OPERATING ON RECEIPTS  , every figure on this page binds to a receipt on disk. The instrument's no-claim boundary is part of what makes its claims credible. real receipt self-test discarded
Active mission · AI receipt discipline

Pre-register what would count. Refuse what will not hold.

Live, awaiting reviewer
4Substrates pointed at
6Verdicts shipped
9RM3 runs audited
4Stable scorer baseline
2Methodologies pre-locked
1Reviewer seat open
realCounts bind to receipts on disk as of 2026-06-23.

No mission, no capture. No receipt, no claim.

This cockpit points an instrument at AI receipts and tells the operator whether the receipts justify the claims the model wants to make. Every figure binds to a receipt on disk. Three of the first four substrates this instrument pointed at produced self-discarding verdicts. That is the system working, not failing.

Operating doctrine

The order the instrument works in.

  • 01Reality is the source.
  • 02Evidence is the record.
  • 03Prediction is the test.
  • 04Verification is the judge.
  • 05Language is the translator.
  • 06Humans remain in command.

Claim boundary

What this instrument refuses to claim, on every substrate, every time.

  • Model lift, held-out pass, benchmark performance
  • Training authorization, adapter promotion, scorer change recommendations
  • Public proof, code-change readiness, capability claims
  • Surgical conclusions about substrates the rig does not own
  • A pattern without a pre-locked prediction
  • Any finding the operator has not approved

The list is enforced in the rig source, not aspirational. What the rig refuses is part of what makes its claims credible.

Loss is allowed. Hidden loss is not.

A selection of receipts shipped on 2026-06-23. Every row is a commit on a branch under TFB main, every artifact carries a no-claim boundary inside it, every routing note pairs to a memo. Nothing enters the instrument as vague data; it enters as accountable evidence.

Evidence stream · 2026-06-23 substrate work
UTCArtifactSourceClassReceiptState
02:15Diagnostic basis memoTelescopeinternalc60858c05cin main
02:45Wound matrix addendum 2Telescopeinternalc98b0c6378in main
03:10Plan v5 vocabulary widening pre-registrationTelescopemethodologyef2a97473bin main
12:30Class C perturbation discriminator verdictTelescopehandoff7d5ba6a2bin main
17:40Distractor receipt observationsTelescopehandoff8380637bbin main
20:15Scorer stability audit across 9 runsTelescopehandoff64a9c016cin main
21:30Class A and B methodology pre-registrationTelescopemethodology38dc70fbfin main

real  Commit SHAs on the working branch · each artifact carries its own no-claim boundary block inside the memo file

A pattern is not truth yet.

Four substrates this instrument has pointed at so far. Three produced self-discarding verdicts on the rig's own internal data, the discipline working at the gate. The fourth is an ongoing read-only application on a separate research surface. Each substrate's verdict was bound to thresholds locked before the data existed.

Constellation → cluster · the methodology
Observation plate, pointed inward

A finding survives only if it holds across two independent vantages  real

A signature seen from one vantage is a constellation, a chance alignment from a single position. Only one that survives a second independent vantage becomes a cluster. Permutation controls (outcome shuffle + event order shuffle) catch the constellations the model would otherwise want to keep.

Four substrates, four verdicts

Sandbox

CONSTELLATION discarded

First pointing. Pattern signature appeared on the named substrate; permutation control reached the same signature on shuffled data. The rig refused to promote.

SWE-bench Verified

NULL · no candidates

K=2 partial pointing. Vocabulary saturated at the lock gate; zero candidates met the pre-locked support fraction. Receipt issued with the null verdict, not silently dropped.

TFB DeepSWE

CONSTELLATION discarded

K=2 anchor pointing on mini-swe-agent traces. Mode 1 outcome shuffle held (0/20 reached 2/2). Mode 2 event order shuffle reached 1/2 on 9 of 20 seeds. Filler-token saturation caught.

RM3 invariant_object

Read-only application

Six artifacts shipped to a separate research surface as third-party calibrated reads against thresholds locked in advance. Today: Class C verdict 6/8 BETWEEN, leaning toward semantic-structure.

Reality is the judge.

A finding becomes credible when independent vantages converge. Today's Class C perturbation experiment shows the full discipline in motion. A non-co-author cross-vendor reviewer PASSED the gate before execution; the rig pre-locked the discriminator thresholds before any data; the receipt landed in one of the pre-locked classes.

Observation plate · Class C wrong-key name perturbation (2026-06-23)

Receipt landed in BETWEEN, leaning toward semantic-structure

base 0/8 → adapter 6/8 → 75 percent of the way from memorization extreme toward semantic-structure extreme. The heal is not pure memorization, not fully name-decoupled. Both semantic structure and name tokens carry weight, neither alone is sufficient.

