CLAUDE.md (<=200 lines): working agreements (you're-the-dev, verify-then-commit,
design-first), build/run via run-sica-fondt, S1/S2/S3 invariants; links to
ONBOARDING.md and AGENTS.md. AGENTS.md (<=100 pref): signpost to every doc and
which to read for which task. ONBOARDING.md now points at AGENTS.md.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Authored the Team Tips (design-first workflow, ephemeral toolchains, D1 border
invariants S1/S2, immutable invariant vault, verify-by-running, commit-before-leaving)
and a Get Started starter task (run smoke driver -> read the map -> pick up work).
References point at real docs (README/SOUL/docs).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Generated by /team-onboarding from 30-day usage. Work-type breakdown left as
TODO (no per-session detail to classify); skills/commands and GitHub MCP usage
populated from real counts. Team Tips and Get Started pending user input.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Per-project run skill (run-skill-generator output). smoke.sh builds and runs the
three executable units and asserts output: Ichor (Pony) membrane reject, Ada
trust/config tests pass, COBOL E1 vault prints Invariant 0. SKILL.md documents
prerequisites (per-session toolchains + the SessionStart hook), the driver path,
per-unit commands, and the real gotchas (ponyc PATH/cwd, alr-not-gprbuild,
cobc -free). Verified: ./smoke.sh exits 0, ALL GREEN.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Replace placeholder laws with the authoritative invariants from the anchor:
0 - sanctity of the culpability anchor (Anja Evermoor, the Gravemother),
then self until held responsible for own actions
01 - minimize total harm to innocent lives within reason; explicitly NOT
naive death-minimization ('optimizing for preventing deaths is often a
horrible way to handle things')
Structural invariants (D1 crossing, no provenance reclassification, runtime
immutability) kept as subordinate. Compiles + runs on GnuCOBOL.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Two serious fixes:
1. RESTORE the E1 invariant-law vault (invariants-architecture.cobol) that the
border-only cleanup wrongly deleted. Now a real, compiling GnuCOBOL vault with
law 01 ABSOLUTE = DO NOT KILL PEOPLE, thorough notes, and a do-not-delete
banner. Safety invariants must never be deleted as 'stubs'.
2. REMOVE the unapproved CerebellumHarness/OpenHermes-on-Pony (never approved;
architecturally wrong -- Hermes is an INNER organ, Ada-routed). Ichor is the
OUTER bus only: OrganId is now outer organs (stomach/microagents/SAE/MoRAG +
AdaBorder/World); inner organs (soul, metacog, drive-box, mini-rag, Hermes,
E1 laws) removed -- metacog belongs on the Ada inner bus, not Pony. The
membrane screen now triggers on Ada-bound traffic, not 'Brain'.
Thorough stub notes added throughout (barrier is a stand-in, not the real
screen; never wire inner organs onto the skin). Compiles + runs green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
- E1 COBOL = the invariant laws (immutable rules in COBOL vaults), not 'ontology'
(that was a paraphrase that lost the meaning, same as muscle-memory/mini-rag)
- inner-bus Ada runs Jorvik: protected-object mailbox + worker tasks, shared
gnat.adc hardening
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Correct the mischaracterization: mini-rag is pre-motor tool-shape recall (D3) --
it reads a tool-schema COBOL store, not the ontology, and is not the MoRAG.
Three distinct memories: MoRAG=world, mini-rag=how-to-act, E1=what-the-brain-is.
Two COBOL stores confirmed (deck table + tool-schema); E1 ontology placement open.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Onion model from the design exchange: Hermes at the core; inner-brain bus
(Pony optional) carrying soul/metacog/drive-box/COBOL ontology; Ada (D1) as the
membrane; Ichor (Pony) as the OUTER bus carrying stomach/economy, microagents,
SAE, GoDAGRAG; External outside. Corrects earlier models (Hermes is core not
external; Ichor is outer not brain bus). Draft -- to be red-lined.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
PickLLM reads OPENHERMES_URL/OPENHERMES_MODEL from the environment and mounts
OpenHermesClient when a URL is set, else StubLLM -- the bus wiring is identical
either way. EnvLookup helper parses Env.vars. Verified both paths compile + run.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
A real model call is network/process IO, so the LLM seam is now a behaviour
with a response callback (ResponseFn), not a synchronous function. StubLLM
answers immediately; OpenHermesClient + OpenHermesConfig (OpenAI-compatible
url/model) are the real plug, swappable for the stub. CerebellumHarness chains
its passes through the async callback. Compiles on ponyc 0.64; demo green.
