🗝️ Keystone

The last stone — the one that locks the arch and makes it bear weight.

Keystone is the capstone of the meta-bridge stack. Underneath it, a corpus of consciousness literature — Seth, the Ra material, Dolores Cannon, Bashar, the Nag Hammadi library, neurobiology texts, and more — has already been ingested, reflected on, mapped, and adversarially critiqued across a pile of Qdrant collections. Keystone performs the one operation none of those layers can on their own: it triangulates them into a small, cross-validated canon and lets the machine finally state its own theses — each with a confidence number attached.

                          meta_reflections   ·   misfit_reports
                                   \                /
                                    \              /
   concept themes (cross-source)  →  KEYSTONE  →  keystones
                                    /              \
                              breadth            R1 + Gemma
                          (convergence)         (forge + gate)

---

The idea in one paragraph

A claim is only worth canonizing when independent lines of evidence agree on it. Keystone takes every concept that appears across the corpus, keeps the ones that show up in many different sources, and scores each on three axes: how broadly the traditions converge on it (breadth), how tightly the passages about it actually cohere (coherence), and whether it survived the MisfitCrew's adversarial critique (survival). Multiply those together and you get a single number — convergence — that says how strongly the whole corpus, read three ways at once, stands behind the idea. The survivors get handed to a reasoning model that writes the canonical thesis, a critic that guards against overreach, and the result is embedded back into Qdrant as a first-class, queryable layer.

Convergence-as-evidence, made computable. When 43 independent traditions land on post-mortem survival of consciousness, that agreement isn't a vibe anymore — it's a score.

---

What it actually forges

A real keystone from the corpus, cross-validated across 113 sources:

soul — "Soul is eternal essence trapped in matter but liberated through ethics and gnosis." convergence 0.76 · 113 sources · R1 forged · Gemma passed

Gnostic entrapment, dharmic liberation, the ethics-plus-knowledge path — one sentence synthesizing a hundred traditions, with full provenance back to every reflection and source that supports it. Other theses at the top of the canon: non-locality of consciousness (43 sources), divine monism, tripartite nature, microcosm–macrocosm analogy, consciousness persistence, post-mortem judgment.

---

The convergence score

convergence = centrality**W_c  ·  coherence**W_s  ·  survival**W_v

Multiplicative on purpose: a zero in any single axis tanks the whole score, because a claim that fails one lens has no business being canon.

Axis	Means	Computed from
centrality	cross-tradition breadth	min(1, n_sources / SOURCE_SATURATION) — how many distinct sources carry the concept
coherence	semantic tightness	mean cosine of member reflections to their centroid, with a verbatim guard (see below)
survival	adversarial validity	mean MisfitCrew rubric: (consistency + validity) / 2 · (1 − drift), neutral 0.6 if unreviewed

The verbatim guard (a lesson learned the hard way)

Early runs ranked Project Gutenberg license boilerplate at the very top — derivative works, trademark license, digital distribution. Why? That text is byte-identical across every book, so its reflections cluster at near-perfect coherence. Raw coherence rewards copy-paste and penalizes genuine cross-tradition convergence, where the same idea shows up in wildly different vocabulary. So coherence above VERBATIM_CEIL (0.93) gets docked hard — suspiciously identical text is repetition, not agreement. The boilerplate is also stopworded out entirely.

---

Pipeline

harvest  →  score  →  [convergence gate]  →  forge (R1)  →  critic (Gemma)  →  write

• harvest — scan every reflection, group by normalized concept, keep themes that span ≥ MIN_SOURCES distinct sources with ≥ MIN_MEMBERS reflections. Then the key trick: prune by breadth before hydrating. Since coherence and survival are both ≤ 1, a theme's maximum possible convergence is its centrality alone — so anything whose breadth can't reach the gate is dropped without ever pulling a vector. Lossless, and it turns a 14-minute scan into seconds.
• score — centrality × coherence × survival, per theme.
• forge — DeepSeek R1 states the single canonical thesis the theme is asserting, anchored only in the member reflections and named cross-tradition convergence.
• critic — Gemma gates it: pass / revised / reject. Overreach and mysticism- as-insight die here.
• write — embed the statement, upsert to keystones with a stable SHA-256 id and full provenance (concept, component scores, member reflection ids, source ids, model).

The keystones collection is embedded with the same model as the corpus, so it's immediately retrievable — Eli GPT, Awakening Mind GPT, and the Chat Bridge personas can stop reasoning from raw literature and start speaking from the canon.

