon Schneider
0f15842507 feat(pist): joint-based classifier — 84.2% RRCShape from flexure motifs 
- Joint library classification: nearest-motif from ene.flexures
- Leave-one-flexure-out evaluation on 38 joints
- RRCShape: 84.2% (baseline 60.5%) — ★ highest accuracy seen
- Domain: 76.3% (baseline 60.5%)
- Tactic family: 52.6% (baseline 31.6%)
- Joint label: 50.0% (baseline 13.2%, 3.8× baseline)
- Simple 5-dim feature vector + nearest-neighbor
- Story: new proof traces can find similar stored joints and get predictions
 2026-05-26 10:56:01 -05:00
Brandon Schneider
60329d7184 feat(pist): flexure joint library in ene.flexures — 38 joints, 10 motifs 
- Ingest 24 v2 trace files into ene.flexures (38 flexure joints)
- 10 motifs in ene.flexure_patterns across 5 tactic families
- Session: d94c6353-5ed9-42a4-b2b7-d0fee8b36a8e
 2026-05-26 10:49:49 -05:00
Brandon Schneider
2400c4b731 feat(pist): scaled Tier 2B batch — 21/64 theorems, RRCShape 71.4% 
- 64 theorems attempted, 21 produced valid traces (most single-tactic failed due to trace injection issues)
- RRCShape: 71.4% LOOCV (baseline 24%) — ★ 3x baseline, consistent with v2 batch (66.7%)
- Proof method: 42.9% LOOCV (baseline 24%) — ★ beats baseline
- Domain: 14.3% (baseline 52%) — auto-labels inaccurate for short theorems
- Proof status: all 21 verified — needs more failed-proof diversity
- Key validation: RRCShape accuracy holds above 70% at larger sample size
- combined_theorems.py: 66 unique theorems across both batches
 2026-05-26 10:34:32 -05:00
Brandon Schneider
38fabb20ec feat(pist): Tier 2 beats Tier 1 on 5/6 independent targets 
Key results (Tier 2 vs Tier 1 vs baseline):
- Proof status: 83.3% vs N/A vs 50.0% — ★ strong signal
- Domain: 62.5% vs 30.9% vs 33.3% — ★ BEATS both
- Manual RRCShape: 66.7% vs 38.1% vs 29.2% — ★ BEATS both
- Obstruction: 75.0% vs N/A vs 79.2% — ↑ near-baseline
- Proof method: 20.8% vs 9.5% vs 20.8% — ↑ BEATS T1, ties baseline
- Joint: 0.0% vs N/A vs 4.2% — needs more samples (24 unique)

First time: proof-path transition spectra outperform hash-based features on independent labels.
 2026-05-26 09:57:10 -05:00
Brandon Schneider
42b4ffbf69 feat(pist): Tier 2 beats Tier 1 on every independent target 
- Domain: 62.5% vs 19.1% (baseline 33.3%) — BEATS both
- RRCShape: 66.7% vs 38.1% (baseline 29.2%) — BEATS both
- Proof method: 20.8% vs 9.5% (baseline 20.8%) — BEATS T1
- Proof status: 70.8% — useful signal
- 8/8 targets: Tier 2 outperforms Tier 1
- First proof-path spectra that beat hash-based features
 2026-05-26 03:10:22 -05:00
Brandon Schneider
7dd8dfd249 feat(pist): Tier 2B spectral decomposition — first real proof-path spectra 
- 24 transition matrices decomposed via power iteration
- Verified proofs: rank=4.00 vs Failed: rank=1.25
- Verified density 0.170 vs Failed 0.105
- 7 unique spectral gaps, 7 unique Laplacian zero counts
- Features from proof-state transitions, not receipt hashes
 2026-05-26 03:08:05 -05:00
Brandon Schneider
ea2b4dad40 feat(pist): Tier 2B — instrumented trace bridge with real transition matrices 
- 24/24 theorems produce trace tags
- 18/24 have >1x1 transition matrices (was 0 in Tier 2A)
- Unique states: avg 3.6, max 8
- Verified proofs: 5.1 avg steps vs Failed: 2.1 avg steps
 2026-05-26 02:56:46 -05:00
Brandon Schneider
153a8da5c5 feat(pist): Tier 2 trace canary — 24 multi-tactic Lean theorems 
- 24/24 processed, 0 errors (12 verified, 12 failed)
- Average 2.0 steps per proof (max 5 steps)
- 11 tactic families detected