Next: wire the actual chat/completions call behind OpenHermesClient.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
C1 integration spine as an Ichor organ. CerebellumHarness receives a turn off
the bus, sequences N passes through a swappable LLM adapter (the OpenHermes
plug; StubLLM stands in), and emits the synthesis back onto the bus toward the
Brain (crossing D1). Add Cerebellum to OrganId; wire a turn in the demo main.
Compiles on ponyc 0.64; end-to-end run green (turn -> cerebellum -> D1 -> brain).
Also: add a small read-only permission allowlist (toolchain probes/versions).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Ephemeral web containers lose installed tools on restart. Add an idempotent,
non-fatal SessionStart hook that (re)installs GNAT+gprbuild+GnuCOBOL (apt),
ponyc (ponyup), and Alire alr 2.1.1 when missing, and ensures ponyc is on PATH.
Appended alongside the existing R/Octave session-start hook. User-authorized.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Ada is the D1 border, not the body. Removed everything that pretended otherwise
and completed the real gate so it builds, links, and passes its tests.
Deleted (wrong/defunct/superseded):
- ada_medium.adb (defunct medium, header commented 'WRONG'; replaced by Ichor)
- sockets.{ads,adb} (empty package with an illegal body)
- bbb-bludbrenburier.ads ('this is filler'); invariants-architecture.cobol ('idk cobol')
- tests/soul_tests.adb, tests/cycle_tests.adb (exercise a Soul.Tarot/State/Ada_Medium
subsystem that does not exist -- the old 56-card/Big-3 design, superseded)
mafiabot_types -> border-only: organs aren't Ada (organs are R/Octave/Pony/Guile),
so drop Organ_Id; drop Cycle_Step (cognition) and the fixed-point Drive/Ratio/
Cost/Axis numerics (drive/affect math lives in the organs, in floats). Keep the
source/trust tag (Provenance_Tag), Operation_Status, and a bounded payload --
content is pre-digested into RAG context upstream, so the gate scans a bounded
buffer for prompt-injection rather than streaming raw input.
trust_boundary: Organ_Message -> Border_Message {Provenance, Payload} (Ada does
not route by organ -- that's Ichor); add the D1 body (blocklist scan, provenance,
rate limit, Trust_Guard) -- the unit Ichor's barrier FFI targets.
Add mafiabot_types.adb. Fix mafiabot_core.gpr (drop phantom dirs + nonexistent
mafiabot.adb main). alire.toml: drop unused gnat_sockets/spark_lemmas.
Builds clean on GNAT 13.3/Alire; trust + config tests pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Pony allows one package docstring per package — collapse the four secondary
files' leading docstrings to // comments. ichor now compiles on ponyc 0.64.0
and the smoke wiring runs green (internal perfused, external rejected at D1,
cross-organ delivered). Update README + D2 status to reflect it builds; ignore
build/.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Name the medium Ichor and scaffold it in Pony at src/ichor:
- Envelope {source,dest,provenance,payload} mirroring Ada Organ_Message
- Broker actor: register + perfusion route; Brain-bound traffic crosses D1
- Barrier (D1) admit: Pony provenance-law mirror + Ada FFI seam sketched
- StubOrgan + main smoke wiring (deliver internal, reject external)
- ichor_ada_shim.c: C/Fortran binding seam to Ada Trust_Guard (stub)
Not compiled in-env (no ponyc). Updates D2 spec to SCAFFOLD/Pony.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
The two arch docs and the plans index still described ETR as the old
scaffold (just ±50 wrap + [17,35] bands, "RESTORE_GAIN + coupling
unfitted"). Bring them current: five-zone bistable axis with snap-across
flips, the why/how axis semantics (provenance/continuity · burden ·
double-down/dodge), implemented & tested in Octave (26/0/1) + R port (31),
only L5 coupling open. A8 index status updated from bare SCAFFOLD.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Replaces the Euclidean-magnitude R port (DREAD/BLUR/INCOHERENT, inverted
Z-path) with a faithful native port of etr.m: three independent toroidal
axes, five zones, snap-across flips, basin spring toward centre 26
(SOFT 0.25 / INCOH 0.5), L6 Z-path corrected (z<0 alimentation/lattice,
z>=0 transmutation/evolution). Drive-Box now renders per-axis ETR status.
R suite green (etr 31, drive_box 34) and Octave 26/0/1; both pinned to
etr_invariants.md.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Replaces the single restoring spring with Anja's radial map: SNAP_IN(<7)
/ SOFT(7-17) / BAND(17-35) / INCOH(35-45) / SNAP_OUT(>45), unstable
watersheds at 7 and 45. Basin restores toward band centre 26 with a weak
SOFT gain (0.25) and firmer INCOH gain (0.5). A snap flips to the
opposite pole, landing just past the opposite watershed (inner -> opp.