Parallax on the Class C run real receipt
Vantage A · Telescope thresholds (pre-locked)operator-readable in advance
≈ 8/8semantic-structure heal
≈ 0/8 (= base)token-co-occurrence memorization
betweenmixed signal, report honestly
Receipt landed6/8 BETWEEN
Vantage B · Gemini gate reviewcross-vendor PASS
Reviewer familyGoogle
Build pair familyAnthropic + qwen3
Gate decisionPASS before execution
Provenance sourceorigin_manifest.json
Instrument self-check · validation battery
recovered
Positive control
Known star
An obvious planted pattern a naive baseline is blind to. v0 self-test recovered it.
in class
Real receipt
Class C perturbation
6/8 result mapped to the pre-locked "between" class. Discipline held in the live application.
discarded
Self-discard
Sandbox, SWE-bench, DeepSWE
Three of the first four substrates produced discard or null verdicts. The rig refuses to promote its own constellations.
Instrument holds at the gate and ships verdicts that land inside the pre-locked classes. Class C: tfb_telescope/reports/rm3_wrong_key_name_perturbation_discriminator_class_verdict_20260623.md

A concept can be useful and still limited.

A finding becomes a concept when its bounds are drawn and named. Today's substrate-wide scorer stability audit promoted one concept to the verified-limited tier, with its applicable domain, its non-applicable domains, and its allowed and refused uses written down before any external use.

Scorer-trace attribution

When a substrate ships its runner source with each receipt, score deltas across runs can be attributed line-by-line to specific scorer-source changes. One such instance found across nine RM3 receipts spanning four days: `no_wrong_key_role_array` is the only predicate added then removed, and the removal maps to a measured 3-point base score drift on one specific surface.

verified, limited real
Promotion ladder
  • 0Observation
  • 1Receipted evidence
  • 2Pattern candidate
  • 3Concept candidate
  • 4Prediction-tested
  • 5Verified limited
  • 6Operational (needs approval)
  • 7Human-named
Edges
Verified in
RM3 deepbench-like substrateruns 2026-06-20 to 2026-06-23
Not validated
SWE-benchDeepSWEsandboxany substrate without runner source
Allowed uses
third-party scorer attribution memocross-run audit baselinescorer-state change detector
Blocked uses
scorer-change recommendationclaim scorer was wrongcross-substrate generalization
Receipt chain

c60858c05c (B.1 attribution) → 64a9c016c (substrate-wide audit) → routing note in handoff/

The H1 wound the doctrine warns about is its own build pair.

The build pair is the operator plus one AI design partner inside the same vendor family. The doctrine says nothing is real until a second independent vantage has seen it. One reviewer seat is open. The package is assembled, the receipts are in numbered order, and the design AI's own letters are sealed in Tier 2 so an independent verdict comes from the data first, the narrative second.

Reviewer seat · open

Read the package, write the cold verdict, send it back

The reviewer reads Tier 1 (the reviewer brief, the methodology summary, eleven receipts in numbered order, the rig spec, and a runnable reference rig that reproduces a calibration receipt exactly) and writes a verdict. The verdict has three honest outcomes: the discipline held, one or more pieces did not hold, or one specific additional artifact would close the question. There is no wrong answer.

What the reviewer reads
Reviewer brief with twelve verdict questions, methodology summary covering four substrate applications, eleven numbered receipts, the v0 SPEC, and the runnable rig.
What the reviewer is asked
An independent cold read, written before the Tier 2 narrative is opened. The Tier 2 seal exists so the verdict comes from the data, not from the build pair's framing.
Who fits the seat
Background
Eval-literate. Has read at least one rigorous benchmark methodology paper and either argued with it or written one.
Independence
Not in the build pair's vendor family (not Anthropic). No prior collaboration on this instrument's design. Cross-family AI vantages also welcome as a first-pass calibration.
Time ask
One read of Tier 1, one verdict, one optional follow-up conversation. Reviewer keeps the receipts; the package is theirs to retain.
What you get
Authorship credit on the H1 ratified receipt, named in the methodology publication if the discipline holds.

Every wound becomes a guardrail.

Failures are not hidden here; they are converted. A wound that is receipted becomes a rule the instrument carries forward, a scar, not a bandage. One real healing from today's run is on the table.

2026-06-23  

Almost issued a verdict the doctrine forbids

Healed inside same commit
Trigger
Drafting the distractor experiment observation memo, the instinct surfaced to apply the Class C name-perturbation pre-locked thresholds to the distractor receipt, since both ran on the same row family and lanes.
Wound
That would have been design-after-data: thresholds locked for one discriminator class retroactively applied to a different one. The dev's 2026-06-22 constraint forbids exactly this shape of move.
Root cause
Two perturbation experiments that look similar at the lane level test different discriminator questions. The Telescope's doctrine binds threshold scope to discriminator class, not to row provenance.
System change
Memo reframed as "receipt observations only, not a discriminator-class verdict." The honest output named the absence of pre-locked thresholds explicitly.
New guardrail
Before issuing any discriminator-class verdict, the memo must cite the specific pre-locked threshold set being applied. Citation-by-paragraph, not assumed from surface similarity.
Scar
The doctrine binds the Telescope's verdict scope as tightly as it binds the RM3 substrate's surgical scope. Neither side gets to retro-fit.

Memo at tfb_telescope/reports/rm3_wrong_key_distractor_vs_bodyline_recompute_receipt_observations_20260623.md · commit 8380637bb