Core Concepts of SemeSeam
Authoritative definitions from the binding SemeSeam Glossary. Invent the look, never the meaning.
Spine Bundles & The Four Spine Kinds
A spine bundle is a portable reasoning state file consisting of an anchor line and an immutable sequence of typed spines. When a conversation gets long, SemeSeam condenses it into a bundle, allowing subsequent work to proceed with a Fresh Context.
Settled Decisions & Facts
Agreements and conclusions that are locked in. The model must treat them as settled ground and never re-litigate or discard them.
+ [settled] Database schema uses SQLite with strict tables.Negative Constraints & Cages
Explicit boundaries, limits, and prohibitions ("must not happen"). A wall forms a hard cage preventing the model from exploring forbidden paths.
! [wall] Never emit external HTTP/HTTPS URLs in carrel CSS.Essential Background
Operational assumptions, system parameters, or domain definitions necessary for interpreting future reasoning steps.
~ [context] System target width is 1600x1000 with 1100px minimum.Open Steps & Next Moves
Concrete opportunities and next steps ready to be explored. An open step provides structured direction for the next turn.
? [open step] Implement deterministic in-browser hardware check.settled, wall, context, and open step. In raw bundle file serialization and code keys, internal tokens are stored as resolved, wall, context, and affordance.Harvest, Pickup, and Fresh Context
In traditional chat tools, context degrades over time. Models succumb to recency bias, forget negative constraints, and generate repetitive filler.
SemeSeam introduces the Harvest & Pickup cycle:
- Harvest: Condenses the active turns into a new bundle and attaches it to the lineage graph.
- Fresh Context: Discards the conversation history. The model sees only the harvested bundle and any explicitly attached source files.
- Pickup: Resumes reasoning in an uncluttered session with maximum attention budget.
The Reasoning Lattice (DAG)
Reasoning is non-linear. Complex investigations fork into alternative hypotheses or converge when multiple findings agree.
The Lattice is a directed acyclic graph (DAG) of reasoning seeds stored in the subscriber Vault. Each edge records the exact relationship:
continues: Direct step forward along the same line of inquiry.forks: Splitting into divergent experiments or hypotheses.merges: Unifying verified findings from separate branches.
Find & Quote (Deterministic Checks)
AI models often hallucinate subtle wording differences in cited text. SemeSeam eliminates this with Find & Quote:
When a reply includes quotes from files, the local harness runs an exact substring search against the actual files on disk. Only when every character matches does Needle bestow the green proven seal (✓).
If a model invents or paraphrases text, the reply is flagged with a calm warning, and fake citations are excised before reaching the bundle.
Greek Operational Verbs
To ensure tools cannot be spoofed or ambiguous in markdown, SemeSeam reserves five classical Greek verbs for file and memory manipulation:
| Greek Verb | Pronunciation | Exact Operational Meaning | Example Form |
|---|---|---|---|
ΓΡΑΦΩ | Grapho | Write a file (closed strictly by ΤΕΛΟΣ) | ΓΡΑΦΩ path/to/file.py ... ΤΕΛΟΣ |
ΤΕΛΟΣ | Telos | End of that file (delimiter for write operations) | ΤΕΛΟΣ |
ΘΥΜΑΜΑΙ | Thimamai | Memory note to keep across context resets | ΘΥΜΑΜΑΙ Remember port 8791 is offline |
ΣΒΗΝΩ | Svino | Erase a file from the carried workspace | ΣΒΗΝΩ old_test.py |
ΑΛΛΑΖΩ | Allazo | Edit: find this exact string, replace with that | ΑΛΛΑΖΩ path/to/file.py [target] >>> [replacement] |
Descents, The Ratchet, & Relate
A Descent is an autonomous multi-level reasoning trajectory. Descents can be convergent (running until a truth check proves a theorem or puzzle) or divergent (open exploration of new intellectual territory).
- The Ratchet (⛓): Prevents backtracking. If a level repeats a recent conclusion or walks backward, the ratchet immediately forces a redraw.
- Relate (➜): Steers the descent along semantic gradients. By relating concept A to concept B, you move the generative vector away from A and toward B.
- Needle Metrics: Needle grades every level on meaning-novelty and tether distance from the root problem. Every 10 levels, a depth ring milestone is logged.