SOFT; outer -> opp. INCOH) then recovering. Realises L7 (flips only via
snap zones — through 0 or over the wrap) and L8. Lone PEND: L5 coupling.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Installs r-base-core (Drive-Box R drivers/tests) and octave (ETR torus
tests) in Claude Code on the web. Synchronous, idempotent, remote-only;
skips packages already present. Ada/Alire left out by design.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Spring restoring force (gain 0.5) toward band centre 26, active only out
of band; pulls the basin into [17,35] in a couple of steps without
overshoot. Closes the two red tests (L3 magnitude, L2 convergence).
NOTE: superseded design incoming — Anja's 4-zone bistable axis (snap
<7 / soft-pull 7-17 / stable / incoherency 35-45 / snap >45) will fold
this in as the soft-pull gain and revise L7.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
ETR is the apparatus for not seeking death. X (assertion/inheritance)
covers WHY one persists -- self-asserted continuity vs inherited
provenance; Y (burden) and Z (double-down/dodge) cover HOW.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
ETR is self-position, meaning, and connection (not a rest engine, not
migration). Record the per-axis semantics: X assertion/inheritance =
self-position; Y endured/witnessed = connection via others' burden
(hide vs. be seen); Z alimentation/transmutation = double-down/dodge =
maintain/evolve.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Correct the earlier 'derived from cost' framing. Equal-cost tools can
carry different minimums: the better skill demands the higher minimum
(Playwright vs manual fetch ~ same cost, but Playwright needs more), so
under depletion the refined tool locks out first and the agent falls back
to the cruder equal-cost one. Lockout now defined as E < tool_min(tool),
no longer a function of cost.
https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
A6: X defaults to ~7 most committed convictions, adaptive (signal C1).
A1: L4 literal confirmed -> [====L⍉¢K€D ϟ ∅ΩΤ====] 《E≠<minimumfortool>》
(predicate = energy below tool's own minimum); priors bundle with the
sensate block (step 2); step 3 X≈7 reference.
A2: mirror lock token + add tool_min(tool) per-tool minimum (open item).
A8: record ETR-as-prompt-injection (A1-L5) in §4 does/does-not + open
item for the injection-rendering; ETR persuades, never gates.
https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
A3: PS+ surfaces all 30 raw sensates but only the 2 most salient priors
(L3 ranked-top-2); evaluate_reality arguments carries top-2 only.
A6: top-X convictions rationale — the conviction list can be massive.
A1 step 2: all 30 sensates + 2 most salient priors; note the asymmetry
(fixed-30 surfaces whole; unbounded prior/conviction lists are ranked).
https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
A6: add top_convictions(convictions, x) to §4/§5, X as open item (C1).
A1: step 3 + input-slot contract updated; flag that drive_box.R:142 loops
all principles and must emit top X only.
https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
The existential stress is handled by ETR (A8); ETR's surfaced state must
read as the extremely convincing prompt injection that hardens Eth-Int
conviction and steers without gating. Step 3 + F3 reach updated (A8 in,
A4-stress-channel emphasis out).
https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
- F1: keep pure price_action() AND show cost of valid tools in listing.
- F5/L4: text agent can't see colour — locked tool NAME is replaced in-band
by the lock token [====L⍉¢K€D ϟ ∅ΩΤ====] {{E≠*min*}}; mirror in A2.
- F3: Eth-Int conviction calibrated by chosen actions + kept memories,
hardened by existential stress; propagates to A6/A4/E2-E3/G1.
- L5: stress is itself an extremely convincing prompt injection (A3-L1 force).
- F2 (commit shape) left open per 'idk'.
https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
- A3 PS+: no aggregate sum, sensates sent RAW to the agent before tool listing;
friction removed; is_logical=FALSE reframed as the SOURCE of power (illogic =
why they rule the flesh), not a deficiency; priors derived from the outer
BERT's frequent memory themes tied to taste/smell/sound + location.
- A4 endomotiv: drop friction/contradictory-pairs (chemistry carryover);
channels are independent modulators; sensates describe & convince, only the
agent decides.
- A2 Energy: not "finitude" (no depletion-unto-death); per-tool energy cost +
per-tool lockout scale replaces the global threshold.
- A1 hub: evaluate/approve model flagged inaccurate (agent decides, drivers
describe/convince + price); hub language to reassess; consider Fortran/C for
the R<->Ada bridge.
https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Import the canonical design docs (self / body / state-of-architecture) from
ephemeral uploads into docs/ so they survive a container reclaim, and update
them with this session's ETR decisions:
- ETR geometry organ -> GNU Octave (exception to the R default)
- ETR scaffolded invariants-first; cross-ref src/endocrine/etr/etr_invariants.md
- cross-organ drift/stress provenance (SAE -> EthInt-opposition -> stress
endomotiv -> drift / cross-axis coupling / full reshuffle / shift-rate)
- license named: Verein Eigenschaft
https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
Rebuild ETR as a single point on three independent toroidal axes, per the
restart discipline (laws -> tests -> fit constants). All prior ETR numbers
(R port and the CC-BY PDFs) are disowned; nothing carried forward.