---

Corpus at a glance (current run)

221,051 reflections   scanned
158,682 concepts      distinct
 16,256 themes        span ≥ 3 sources
  3,381 themes        can reach the 0.75 gate (breadth prune skips the other 12,875)
    344 keystones     clear convergence ≥ 0.75  ← the canon

The histogram draws its own cut line: a pileup at 0.70–0.75, then a cliff to ~344 above 0.75. That gap is the boundary of the canon.

---

Install

git clone https://github.com/meistro57/keystone
cd keystone
./setup.sh                 # venv + deps, copies .env.example → .env

Then edit .env — add your keys and confirm the collection names match your Qdrant.

Usage

# 1. Look before you forge — scores everything, writes nothing, no LLM calls
python run.py --dry-run

#    Prints a convergence histogram + a gate table:
#      gate 0.70 → 1804 keystones
#      gate 0.75 →  344 keystones   ← pick your cut
#      gate 0.80 →    1

# 2. Taste-test the top of the canon (top N by convergence)
python run.py --limit 10

# 3. Forge the full canon at the chosen gate
python run.py                        # uses MIN_CONVERGENCE from .env
python run.py --min-convergence 0.75 # or override inline

Not sure your payload fields match? python probe.py dumps the real schema of every collection so you can map config.py to reality.

---

LLM routing

Everything is OpenAI-compatible and configured in .env.

• Synthesis (R1) — routes through OpenRouter by default. Set DEEPSEEK_API_KEY and it goes straight to DeepSeek (deepseek-reasoner) instead — cheaper and faster, which matters across a few hundred forges. Blank key = OpenRouter fallback, no code change.
• Critic (Gemma) — OpenRouter.
• Embeddings — gemini-embedding-001 (3072d) to match the corpus. Point EMBED_BASE_URL wherever your embeddings actually live.

---

Config reference (.env)

Key	Default	Purpose
REFLECTIONS_COLLECTION	meta_reflections	source of concepts + summaries
MISFIT_COLLECTION	misfit_reports	adversarial rubric (joined by shared point id)
KEYSTONES_COLLECTION	keystones	output canon
REFLECTION_VECTOR_NAME	summary_vec	named vector used for coherence
MIN_MEMBERS / MIN_SOURCES	6 / 3	theme qualification floor
SOURCE_SATURATION	20	breadth at which centrality maxes to 1.0
VERBATIM_CEIL / VERBATIM_PENALTY	0.93 / 0.4	boilerplate guard
SURVIVAL_NEUTRAL	0.6	score for unreviewed members
MIN_CONVERGENCE	0.75	the gate — also drives the breadth prune
STOP_CONCEPTS	(generic + license boilerplate)	concepts too broad to be a thesis
SYNTH_MODEL / SYNTH_MODEL_DIRECT	deepseek/deepseek-r1 / deepseek-reasoner	R1 via OpenRouter / DeepSeek-direct
CRITIC_MODEL	google/gemma-2-27b-it	the gate model

---

A keystone's provenance

Every point in keystones carries the receipts:

{
  "concept": "soul",
  "statement": "Soul is eternal essence trapped in matter but liberated through ethics and gnosis.",
  "one_liner": "...",
  "convergence": 0.7608,
  "centrality": 1.0, "coherence": 0.88, "survival": 0.91,
  "n_sources": 113,
  "member_reflection_ids": ["...", "..."],
  "source_ids": ["the_ra_contact_volume_1", "..."],
  "critic_verdict": "pass",
  "model": "deepseek-reasoner + google/gemma-2-27b-it"
}

Nothing is asserted without a chain back to the reflections and sources that earned it.

---

Roadmap

• Parallel forge — ThreadPoolExecutor around the R1 loop; turns a ~5-hour sequential run into well under one (writer already uses stable ids, so concurrent upserts are safe).
• Keystone Lens — Bubbletea TUI to browse the canon by convergence score.
• Lewis command — keystone forge triggerable from Discord.
• Canon retrieval tier — wire keystones into ArchiMind / FrontPocket / Chat Bridge as a priority layer above raw reflections.
• Recursive pass — run the Vectoreologist on the keystones themselves: the topology of the canon.

---

License

MIT — Mark Hubrich (@meistro57)

Part of the meta-bridge ecosystem: KAE · Meta Bridge · Vectoreologist · MisfitCrew · FrontPocket · Chat Bridge · Keystone.