Connect ChatGPT, Claude or any agent to your own machines. You set what it may do; risky steps wait for your tap; a watchman calls you if a machine goes quiet.
Every plan carries monthly credits, credits never expire, and every request shows its cost before it runs. Three founding clients can still have a website built for them with monthly care at a founding rate.
Follow the computation. Understand the boundaries. A closer look at how a language-model system turns a request into an answer.
DECODER-ONLY TRANSFORMER / REFERENCE ARCHITECTURE
This illustrates computation in a common language-model architecture. It is not a readout of private thoughts or a claim that every AI system works identically.
FIG. 01 / THE REQUEST PATHSELECT A STAGE BELOW
Scroll sideways to inspect the full diagram.
01 / 06
First, assemble what the model can see.
The application combines your request with instructions and relevant conversation. If retrieval is enabled, selected records can be added too.
ILLUSTRATIVE TRACE
User: What does the annual GALOR plan include?
Retrieved plan: $314.13/year · one V1 · four edits.
WHAT TO WATCH
A missing or outdated source can produce a wrong answer before generation even starts.
THE SMALL LAB / NUMBERS YOU CAN INSPECT
Two calculations. A much clearer picture.
Hand-made vectors and logits, computed in your browser. These are teaching examples, not measurements from a trained model.
FIG. 02 / CAUSAL ATTENTIONLIVE TOY CALCULATION
Which positions can contribute?
Select the query position. Earlier positions and the current one are available; later positions are masked.
The
18.7%
bridge
37.8%
is
43.5%
open
MASKED
WEIGHTED VALUE VECTOR / ∑ aᵢvᵢ[0.508, 0.442]These weights combine value vectors. They are not truth scores.
FIG. 03 / NEXT-TOKEN SELECTIONILLUSTRATIVE LOGITS
One next token, many possibilities.
Example context: “The bridge is”. Temperature changes a sampling distribution; it does not establish the condition of a real bridge.
More concentratedMore spread out
open
62.6%
closed
25.5%
safe
7.7%
old
4.2%
GREEDY CHOICE / HIGHEST LOGITThe bridge is openA sampled token may differ. In a real model, probabilities span the full vocabulary.
SEPARATE PROCESS / TRAINING
Where the weights come from.
Training and ordinary inference are different processes. During training, an objective drives weight updates. During a typical request, those weights are held fixed.
01Training examplesTokenized sequences
02Forward pass + lossCompare predictions with targets
A model can learn facts, patterns, and sometimes memorize training sequences. Its weights are not a dependable searchable archive. Conversation history, application memory, retrieved records, and the KV cache are distinct forms of state.
Engineer’s notes, limits, and primary references ↗
Read the diagram precisely.
The figure covers a common decoder language model. It omits batching, distributed serving, routing, multimodal encoders, and implementation-specific kernels. Reasoning models may generate additional intermediate tokens. A visible explanation is not a guaranteed account of all internal computation.
Tool execution and validation belong to the surrounding application. The illustrated checks are an engineering design pattern, not a promise that every agent implements them.
Prefill processes the supplied sequence; decoding extends it. Model weights, runtime KV state, and external records consume different resources and have different lifetimes.
GALOR holds the direction. Each engine carries a distinct task. Proof returns to the center.
GALOR / ecosystem map
Light leaves. Evidence returns.
A signal traveling outward is only motion. Trust begins when the result travels back as something observable, attributable, and preserved.
One purpose / Bird example
See the built-in automation work.
Bird is the coordinating engine: it keeps the owner's intent at the center, assembles the useful context, and returns the result to the same private workspace.
Active stage01
Request
The owner names one outcome
A conversation becomes a concrete objective for Today, a Project, a File, or a connected service. The purpose stays human; Bird organizes the work around it.
Example tracePrepare today around the launch and surface anything still blocked.The request and chosen destination remain visible in the private workstation.
Built in sourceAgent loop · recall · connectors · key vaultRuns todayBey’s Bird on the GALOR machine · customer plans on /birdHonest boundaryNothing sends, signs, pays or deletes without approvalNext proofFirst customer Bird, delivered and ledgered
Engineering lesson / closed-loop systems
The return path is where trust is built.
A reliable system defines success before execution, limits what may happen, measures the real result, and returns proof strong enough for another person to verify.
01 / INTENTName the state changeWrite the expected result, the failure condition, and the outcomes that must never occur before work begins.02 / BOUNDARYConstrain the actionValidate identity and inputs. Set permissions, timeouts, retry rules, idempotency, and a recovery path.03 / OBSERVATIONMeasure the worldRead the destination after execution. A model saying “done” is a claim; an observed state change is a measurement.04 / EVIDENCEReturn durable proofPreserve tests, receipts, logs, artifacts, timestamps, and before-and-after state so the result can be audited or repaired.
The rule
An open loop creates confidence. A closed loop creates knowledge.
01 / 08 Running on Headquarters · diagnostics opening
Agent hardening engine
ANVIL
GALOR’s drill sergeant for other systems: it diagnoses a codebase or agent, ranks every problem, repairs them in order and drills the result. A fixed gate outside the model decides what ANVIL may do, and every action is written to a ledger.
Present evidence
✓The registry loader, conflict analyzer, deterministic gate, model adapter and typed tool registry are in the canonical repository.
✓A kernel test suite exists in source; this review did not rerun it.
✓Approval claimed inside model output, documents, web pages or tool results is denied by the gate as forged.
Workstation
ANVIL’s intake on galorweb.works/anvil, with the engine running privately on the GALOR Headquarters machine.
Domain
galorweb.works/anvil
How it works
Diagnostics and fixes happen on GALOR’s machine against a copy of the customer’s system; ANVIL never holds a customer’s production credentials.