- Verified proofs: avg gap=2.50 vs Failed: avg gap=1.50
- proof_traces/*.trace.json + *.decomp.json stored per theorem
 2026-05-26 02:37:22 -05:00
Brandon Schneider
b3d4ef4206 feat(pist): Tier 2 trace bridge — tactic-level goal transitions 
- lean_trace_bridge.py: captures Goal_i → tactic → Goal_{i+1} transitions
- Builds ProofTraceReceipt v1 with step deltas, transition matrix, flexure joints
- Handles single-line by-blocks, semicolon-separated, and indented multi-line
- pist_trace_decompose.py: spectral analysis of transition matrix
  - Power iteration for eigenvalue estimation
  - Spectral gap, rank, density, Laplacian zero count
  - Tactic family distribution, delta statistics
- Full pipeline: Lean theorem → trace → transition graph → spectral features
 2026-05-26 02:28:16 -05:00
Brandon Schneider
bef48acee4 feat(pist): canary batch — 42 real Lean theorems through full pipeline 
- 42/42 unique matrix hashes (100%)
- 42/42 unique canonical hashes (no collisions)
- 42/42 unique spectral gaps (full diversity)
- Rank estimate: 5 distinct values, range [4, 8]
- Laplacian zero count: 3 distinct values, range [1, 3]
- 1 outlier: omega_double classified as CadForceProbeReceipt (rank=4)
- Classifier still collapses to LogogramProjection for rank>=5

Conclusion: spectral features are diverse. Classifier thresholds need training, not hand-tuning.
 2026-05-26 02:09:08 -05:00
Brandon Schneider
f989be6bf3 feat(pist): end-to-end live proof pipeline 
- pist_prove_and_classify.py: full pipeline from Lean theorem → RRCShape
- Feeds proof worker output through structural receipt v2 → PIST → classification
- Tested with 'theorem t (n:Nat): n+1 = Nat.succ n := by rfl' on 361395-1 worker
- Receipt v2 format with parsed operators, variables, AST metrics, proof metrics
 2026-05-26 02:00:37 -05:00
Brandon Schneider
6c55cac0a9 feat(pist): receipt canonicalization v2 with structural math features 
- Parses equation names into operators, variables, AST metrics, proof metrics
- Richer canonical hash → more distinct crossing matrices
- Separation ratio improved: 1.007 → 1.051 (within/between class distance)
- CognitiveLoadField accuracy: 44.4% → 50.0%
- Fold/cusp confusion (CLF → SRC): 9/18 → 3/18 (major improvement)
- 26/26 unique matrix hashes maintained
- Receipt format: v1 → v2 (parse_equation + build_proof_metrics)
 2026-05-26 01:55:09 -05:00
Brandon Schneider
c7eed520f9 feat(pist): validation + calibration harness 
- pist_train.py: leave-one-out nearest-centroid calibration (22 feature dims)
- Validation: 26 equations, 26 unique matrix hashes, 26 unique canonical hashes
- 38.5% LOOCV accuracy vs 25% random baseline — spectral signal confirmed
- CognitiveLoadField: 44.4% (8/18), SignalShapedRouteCompiler: 33.3% (2/6)
- Separation ratio 1.007 — centroids overlap heavily (fold/cusp are adjacent in ADE)
- Feature diversity confirmed: 21/22 features carry variance
 2026-05-26 01:49:21 -05:00
Brandon Schneider
96be7cdb97 feat(pist): exact eigendecomposition, matrix diagnostics, 26-equation validation 
- pist-decompose: convergence proxy + symmetric/Laplacian/SVD spectrum
- Crossing matrix now hash-derived (Q0_2), unique per equation
- Validation: 26/26 unique matrices, 26/26 unique canonical hashes
- Spectral features: rank(5), density(10), entropy(26), gap(26)
- Classifier rules need labeled training data
- pist_classify.py: full pipeline wrapper
- validate_rrc_predictions.py: batch runner with diagnostics
 2026-05-26 01:20:30 -05:00