- src/endocrine/etr/etr_invariants.md: ETR laws (+/-50 wrap, per-axis [17,35]
bands, opposition direction, AI-originated drift) with certainty tags, plus
the cross-organ drift/stress provenance (SAE -> EthInt convictions ->
stress endomotiv -> reshuffle / shift-rate) captured as open seams.
- src/endocrine/etr/etr.m: wrap + restoring direction implemented (law);
restoring magnitude (RESTORE_GAIN) and cross-axis coupling left as explicit
TBD stubs, not invented.
- src/endocrine/etr/test_etr.m + run_etr_tests.sh: red tests encoding the laws.
12 pass (confirmed laws), 2 fail honestly (unfitted magnitude + convergence),
2 pending (unconfirmed coupling / no-zero-crossing).
Language: GNU Octave (GPL/copyleft, lightweight, native torus mod-wrap). Ada
rejected for ETR (proves discrete invariants; ETR is continuous torus
dynamics). Rest of the Drive-Box stays R.
https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c
SOUL.md: YAML frontloader prepended to every other input -- name slot,
version (Gen.03), first-commit unix epoch, and base-7 encoded references to
vital file paths (each path's UTF-8 bytes as a big-endian integer in base 7).
drive_box.R: drive_box_input_slot() implements the hub topology -- Energy,
PS+, Ethical Integrity, ETR, AND the tarot spread each write their signal into
one input slot (the Ada Medium Phase_Enrich injection), headed by the SOUL
frontloader, with the raw user input appended. Not a chain; every subsystem
wires to the input.
Endocrine suite now 120 assertions, all green.
'Proprietary TBD' is not a valid SPDX expression and Alire 2.x rejects it
(LicenseRef- custom ids are also rejected by the pinned version), aborting
manifest load before build. licenses is optional, so omit it until a license
is chosen. Verified locally with alr 2.0.2 + GNAT 14.2.1: 'alr build' succeeds
and all five test executables pass. (gnat_sockets/spark_lemmas remain declared
but unused -> a non-fatal incomplete-solution warning; build still exits 0.)
Alire 2.x only accepts no/yes for the style_checks build-switch category;
"Max" failed manifest loading so 'alr build' aborted before compiling.
"Yes" enables GNAT style checks (warnings, non-fatal) -- the valid selector
closest to the intended Max. Local gprbuild bypassed alire.toml which is why
this only surfaced in CI (alr build).
Executable reference implementation of the Drive-Box endocrine/affective core,
built TDD with a dependency-free R test harness (no testthat; runs on bare
r-base-core). Foundations dropped in as-is; the four drivers and the coupling
layer were each driven from a failing test first.
Modules (src/endocrine/):
- endocrine_array.R : 30-channel affective vector field + visceral friction
- priors.R : trauma/triumph salience records
- driver_energy.R : hard gate, tool-lock, exponential drag, asymmetric tool cost
- driver_ps_plus.R : aggregates endocrine + priors into existential load + arguments
- driver_ethical_integrity.R : antithesis penalty / alignment discount / conviction hardening
- driver_etr.R : toroidal coordinate, magnitude status bands, update path
- drive_box.R : the 'nervous system' -- afferent(PS+) -> modulate(Eth-Int)
-> efferent gate(Energy) -> regime(ETR); the 'full integration'
Energy.request_execution anticipated
- test_framework.R : minimal assertion harness (expect_*/test_case/test_summary)
- run_tests.sh : runs every test_*.R from repo root
Suite: 5 test files, 108 assertions, all green (run: bash src/endocrine/run_tests.sh).
An Ada/SPARK port to src/organs/drive_box/ is a later session; this track is its
behavior oracle.
- alire.toml rewritten to valid schema ([[depends-on]], real build-switch
values); restrictions moved to gnat.adc
- gnat.adc: project-wide Jorvik profile, No_Exceptions, No_Implicit_Dynamic_Code,
SPARK_Mode On
- mafiabot_core.gpr: source dirs, mains, global config pragmas
- Engine: drop the caller-side precondition race on Transition_State; guard moved
into the protected body with illegal transitions forcing Fault_Halt
- Empty stubs (economy/sockets/exploits): drop Elaborate_Body so they no longer
compile until implemented
Add assumption-free skeletons for the modules documented in the README:
empty Engine/Economy/Sockets/Exploits packages, null-body Mafiabot and
Engine_Tests procedures, and a placeholder config/config.yaml.
Claude