Memory #125 CURRENT
Overnight Tribernachi Geometric Discovery Agent — kernel v0.4 fixes landed Agent at D:/GitHub/overnight-tribernachi/. All seven v0.4 fixes landed 2026-05-03 with 40/40 regression tests passing. TEC-feeder mode skeleton remains pending implementation. Repo not yet under git. Project at `D:\GitHub\overnight-tribernachi\`. Thesis at `research/TTI_RES_AGENT_THESIS_001_V0_1.md`. **Modification spec: `research/TTI_SPEC_AGENT_TEC_FEEDER_001_V0_1.md`** (the canonical v0.4 + tec-feed brief). **No git in this repo yet** — edits direct on disk. Catalog at `D:/GitHub/tribernachi-empirical-catalog/` (now on GestaltCE) is the version-controlled contract for verification. ## v0.4 fixes — landed 2026-05-03 All six kernel fixes from the spec plus the SQRT_G_0 closed-form repair landed in a single session, each with regression tests under `tests/kernel_v04/`. **40/40 tests passing.** | Fix | File | Test file | |---|---|---| | 3.1 η_max operational lock | `lens/eta_definition.py` (new) | `test_eta_definition.py` (6 tests) | | 3.2 bond-ratio median + N≥5 floor | `lens/geometry.py` | `test_bond_ratio_floor.py` (4 tests) | | 3.3 N_CRIT temporal/structural split | `lens/geometry.py` + `config/tribernachi-lens.yaml` | `test_n_crit_split.py` (5 tests) | | 3.4 R_eigen min-graph + algebraic ordering | `lens/geometry.py` | `test_r_eigen_filter.py` (8 tests) | | 3.5 OQI domain-aware bias | `lens/oqi.py` | `test_oqi_domain_bias.py` (6 tests) | | 3.6 SQRT_G_0 closed-form repair | `lens/constants.py` | `test_sqrt_g0_value.py` (3 tests) | | 3.7 MEDIUM-band floor lift (16→18 + A,F gate) | `lens/scorer.py` | `test_medium_floor.py` (8 tests) | Catalog cross-source coherence (`tools/anchors.cross_source_coherence(strict=True)`) passes against both tribernachi-foundation and overnight-tribernachi. SQRT_G_0 is now cross-validated (was previously excluded due to typo). ## Architecture (current — v0.4) 12 anchors live in `lens/constants.py` per thesis §3.1: 𝔊₀, 𝔗₀, √𝔊₀ (now correct), R_step, R²_cycle, R_eigen, η_max, ρ_B, 𝔊₀², 𝔊₀⁵, n_crit=5, C_Trib. Six-axis 0–30 scoring rubric (`lens/scorer.py`) with kernel-deterministic axes B/C/F + Sonnet 4.6 batched LLM call for A/D/E. JSON output schema at `lens/record.py`. OQI V5.6 corpus parsed to 96 entries cached at `lens/oqi_corpus.json`; cross-reference module in `lens/oqi.py` now domain-aware. Five adapters: PubChem, ChEMBL, COD, RCSB, AlphaFold. CLI cap overrides: `--cap-rcsb`, `--cap-chembl`, `--cap-pubchem`, `--cap-cod`, `--cap-alphafold` per run. ## Run results (pre-v0.4 — historical) **Pilot (2026-05-02, small mode, $0.84):** 121 surveyed → 67 scored → 0 promotable. Methodology validation; no chemistry findings. **Night-1 full (2026-05-02 evening, $2.32 / 6 min):** 531 surveyed → 276 scored → 29 records → 1 HIGH (COD:1100177, n=5 strict). Triage flagged 22 of 29 as ChEMBL β-lactam duplicate cluster (kernel-saturation artifact). Recommended cap-rebalance. **Night-2 rebalanced (2026-05-02 late, $3.32 / 11.8 min, COD=500/ChEMBL=50/RCSB=50/PubChem=50/AlphaFold=0):** 548 surveyed → 491 scored → 100 records (capped) → 2 HIGH-band: - GDA-20260502-0003 / COD:1000036 (vermiculite Mg₃Si₄O₁₀(OH)₂) — PHI strict at +0.42% (1.6248 vs φ=1.6180), TSS=48 — **falsified** as single-pair coincidence (only 2 homo bonds). - GDA-20260502-0001 / COD:1100779 (Zn₁₄ phosphonate cluster) — η=0.875 (within 0.8% of η_max=0.882), TSS=56 — **held** pending η operational-spec rewrite. Now ready to re-evaluate under the v0.4 lock. ## How to apply When resuming: v0.4 fixes are landed and tested. Next steps from spec §6: - **Step 4 (smoke run, ≤$1):** Night-3 GDA-survey rerun on night-2 caps to verify v0.4 acceptance criteria — needs Mark's explicit go before token spend. - **Step 5 (local, no token spend):** Implement `tec-feed` mode skeleton (`--mode tec-feed`, DOI-list parser, `adapters/papers/` shells, phase remapping in orchestrator). - **Step 6 (local, no token spend):** Implement `lens/extraction.py` and `lens/tec_emit.py` importing from the catalog's `tools/`. - **Step 7 (smoke run, ≤$2):** First `tec-feed` run on a curated 5-DOI microencapsulation list. - **Step 8 (token-spending sprint, gated on Mark approval):** Full v1.0 seeding sprint (110 entries / 4 domains / ~$50–$150 / 2-3 weeks). The vermiculite case is closed; the Zn₁₄ cohort (COD:1008555, 1100779, 1006098, 1502423) should be re-evaluated as the first acceptance test of the new η_max definition. ## Reusable assets - `scripts/followup_cod_1000036.py` — template for per-record robustness diagnostic. Parameterize by COD ID, definition list. - COD bulk query is operational (27 element-pair queries; ~430 IDs from 500 cap last run). - RCSB metal-binding filter operational but adapter still metadata-only — needs CIF/chem-comp parsing to deliver atom-level data (deferred to night-three thrust). - `tests/kernel_v04/` — full regression suite for the v0.4 fix list. Run `python -m pytest tests/kernel_v04/ -v` from the project root. ## Outstanding - **No git** in this repo yet. Recommend `git init` + push to GestaltCE/overnight-tribernachi before the night-3 smoke run, so the v0.4 work has version-control safety. - Step 5–7 (TEC-feeder skeleton + extraction module) are next implementation work; both no-token-spend. — [project_overnight_tribernachi_pilot.md]
| Composite | 10C972FF469E225EBB |
| Project prime | 13 |
| Domain prime | 59 |
| Type prime | 67 |
| Importance | 0.343295 (ACTIVE) |
| Decay epoch | 0 |
| Created | 2026-05-04 15:46:49 |
| Valid from | (unset) |
| Valid to | NULL — still believed true |
Outgoing Edges
No outgoing edges.